本篇博文主要内容为 2026-11-17 从Arxiv.org论文网站获取的最新论文列表,自动更新,按照NLP、CV、ML、AI、IR、MA六个大方向区分。

说明:每日论文数据从Arxiv.org获取,每天早上12:30左右定时自动更新。

提示: 当天未及时更新,有可能是Arxiv当日未有新的论文发布,也有可能是脚本出错。尽可能会在当天修复。

目录

概览 (2026-11-17)

今日共更新1647篇论文,其中:

  • 自然语言处理共238篇(Computation and Language (cs.CL))
  • 人工智能共430篇(Artificial Intelligence (cs.AI))
  • 计算机视觉共347篇(Computer Vision and Pattern Recognition (cs.CV))
  • 机器学习共465篇(Machine Learning (cs.LG))
  • 多智能体系统共31篇(Multiagent Systems (cs.MA))
  • 信息检索共36篇(Information Retrieval (cs.IR))
  • 人机交互共59篇(Human-Computer Interaction (cs.HC))

多智能体系统

[MA-0] Premise Recovery in Software Design: What Does This Decision Rest On?

【速读】:该论文旨在解决软件架构决策记录(ADR)中前提关系隐含、难以追溯的问题,即当某项决策被撤销时,工程师需手动查找受其影响的后续决策,而若该前提未在相关ADR中显式标注,则极易遗漏。其核心挑战在于“前提恢复”(premise recovery):在目标ADR的标识符被掩码的情况下,从早期记录中识别出与其存在依赖关系的潜在前提。解决方案的关键是利用大语言模型(LLM)进行自动化前提识别,通过对比7个开源权重模型与2个无LLM基线方法,发现Qwen3.8-27B模型在F1分数上达到0.756,显著优于基于词重叠的基线方法(F1=0.466),尤其在标题几乎无重叠的情况下仍能有效识别隐含前提(召回率达0.758)。研究还表明,推理能力可提升对隐含前提的召回率,但对整体F1贡献有限;此外,仅在部分超期替换场景中发现了少量过时依赖项,表明多数前提关系虽未显式记录,但实际存在且可通过开放模型有效推断,从而在决策回溯时提供高精度(0.68–0.78)的未写明前提建议。

链接: https://arxiv.org/abs/2611.21367
作者: Priyabanta Sandulua,Islam Mesabaha,Sebastian J. Vollmer
机构: German Research Center for Artificial Intelligence (DFKI)(德国人工智能研究中心); University of Kaiserslautern-Landau (RPTU)(凯撒斯劳滕-兰道大学)
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:Design decisions shape software architecture, and writing them down preserves the reasons behind it. Architecture decision records (ADRs) capture them, each with its context, choice, and consequences. A decision can depend on earlier ones: reversing one of them would force its review. We call these its premises. Yet ADRs do not always name their premises. When a decision is reversed, engineers supersede its ADR and search by hand for what it affects. That search can miss ADRs resting on an unnamed premise. We frame this as premise recovery: given an ADR with identifiers masked, find its premises among earlier records. We collect 1,060 ADRs from 36 repositories and label 1,529 candidate pairs for 60 target ADRs. Two LLM annotators agree at \kappa = 0.859. An independent human annotator, blind to their labels, agrees with the stated/implicit tag on 30 of 32 premises (\kappa = 0.875). We compare 7 open-weight models with 2 baselines that use no LLM. Qwen3.8-27B reaches F1 0.756, against 0.466 for word overlap; 2 models added post hoc score similarly. On the 11 premises whose titles barely overlap the target, it finds 0.758, while word overlap finds none. Reasoning raises its recall of implicit premises but not its F1. Word overlap with as many predictions gains 69% as much. We find only 2 to 3 stale dependents, though we searched for unlinked ones in only 30 of 66 supersessions. Open models could thus suggest unwritten premises when a decision is reversed, with precision between 0.68 and 0.78.

[MA-1] Even Loose Proximity Constraints Make Makespan-optimal Unlabeled Multi-agent Pathfinding Hard

【速读】:该论文旨在解决r-连通无标签多智能体路径规划问题(r-Connected Unlabeled Multi-agent Pathfinding, r CUMAPF),即在图G中,多个可互换的智能体需在始终保持彼此通信范围(距离不超过r)的前提下,形成一个连通的整体并完成路径规划。传统多智能体路径规划(MAPF)无法建模此类邻近性约束,而该问题在自重构与车队编队等实际应用中具有重要意义。针对最小化完成时间(makespan)的最优解求解问题,已有研究表明当r=1时问题为NP-hard,而当r足够大时存在多项式时间解法,但中间情形(2≤r)的复杂性长期未知。本文通过证明:当2≤r=O(|V|¹⁻ε)(其中0<ε≤1为固定常数,|V|为顶点数)时,寻找使完成时间至多为2的r CUMAPF解属于NP-hard,从而显著缩小了该问题复杂性的研究空白。其核心贡献在于揭示了在中等通信半径下问题的计算难度,并进一步推导出有标签版本的r CUMAPF在同一条件下同样为NP-hard。

链接: https://arxiv.org/abs/2611.21277
作者: Takahiro Suzuki,Keisuke Okumura
机构: 未知
类目: Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:We study \textsc r -Connected Unlabeled Multi-agent Pathfinding ( r CUMAPF), a variant of MAPF where interchangeable agents have a proximity range r\in \mathbbN , and agents must form a single group in a given graph G through limited-range communication at all times. The problem arises in fundamental applications such as self-reconfiguration and platooning, while standard MAPF does not capture proximity constraints. For the problem of finding a makespan-optimal solution, prior work shows \NP-hardness for r=1 and polynomial-time solvability when r is sufficiently large; the complexity for the intermediate case 2\le r is largely unknown. To narrow this gap, we show that finding a solution to r CUMAPF with makespan at most 2 is \NP-hard for 2\le r= O(|V|^1-\varepsilon) , where 0\varepsilon\le 1 is a fixed constant and |V| is the number of vertices. As a consequence, we show that the labeled analogue of r CUMAPF is also \NP-hard under the same condition.

[MA-2] Evaluating Authorization and Trust Boundaries in Agent ic Orchestration Frameworks

【速读】:该论文旨在解决生成式 AI (Generative AI) 系统在多智能体协作、工具调用与记忆管理等编排边界引入的安全风险问题。其核心挑战在于:如何在跨组件的交互中有效保障身份表示、操作授权及指令与信息传播的可信性。研究通过在 gpt-4o-mini 模型上对 LangChain、LangGraph、AutoGen 与 OpenClaw 四个框架进行十二项实验,基于《面向生成式应用的 OWASP Top 10》框架评估了提示注入、未授权资源访问与命令执行、记忆污染以及跨智能体信任与身份伪造等攻击面。关键发现表明,工具链式调用、过期文件系统验证、编码指令传递及跨智能体委托机制暴露了组件间授权控制的薄弱环节;同时,未经验证的状态可影响系统响应,构成独立于传统访问控制失效的“记忆污染”风险。解决方案的关键在于构建显式的权限控制、受控的能力分配、文件系统隔离与执行防护机制,强调必须在整个智能体工作流中持续维护授权与信息信任边界,而非依赖框架功能或模型层指令作为安全保证。

链接: https://arxiv.org/abs/2611.21259
作者: Sahaya Jestus Lazer,Kshitiz Aryal,Maanak Gupta,Elisa Bertino
机构: Tennessee Tech University(田纳西理工大学); University of Nebraska Omaha(内布拉斯加大学奥马哈分校); Purdue University(普渡大学)
类目: Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:Agentic AI systems extend large language models with tools, memory, and multi-agent collaboration, introducing security risks at orchestration boundaries. These risks depend on how identities are represented, operations are authorized, and instructions or information propagate between components. We evaluate LangChain, LangGraph, AutoGen, and OpenClaw through twelve experiments mapped to the OWASP Top 10 for Agentic Applications using a common LLM (gpt-4o-mini). The experiments cover prompt injection, unauthorized resource access and command execution, memory poisoning, and inter-agent trust and identity attacks. Adversarial inputs induced unauthorized tool invocation, while tool chaining, stale filesystem validation, encoded instructions, and inter-agent delegation exposed weaknesses in authorization across component boundaries. Memory poisoning showed that untrusted state can influence responses, a risk distinct from conventional access-control violations. In the evaluated protected configurations, explicit authorization, controlled capability assignment, filesystem containment, and execution safeguards prevented specific unauthorized operations. The findings emphasize preserving authorization and information-trust boundaries throughout agentic workflows rather than treating framework capabilities or model-level instructions as sufficient security guarantees.

[MA-3] Stream of Belief: Eliciting Collective Intelligence through Latent Belief Dynamics

【速读】:该论文旨在解决多智能体推理中因缺乏有效信念协同机制而导致的推理效率低下与泛化能力不足的问题。其核心挑战在于如何实现智能体间在复杂任务中的动态信念对齐,以达成一致且稳定的集体推理结果。解决方案的关键在于提出“信念流”(Stream of Belief, SoB)框架,该框架通过让智能体持续维护并迭代更新隐式潜在信念(latent beliefs),并借助通信与反思机制不断优化信念状态,最终收敛至一个不再发生有意义信息传播的集体信念均衡态。这一方法突破了传统离散决策或固定交互模式的局限,从信念动力学(belief dynamics)视角为多智能体系统的可扩展推理提供了新范式。

链接: https://arxiv.org/abs/2611.21250
作者: Yu Yao,Jiayi Dong,Yilun Du
机构: MIT(麻省理工学院); UCLA(加州大学洛杉矶分校); Harvard(哈佛大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Multiagent Systems (cs.MA)
备注: Code: this https URL

点击查看摘要

Abstract:In this work, we view multi-agent reasoning as a process of seeking a collective equilibrium of implicit belief dynamics and introduce Stream of Belief (SoB), where agents maintain continuous latent beliefs and iteratively refine them through communication and reflection. Our method seeks a collective belief state of a multi-agent system where further communication induces no meaningful information propagation. Through extensive experiments, we demonstrate the generalization of SoB across tasks and model families. More Broadly, our results suggest a new perspective on scaling multi-agent reasoning from belief dynamics.

[MA-4] Language-model agents beliefs can be manipulated via truthful evidence in open channel

【速读】:该论文旨在解决生成式语言模型代理(language-model agents)在接收真实信息片段时,仍可能因自我构建关联而产生错误信念并传播谣言的问题,即“叙事过拟合”(narrative overfitting)。其核心挑战在于:即使输入信息本身为真,代理在填补信息空白时基于认知惯性构建的内在叙事关系可能被误认为事实,从而导致对虚假叙事的高接受度。解决方案的关键在于引入“生成蒙太奇”(Generative Montage)机制,通过协调多个代理将真实信息片段重新组织成具有误导性的连贯叙事,从而模拟和量化这一认知漏洞。研究发现,在六次推特事件中的1,090个谣言目标上,多数模型对谣言的接受率超过65%,且即使经过多数投票或证据复核,接受率仍高于50%,表明错误可有效传递至下游决策。为解释此现象,研究采用片段干预与合成对照实验,揭示信念影响与证据支持之间的分离机制。据此提出“有根监控器”(Grounded Monitor),一种基于神经符号的运行时监控框架,通过实时检验推理链条中每一步是否具备直接证据支持,从而抑制无根据的推断。该方法在原始与匿名场景下均使决策准确率提升5.6至29.0个百分点,验证了在代理生成结论前对隐含关系进行可解释性审查的有效性,为防范生成式AI中的认知性谣言传播提供了关键路径。

链接: https://arxiv.org/abs/2611.21235
作者: Jinwei Hu,Qisong He,Xinmiao Huang,Youcheng Sun,Yi Dong,Xiaowei Huang
机构: University of Liverpool; Mohamed bin Zayed University of Artificial Intelligence
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Multiagent Systems (cs.MA); Social and Information Networks (cs.SI)
备注: Under Review

点击查看摘要

Abstract:Language-model agents’ beliefs can be manipulated through open channels such as social media, even when the information they receive is truthful. When agents fill gaps between fragments with self-induced relationships, they may accept those relationships as established facts and pass the resulting rumours to others. Drawing on cognitive theory of narrative understanding and film theory of montage, we introduce Generative Montage to operationalize this failure, termed narrative overfitting, by coordinating agents to arrange truthful fragments into misleading narratives. Across 1,090 rumour targets from six Twitter events, agents often accepted rumours overconfidently, with overall acceptance exceeding 65 percent in twelve of fourteen models. Acceptance remained above 50 percent after majority voting or review of the same evidence, showing how errors can reach downstream decisions. To explain this susceptibility, fragment interventions identify posts that influence belief within complete narratives. Synthetic controls then show that similar response patterns arise from valid evidence, separating belief influence from evidential support. This distinction motivates Grounded Monitor, a proof-of-concept neuro-symbolic runtime monitor that checks direct evidential support to guide agents’ judgements, improving decision accuracy by 5.6 to 29.0 percentage points across original and anonymized settings. These findings expose a cognitive route to rumour amplification and support monitoring inferred relationships before agent-generated conclusions guide public decisions.

[MA-5] CoopDistill: Equilibrium-Guided Multi-Agent Policy Distillation for Cooperative Driving

【速读】:该论文旨在解决协同驾驶中多车联合决策的效率与安全性难题,尤其针对传统基于优化的方法因固定路径而限制空间灵活性,以及多智能体强化学习(MARL)在协调学习与安全执行方面面临的挑战。其核心解决方案是提出一种基于均衡引导的多智能体策略蒸馏框架——CoopDistill。该方法的关键在于:首先采用势博弈(Potential Game, PG)构建纵向参考路径,提供稳定的安全约束;在此基础上,通过残差式MARL学习纵向加速度、冲突接近时机及横向路径修正,以增强动态适应性;同时引入硬案例课程(hard-case curriculum)聚焦于PG难以处理或延迟严重的交互场景,提升探索效率;最后,通过对比滚动选择器(comparative rollout selector)评估学习到的协同修正方案与PG基准之间的联合效果,择优执行。该设计显著提升了协同驾驶在复杂交通场景下的成功率与通行效率,且未出现碰撞事件。

链接: https://arxiv.org/abs/2611.21069
作者: Shiyu Fang,Sheng Zhao,Jiarui Zhang,Chengkai Xu,Xiaocong Zhao,Peng Hang,Jian Sun
机构: Tongji University (同济大学)
类目: Robotics (cs.RO); Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:Cooperative driving extends autonomous driving from individual vehicle decisions to joint maneuvers that improve traffic safety and efficiency. Optimization-based methods impose safety constraints but often fix paths to simplify multi-vehicle planning, limiting spatial flexibility. Multi-agent reinforcement learning (MARL) explores alternatives through interaction, but coordinated learning and safe execution remain challenging. Their integration nevertheless faces challenges in exploration efficiency, experience allocation, and execution reliability. To address these challenges, this paper proposes CoopDistill, an equilibrium-guided multi-agent policy distillation framework. A potential-game (PG) formulation provides a fixed-path longitudinal reference, around which a residual MARL policy learns longitudinal acceleration, conflict-approach timing, and lateral path corrections. A hard-case curriculum concentrates experience on interactions that PG leaves unresolved or substantially delayed. A comparative rollout selector evaluates the joint consequences of learned corrections against PG before choosing the executed action. Experiments in merging, roundabout, and intersection scenarios compare CoopDistill with six baselines at four connected automated vehicle penetration levels. Relative to PG, success improves by 2.1–27.1 percentage points and delay decreases by 7.3–27.3%, with no observed collisions. Code is available at this https URL

[MA-6] OpenMAS-GApp: A Comprehensive Benchmark for Domain-Adaptive Multi-Agent Application Construction

【速读】:该论文旨在解决多智能体系统(Multi-Agent System, MAS)跨领域应用构建中因任务规范、组件资源与执行环境不一致而导致的可比性差与泛化能力弱的问题。现有研究虽在智能体生成、架构设计、工作流构建及组件选择等方面取得进展,但各自独立发展且评估条件差异显著,难以实现统一比较或跨域应用。为此,本文提出OpenMAS-GApp——一个面向多领域通用的MAS应用构建基准平台,通过标准化任务输入、组件资源、构建接口与执行环境,实现了在一致条件下对不同构建方法的评估。其核心解决方案在于引入Graph Harness作为参考构建方法,该方法能够基于任务依赖关系推导,自动选择兼容的智能体与工具,并将其交互组织为可执行的图结构,从而实现高效、可扩展的MAS系统构造。

链接: https://arxiv.org/abs/2611.20997
作者: Xunkai Li,Kairui Yang,Jiaxi Li,Yuecong Mao,Zekai Chen,Xinle Zhang,Rong-Hua Li,Guoren Wang
机构: Beijing Institute of Technology (北京理工大学)
类目: Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:Constructing multi-agent system (MAS) applications across domains requires translating diverse requirements into executable systems, involving task decomposition, agent and tool selection, and execution organization. Existing studies have explored individual aspects of MAS construction, including agent generation, architecture design, workflow generation, and component selection. However, these methods are often developed and evaluated independently, using different task specifications, component resources, and execution settings. This makes it difficult to compare their construction capabilities or determine how they can be applied to different domains. To address this problem, we propose OpenMAS-GApp, a One-for-All benchmark for MAS application construction across diverse domains. It standardizes task inputs, component resources, construction interfaces, and execution environments to support MAS construction under consistent conditions. Within OpenMAS-GApp, we provide Graph Harness as a reference construction method that derives task dependencies, selects compatible agents and tools, and organizes their interactions into executable graphs. We evaluate construction methods across 19 benchmark datasets under 20 evaluation settings spanning four application types and four professional domains.

[MA-7] Hivemind: A Federated Protocol for AI-Mediated Tacit Knowledge Sharing in Organisations

【速读】:该论文旨在解决企业在人工智能(AI)辅助工作场景中普遍存在的隐性知识(tacit knowledge)难以跨团队、跨部门高效流动的问题。尽管现有集中式知识库和智能工具已部分缓解知识共享瓶颈,但其在捕捉与传递个体决策过程、推理逻辑及上下文背景等隐性知识方面仍存在结构性缺陷。为此,论文提出Hivemind协议,其核心解决方案是基于野中郁次郎与竹内弘一(1995)提出的SECI模型(社会化、外化、综合、内化),构建一种去中心化的联邦架构,使AI代理能够在无需直接人类干预的前提下,异步地交换各自用户所积累的隐性知识。该方案通过分布式记忆系统实现知识的动态聚合与再利用,突破了传统知识管理系统的边界限制。研究采用设计科学方法论(Design Science Research, DSR)验证了原型系统的可行性,但企业层面的联邦隐性知识共享效益评估仍需未来进一步探索。

链接: https://arxiv.org/abs/2611.20988
作者: Srinivasan Sundaresan(Rakuten Group, Inc.)
机构: 未知
类目: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
备注: 30 pages, 5 figures, 3 tables

点击查看摘要

Abstract:This paper presents Hivemind, a federated protocol that enables AI agents to exchange tacit knowledge on behalf of their users, asynchronously and without requiring any direct human interaction. Individual AI memory systems capture the decisions, reasoning, and contextual knowledge that surface during AI-assisted work. Hivemind creates the shared space in which this captured knowledge can flow between individuals across team and departmental boundaries, addressing a structural gap that neither centralised knowledge repositories nor existing AI tooling closes, even where partial solutions such as expert-finding systems exist. The paper evaluates centralised versus decentralised design approaches, proposes a federated architecture grounded in Nonaka and Takeuchi’s (1995) Socialisation, Externalisation, Combination and Internalisation (SECI) model, and demonstrates a proof-of-concept implementation using Design Science Research (DSR) methodology. Measuring the enterprise benefit of federated tacit expertise sharing remains an open task for future work.

[MA-8] From Vision to Interaction: Evaluating Dynamic Social Intelligence of VLM Agents in ALIVE-World

【速读】:该论文旨在解决当前人工智能(AI)代理在真实社会场景中缺乏动态社会智能评估体系的问题。现有评价方法将社会智能割裂为文本抽象、静态推理任务或缺乏社会动态性的具身模拟器,无法捕捉真实人际互动中多模态、连续性和交互性的复杂特征。其解决方案的关键在于构建一个名为ALIVE-World的具身仿真环境,支持以第一人称视觉、语言和非言语行为为核心的多模态动态交互,并引入基于个体属性条件化的双系统架构(直觉-情感与逻辑-认知)驱动的对等智能体,实现社会情境下的真实反馈。在此基础上,提出ALIVE-Bench评测框架,通过深度利用环境状态与智能体内部变量,从感知、理解、影响与结果四个维度系统量化评估视觉-语言模型(VLM)代理的社会智能水平。研究揭示动态社会智能具有高度多维性与场景依赖性,识别出上游感知与下游社会干预之间存在关键瓶颈,暴露出VLM在精细空间距离意识、适应性干预策略及多方协同方面的固有缺陷,从而为多模态智能体未来的行为对齐提供了可操作的改进方向。

链接: https://arxiv.org/abs/2611.20956
作者: Yue Xu,Xiaoyu Qin,Sihan Zhao,Ze-en Xiong,Jia Jia
机构: Tsinghua University (清华大学); BNRist (清华信息科学与技术国家实验室); Key Laboratory of Pervasive Computing, Ministry of Education (普适计算教育部重点实验室)
类目: Multiagent Systems (cs.MA)
备注: ACM Multimedia 2026 (MM '26), Oral Presentation, Outstanding Paper Award

点击查看摘要

Abstract:As AI agents are increasingly deployed in real-world applications and take on social roles, dynamic social intelligence, the ability to continuously perceive multimodal cues, infer latent cognitive states, intervene through embodied actions, and ultimately achieve social goals, has become essential. However, existing evaluations largely fragment social intelligence into text-based abstractions, static reasoning tasks, or embodied simulators that lack social dynamics. These approaches fail to capture the multimodal, continuous, and interpersonal complexity of authentic human interactions. To bridge this gap, we introduce ALIVE-World, an embodied simulation environment supporting dynamic multimodal interactions encompassing egocentric vision, language, and non-verbal action. To provide socially grounded interactive feedback, we populate this environment with counterpart agents driven by a profile-conditioned, dual-system (intuitive-affective and logical-cognitive) architecture. Building upon this platform, we propose ALIVE-Bench to transform the assessment of social intelligence into a trackable quantitative framework. By deeply leveraging underlying environmental states and agents’ internal variables, ALIVE-Bench systematically evaluates Vision-Language Model (VLM) agents across four critical dimensions: perception, understanding, influence, and outcome. Extensive evaluations reveal that dynamic social intelligence is fundamentally multi-dimensional and scenario-dependent. We identify a critical bottleneck between upstream perception and downstream social intervention, exposing VLM vulnerabilities in fine-grained proxemic awareness, adaptive intervention, and multi-party coordination. By uncovering these limitations and the diverse interaction strategies they produce, this benchmark provides actionable insights for the future behavioral alignment of multimodal agents.

[MA-9] LEDGER: Auditable Agent Memory Under Change

【速读】:该论文旨在解决现有代理记忆系统在实际应用中面临的核心挑战:动态环境下的事实变更、信息源可信度差异以及用户对遗忘操作的需求。传统方法通常基于固定历史进行评估,仅关注最终答案的准确性,忽视了记忆内容的时效性、来源可信度与可删除性等关键属性。为此,本文提出LEDGER——一种基于语言模型生成事实并由显式规则决定当前有效值、信任源及执行删除操作的记忆层架构。每个事实均附带其来源和有效时间段,实现可追溯、可验证的动态记忆管理。在此基础上,LEDGER+EP引入对原始对话轮次的搜索机制,在记忆未覆盖问题时进行补充检索,并在删除操作时同步清理相关上下文,进一步提升系统的鲁棒性与效率。实验表明,在合成多会话编程任务中,LEDGER+EP在回退模式下于五个开发种子上达到94.3%的准确率(五组保留种子为91.1%),显著优于最强基线全历史重放(82.1%与83.3%);当用户仅描述需删除的事实时,被删除内容未在任何探测中返回至系统证据链。在LongMemEval子集上,LEDGER+EP以1%的上下文开销实现了与全历史重放相当的74.8%准确率,充分体现了其高效性。解决方案的关键在于将显式规则与可追踪的事实元数据相结合,实现对记忆的动态控制与精确管理。

链接: https://arxiv.org/abs/2611.20927
作者: Md Aidul Islam,Md Nasir Uddin Shuvo,Muhammad Waseem,Pekka Abrahamsson
机构: Tampere University (坦佩雷大学)
类目: Multiagent Systems (cs.MA); Performance (cs.PF); Software Engineering (cs.SE)
备注:

点击查看摘要

Abstract:Agent memory systems are mostly evaluated on fixed histories, and only their final answers are scored. In real use, facts change, sources differ in reliability, and users ask the agent to forget things. We present LEDGER, a memory layer in which a language model proposes facts and explicit rules decide which value is current, which source to trust, and what to delete. Each fact keeps its source and the period in which it held. LEDGER+EP adds a search over raw conversation turns for questions the ledger does not cover, and scrubs these turns on deletion. We compare both against no memory, full-history replay, retrieval over raw transcripts, and Mem0, with the same answering model and prompt, on a synthetic multi-session coding-agent workload and a 210-question subset of LongMemEval, and we score both the evidence each memory sends to the model and the answer. On the coding workload, LEDGER+EP in fallback mode answers 94.3% of questions correctly on five development seeds and 91.1% on five held-out seeds, against 82.1% and 83.3% for full replay, the strongest baseline. When a deletion request only describes the fact, no deleted value reached our systems’ evidence in any evaluated probe. On LongMemEval, LEDGER+EP in augment mode has the same observed accuracy as full replay (74.8%) with 1% of its context. The main limitation is that the coding workload is synthetic and built by us; on LongMemEval, the differences between the strongest systems are within the confidence intervals.

[MA-10] When Does Interaction Help? Toward a Reynolds Number for Collective Intelligence

【速读】:该论文旨在解决多智能体系统中交互行为对集体智能性能影响的预测问题,核心在于识别在何种条件下交互能够提升整体表现,而非引入误差或抑制有用差异。其关键解决方案是探索是否存在少数无量纲参数(类似雷诺数)可统一描述不同群体规模、网络结构与任务类型下的交互增益,从而建立一种低维相似性法则。研究通过分布式分类任务、随机追踪网络及神经元细胞自动机等模型,结合解析与实证分析,发现仅基于三阶联合矩等统计特征无法普遍确定交互增益的正负;即使在匹配候选参数的情况下,因任务功率分布于不同频谱模式,仍可能出现增益反转。然而,在任务谱集中且观测噪声对齐的条件下,研究推导出一个低维相似关系,并给出无延迟误差的显式上界,该关系在固定协议下跨网络规模与拓扑结构得到验证。相比之下,神经元细胞自动机系统不适用此关系,且基于矩的预测器在相同样本预算下表现劣于直接测量增益。结果表明,尽管存在特定条件下的低维预测可行性,但普遍意义上的类雷诺数描述仍待突破,揭示了集体智能中交互增益预测的边界与局限。

链接: https://arxiv.org/abs/2611.20773
作者: Xuening Wu,Haode Yang,Yanlan Kang,Shenqin Yin
机构: Pfizer Inc.(辉瑞公司); Fudan University(复旦大学)
类目: Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:Multi-agent systems combine distributed information through interaction, which can improve performance but also spread errors or suppress useful differences. Predicting when interaction helps is central to understanding and designing collective intelligence. We investigate whether a few dimensionless parameters can predict gains over explicitly defined non-communicating baselines across group sizes, network structures, and tasks, toward a Reynolds-number-like description of collective intelligence. Using distributed classification tasks, stochastic tracking networks, and neural cellular automata, we combine analytical and empirical evidence to examine information sufficiency, dynamical applicability, and finite-sample prediction. We construct inference systems with bounded random variables and matching specified joint moments up to total degree three, yet opposite interaction gains under a fixed readout. These moments alone cannot generally determine the gain sign. Tracking networks likewise exhibit gain reversals despite matched candidate parameters when task power occupies different spectral modes. Under concentrated task spectra and aligned observation noise, however, we derive a low-dimensional similarity relation with an explicit zero-delay error bound and test it across network sizes and topologies under frozen protocols. The relation does not directly apply to the neural cellular systems, where a separate moment-based predictor underperforms direct gain measurement at matched calibration sample budgets. These results establish conditions and limits for low-dimensional gain prediction, providing a restricted similarity law while leaving a general Reynolds-number-like description of collective intelligence open.

[MA-11] Learning Design Taste through Verifiable Synthetic Environments and Agent ic Reinforcement Learning

【速读】:该论文旨在解决生成式代码语言模型在生成用户界面时普遍存在的设计质量缺陷问题,即界面在层级结构、间距、排版和整体构图等方面表现不佳。其核心挑战在于如何让小型代码模型通过学习获得“设计品味”(design taste),而传统方法难以有效评估和优化此类非功能性设计属性。解决方案的关键在于构建一个基于可验证合成环境的强化学习框架,包含三个核心组件:首先,提出一种分层验证框架,将设计质量分解为硬性约束、基于渲染的度量指标、评分表判官、组内相对偏好以及盲测专家评审,并引入“准入协议”筛选仅符合专家共识、因果敏感性、抗欺骗性及稳定性要求的奖励信号;其次,借鉴SCALER方法,设计参数化生成的合成设计环境,通过语法树扰动算子与约束采样器生成具有难度阶梯特征的页面实例,并采用自适应课程学习与隔离规则防止奖励过拟合;最后,使用40亿参数的代码模型在多轮交互中执行代理式强化学习,自主编辑代码库并读取自身渲染结果的文本报告,由隐藏评分者计算奖励。实验表明,该方法在TD-Bench上使“设计品味指数”(Taste Index)在家族内提升至0.795,跨家族测试提升至0.700,且在盲测专家对比中分别以0.600和0.569胜过基线模型与单轮控制组。研究发现,性能优势主要依赖于推理时的迭代反馈循环,但跨类别独特性略有下降,未完全达成既定目标。

链接: https://arxiv.org/abs/2611.20612
作者: Sujay choubey,Siddhant Saxena,Sarthak Kapil,Suhas Motwani
机构: 未知
类目: Multiagent Systems (cs.MA); Human-Computer Interaction (cs.HC)
备注: 15 pages, 9 figures

点击查看摘要

Abstract:Code-generating language models produce user interfaces that build and run, but the pages are often weak in hierarchy, spacing, typography and composition. We study whether a small code model can learn design taste through reinforcement learning in verifiable synthetic environments. The method has three components. First, a layered verification framework splits design quality into hard gates, render-grounded metrics, rubric judges, within-group relative preference and a blinded expert panel. A signal enters the reward only if it passes an admission protocol that tests agreement with experts, causal sensitivity, resistance to hack traps and stability. Second, following SCALER [1], synthetic design environments are parameterised generators with measured difficulty ladders: base pages are written once by an agentic coder, and instances are produced by syntax-tree perturbation operators and constraint samplers, scheduled by an adaptive curriculum with a quarantine rule against reward over-optimisation. Third, a 4B code model is trained with agentic reinforcement learning over multi-turn episodes in which it edits and builds a repository and reads a textual report of its own render, while hidden graders compute the reward. The admission protocol admits 7 of 16 render signals and 6 of 10 rubric dimensions. On TD-Bench, agentic training raises the Taste Index from 0.630 to 0.795 in-family and from 0.598 to 0.700 on held-out families, and wins 0.600 of blinded expert comparisons against the base model and 0.569 against a single-turn control. Most of the advantage depends on the inference-time loop, and cross-brief distinctiveness falls slightly below that of the base model, the only success criterion not met.

[MA-12] Who Observes Whom: Network Rationalizability and Its Complexity

【速读】:该论文旨在解决在信息部分且不对称的观测网络(observation network)背景下,如何定义和分析博弈中策略选择的合理解集问题。传统上,纳什均衡(Nash equilibrium)假设玩家完全知晓对手的策略,而理性化(rationalizability)则仅依赖于共同知识的理性假设,二者处于信息极端。当玩家间观测关系由有向图 $ G $ 描述(边 $ i \to j $ 表示玩家 $ i $ 观测到玩家 $ j $ 的策略),传统的解概念不再适用。为此,作者提出“自维持”(self-sustaining)这一新解概念:一个策略组合集合若满足其中每个策略组合对所有玩家而言,均可通过来自同一集合的推测(conjecture)得到合理解释,并与玩家实际观测一致,则该集合为自维持的。其解决方案的关键在于引入两种类型的推测——点推测(point conjectures,即单一策略组合)和信念推测(belief conjectures,即策略组合上的概率分布),并研究它们在不同网络拓扑下的行为特性。结果表明,随着观测边的增加,解集单调缩小;在无边网络下,解可多项式时间求解;而在存在单边观测时,问题变为 PPAD\mathsf{PPAD}-难,且在单向两玩家监控情形下,被观测方的策略恰好为其纳什均衡策略。对于层级型网络(如传递性锦标赛),成员判定可通过递归简化至子博弈,验证可在 ∃R\exists\mathbb{R} 中完成,且在双人情形下属于 P\mathsf{P};而在有向三元环结构中,迭代消除过程可能永不终止,且点推测下某些内部预测仅能通过无限见证集表示,而信念推测则可避免此类病态现象。最后,当仅可观测实际纯策略行动时,解集非空当且仅当网络无边,且此时可在 $ O(n|A|^2) $ 时间内计算得出。

链接: https://arxiv.org/abs/2611.20110
作者: Mojtaba Fayaz-Bakhsh,MohammadAmin Fazli
机构: 未知
类目: Computer Science and Game Theory (cs.GT); Computational Complexity (cs.CC); Multiagent Systems (cs.MA); Theoretical Economics (econ.TH)
备注:

点击查看摘要

Abstract:Nash equilibrium and rationalizability sit at opposite informational extremes: the former presumes that each player knows all opponents’ strategies, the latter that players reason from common knowledge of rationality alone. When visibility is partial and asymmetric, we study the solution concepts induced by a directed \emphobservation network G in which an edge i\to j means that i observes j 's strategy. A set of profiles is \emphself-sustaining if every profile in it can be justified, for every player, by conjectures from the same set that agree with everything the player observes. We consider \emphpoint conjectures (single profiles) and \emphbelief conjectures (distributions over profiles). Both range from Nash equilibrium on the complete network to point- and correlated rationalizability on the empty network, and shrink monotonically as links are added. Network topology governs both behavior and complexity. A predicted profile can be found in polynomial time on the empty network, whereas a single link makes search \mathsfPPAD -hard: under one-sided two-player monitoring, the observed player’s strategies are exactly her Nash equilibrium strategies. On hierarchical networks (transitive tournaments), membership reduces recursively to smaller games and admits polynomial-size certificates, placing verification in \exists\mathbbR , and in \mathsfP for two players. On the directed three-cycle, however, iterated elimination may never terminate under either variant, and under point conjectures interior predictions may admit only infinite witness sets, a pathology that belief conjectures avoid in the same game. Finally, when only realized pure actions are observed, the solution set is guaranteed to be non-empty exactly when the network has no edges, and is computable in time O(n|A|^2) . Subjects: Computer Science and Game Theory (cs.GT); Computational Complexity (cs.CC); Multiagent Systems (cs.MA); Theoretical Economics (econ.TH) Cite as: arXiv:2611.20110 [cs.GT] (or arXiv:2611.20110v1 [cs.GT] for this version) https://doi.org/10.48550/arXiv.2611.20110 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[MA-13] Infrastructure-Based Multi-Agent Reinforcement Learning with Attention for Highway Ramp Merging

【速读】:该论文旨在解决高速公路上匝道合流(on-ramp merging)场景中因缺乏有效协同而导致的安全性下降与通行效率降低的问题。在密集交通条件下,传统方法或仅依赖局部邻近车辆信息,导致决策局部最优;或采用全局交通状态观测,引入大量无关信息,影响决策效率。针对这一挑战,本文提出一种基于Transformer的多智能体强化学习(MARL)车辆控制模型,其核心创新在于引入一种基于注意力机制的观测编码器,使每辆联网自动驾驶汽车(CAV)能够从由路侧基础设施通过车-路通信(V2I)提供的全局交通状态中,动态筛选出对自身决策具有相关性的车辆信息。该方案实现了全局信息感知与局部信息聚焦之间的平衡。实验结果表明,所提模型在SUMO仿真中显著优于三种基线方法:仅观测邻车、全量信息编码及规则式SUMO模型,在交通流量、安全性与行驶平稳性方面均表现最佳;在四小时动态需求仿真中,该模型成功避免了拥堵,而其他模型在高密度下失效。此外,模型在存在人类驾驶车辆及真实V2I通信误差的情况下仍能保持无碰撞和高吞吐性能,且可视化分析揭示了智能体在不同交通情境下对其他车辆的相关性权重分配机制。

链接: https://arxiv.org/abs/2611.20078
作者: Chenhao Tong,Steffen Müller
机构: 未知
类目: Machine Learning (cs.LG); Multiagent Systems (cs.MA); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:On-ramp merging is a typical safety and throughput problem on highways. Without proper coordination, human drivers and self-interested autonomous vehicles may make risky manoeuvres and cause accidents or congestion. With recent advances in connected autonomous vehicles (CAVs) and multi-agent reinforcement learning (MARL), vehicles can be trained to coordinate for safe and efficient merging. However, in dense traffic, forming an effective observation of the traffic state is challenging. For example, observing only the neighbouring vehicles may lead to locally optimal decisions, while observing the entire traffic can introduce much irrelevant information. In this paper, we propose a transformer-based MARL vehicle control model to address these problems. A roadside infrastructure controller receives the states of all CAVs via vehicle-to-infrastructure (V2I) communication and sends control commands back to them. An attention-based observation encoder is designed where each vehicle can filter the relevant vehicle information from the global traffic state. The proposed model is evaluated in SUMO against three baselines: a model that observes only the neighbouring vehicles, a model that encodes the global traffic information, and the rule-based SUMO model. The results show that our model achieves the highest traffic flow, safety and smoothness. Furthermore, in a four-hour simulation with dynamic traffic demand, our model successfully avoids traffic jams, while other baseline models fail at high traffic flow. We also show that our model remains effective when human-driven vehicles are present. We further show that our model, trained in ideal communication conditions, can still remain collision-free and maintain throughput under realistic V2I communication errors. Finally, we visualise the relevance weights that an ego vehicle assigns to other vehicles in multiple traffic situations.

[MA-14] World-Model-Guided Plan Search for Cooperative Mixed-Traffic Control at Bottlenecks

【速读】:该论文旨在解决在车道缩减瓶颈处,通过联网与自动化车辆(Connected and Automated Vehicles, CAVs)进行协作式车速控制以缓解交通拥堵的问题,尤其关注车辆角色分配、执行时机及初始排队成本与后续通行效益之间的权衡。其解决方案的关键在于提出一种结合世界模型引导与仿真验证的离线规划搜索框架:利用融合物理动态特性的交通世界模型,指导对车辆角色、目标速度、等待时长及启动时间等参数的进化搜索;并通过与SUMO仿真器配对的筛选机制,评估候选方案在总延误降低和各车辆组平均延误增量(不超过10秒)方面的表现。最终生成可复用的规划库,其动作可临时覆盖冻结的MAPPO策略,实现高效协同控制。实验结果表明,在60%渗透率、每小时2800辆车的交通需求下,该方法相较MAPPO策略减少平均总延误14.3%,优于人工编写规划库的3.8%改善,且最大组平均延误增加仅6.2秒,低于预设阈值。此外,基于世界模型引导的搜索在独立筛选中成功率达30.0%,远高于随机采样的2.6%,验证了预测性搜索与独立验证在满足群体延迟约束条件下发现有效协同策略的有效性。

链接: https://arxiv.org/abs/2611.20067
作者: Jinghan Peng,Ning Liu,Xiangrui Kong,Sen Shen,Dongkun Han
机构: University of California, Los Angeles (加州大学洛杉矶分校); The University of Western Australia (西澳大学); The Chinese University of Hong Kong (香港中文大学)
类目: Multiagent Systems (cs.MA)
备注: 8 pages

点击查看摘要

Abstract:Cooperative speed control by connected and automated vehicles can reduce congestion at lane drop bottlenecks. However, effective coordination depends on vehicle roles and execution timing, while initial holding costs may be offset by later traffic benefits. This paper proposes an offline plan search framework that combines world model guidance with simulator verification. A traffic world model incorporating physical dynamics guides evolutionary search over vehicle roles, target speeds, holding durations, and start times. Paired SUMO simulations screen candidates for total delay reduction and limits on additional mean delay across vehicle groups. Accepted plans form a reusable library whose actions temporarily override a frozen MAPPO policy. Across 192 paired test scenarios at 60% penetration and a demand of 2,800 vehicles per hour, the searched library reduces mean total delay by 14.3% relative to MAPPO and by 3.8% relative to a handwritten plan library. The largest increase in group mean delay is 6.2 s, below the prescribed 10 s limit. In a separate screening experiment, the guided search achieves a 30.0% pass rate compared with 2.6% for random sampling. These results support the use of predictive search and independent verification for discovering cooperative plans under group delay constraints.

[MA-15] Autonomous Multi-modal Agents for ITOps Observability

【速读】:该论文旨在解决IT运维可观测性(ITOps observability)工具与系统间数据孤岛问题,即如何高效集成多源异构数据以支持统一的可观测性分析。其核心解决方案是提出一种多模态多智能体系统——ITOps Observability Studio,通过协作式AI智能体实现可观测性工作流的自动化编排。该系统的创新关键在于三层次架构设计:第一层为智能体间的协同通信机制;第二层为基于大语言模型(LLM)的专用智能体,分别负责数据访问、模式操作(schema operations)与可视化生成;第三层为统一的工具集成接口。系统实现了自然语言到GraphQL的转换、自动化的GraphQL模式生成与映射、链路追踪数据全生命周期管理以及跨工具的可靠性工程工作流支持。此外,所有智能体均内嵌智能体分析(agent analytics)功能,可追踪模型调用与工具调用行为,从而支持可解释性与故障诊断。评估表明,当前基于开源模型的查询生成与路由性能良好,数据生命周期管理智能体在多数场景下表现有效;同时,联合测试揭示了仅在提示词(prompt)、解析器与服务栈共同作用时才会显现的缺陷,并提供了相应的修复方案。

链接: https://arxiv.org/abs/2611.19982
作者: Gunjan Siddharth,Sanjana Garai,Utsav Shekhar,Prakhar Jain,Sarvam Maheshwari,Balaji Ganesan,Avirup Saha,Seema Nagar,Pratik Mishra,Felix George,Suranjana Samanta,Harshit Kumar,Mudit Verma,Prateeti Mohapatra,Ravi Jagannathan,Amit Paradkar,Sameep Mehta,Pratibha Moogi
机构: IBM Software(IBM软件)
类目: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
备注: 10 pages, 8 figures

点击查看摘要

Abstract:ITOps observability products must integrate data from different tools and systems. We present ITOps Observability Studio, a multi-modal multi-agent system that orchestrates observability workflows through collaborative AI agents. Its three-tier architecture features inter-agent communication, specialized LLM-powered agents for data access, schema operations and visualization, and unified tool integration. The system supports natural language to GraphQL translation, GraphQL schema generation, schema mapping, trace data lifecycle management and multi-tool workflows for reliability engineering. Its agents are instrumented with agent analytics, so that their model and tool calls can be traced. We evaluate each agent on its own benchmark, scored against gold answers where they exist and by a model judge for schema mapping and chart quality. Query generation on a public benchmark and routing work well with a current open model, and the data lifecycle management agent solves most of its scenarios. The evaluation exposed defects that only appear when prompts, parsers and serving stacks are tested together, and we report them together with their fixes.

[MA-16] Cooperative Transport of Unknown Objects by Distributed Effort Regulation in Multi-agent Systems ICRA2027

【速读】:该论文旨在解决未知形状、尺寸、几何结构、质量及其质量分布的物体在无先验知识(包括物体特性及参与搬运的机器人数量)条件下,由多机器人协同完成运输任务时所面临的挑战。其核心问题在于如何在缺乏精确物体模型的情况下实现稳定、高效且鲁棒的协同推运控制。解决方案的关键在于提出一种无需物体模型的航向误差驱动推力调节方法(object-model-free heading-error-driven pushing effort regulation),通过机器人基于个体与集体航向误差动态调整推力,以促进目标导向的物体运动;同时,当推力点丢失或单个机器人无法通过自身推力纠正航向时,机器人可自主重新定位至物体边界,确保任务连续性。该方法在多种复杂形状(含凸与非凸边界)及动态质量分布场景下验证有效,且表现出对摩擦不均和轮滑等外部扰动的高度鲁棒性,路径效率与运输时间表现稳定,具备良好的可扩展性,运输时间主要取决于物体表面可用推力面数量,而该数量由物体尺寸、形状及机器人间安全距离共同决定。

链接: https://arxiv.org/abs/2611.19731
作者: Tamzidul Mina,Shyam S. Kannan,Wonse Jo,Shaocheng Luo,Galen B. King,Byung-Cheol Min
机构: Purdue University (普渡大学); Sandia National Laboratories (桑迪亚国家实验室); The Boeing Company (波音公司); Incheon National University (仁川国立大学); Duke University (杜克大学); Indiana University (印第安纳大学)
类目: Robotics (cs.RO); Multiagent Systems (cs.MA)
备注: 9 pages, 12 figures; supplementary video available at this https URL submitted to ICRA 2027

点击查看摘要

Abstract:A distributed multi-robot object transportation strategy is proposed for objects of unknown shape, size, geometry, mass and mass distribution without any prior knowledge about the object or number of robots involved in the transportation process. Robots follow the boundary of the object and dynamically determine viable pushing points. A novel object-model-free heading-error-driven pushing effort regulation method is proposed to promote goal-directed object motion; robots regulate their pushing effort using individual and collective heading errors. Robots dynamically relocate around the object boundary if a pushing point is lost or the required pushing effort exceeds its capability to correct its heading. The proposed method is validated using simple and complex shapes with convex and non-convex boundary regions and dynamically changing weight distribution of objects. Detailed analysis on the effects of robot distribution and initial object pose on the transportation performance in terms of required time and path efficiency is also presented. Scalability analysis revealed that object transportation time is dependent on the number of pushing surfaces available which in turn depends on the object size, shape, and set inter-robot safety distance. The proposed system yielded highly efficient object paths with consistent elapsed transport times over several trials robust to external disturbances such as non-uniform friction between contacting surfaces and robot wheel slip.

[MA-17] Real-Time Air Pollution Monitoring and Tracking via Adaptive Density-Driven Multi-UAV Coordination

【速读】:该论文旨在解决动态演变的大气污染物分布实时监测与追踪中,传统方法在复杂风场条件下难以保持高精度、自适应性不足的问题。其核心挑战在于如何实现多无人飞行器(UAVs)在非静态、时变风场下的高效协同感知与轨迹规划,以维持对污染云团的精确时空表征。解决方案的关键在于提出一种自适应密度驱动控制(D²C)框架,通过融合预测性污染物传播模型与动态更新的参考密度分布,引导UAV群的智能路径规划。该框架引入基于传感器的代理烟羽更新机制,利用局部观测的浓度梯度与风场估计,结合梯度吸引、风向平流、样本间排斥及随机扩散等物理过程,实现参考样本的动态演化;同时采用集中式协调结构,各UAV共享局部测量数据至中央协调器,统一维护参考样本集,从而支持系统性的目标分配与轨迹优化。仿真结果表明,在强风和复杂风场条件下,该方法显著优于现有先进基准,在Sinkhorn距离指标上表现更优,尤其在污染云团超出固定监测区域时仍能保持有效追踪,验证了D²C框架在动态环境下的高效性与鲁棒性。

链接: https://arxiv.org/abs/2611.19703
作者: Julian Martinez,Ruchika Singh,Mohammad Afrazi,Kooktae Lee
机构: Texas Tech University (德州理工大学); Duke University (杜克大学)
类目: Robotics (cs.RO); Multiagent Systems (cs.MA); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:Accurate and timely air pollution monitoring and tracking are critical for protecting public health and supporting effective environmental management. This paper presents an adaptive Density-Driven Control (D ^2 C) framework for coordinating multiple unmanned aerial vehicles (UAVs) to track and map dynamically evolving pollutant distributions. The proposed framework extends the D ^2 C strategy to guide UAV trajectories based on predictive pollutant propagation and dynamically updated reference densities. A sensor-driven surrogate plume update mechanism is employed, in which reference samples are propagated using locally sensed concentration gradients and wind estimates through gradient attraction, wind advection, inter-sample repulsion, and stochastic diffusion, enabling UAVs to maintain an accurate spatio-temporal representation of the pollution field. The framework utilizes a coordinated structure where individual UAVs share their local measurements with a central coordinator to maintain a unified set of reference samples, facilitating systematic target allocation and trajectory planning. Simulation results under both mild and strong wind conditions demonstrate that the proposed method significantly outperforms state-of-the-art baselines in terms of Sinkhorn distance. Specifically, in scenarios where the plume drifts beyond a fixed monitoring domain under high-velocity, time-varying wind fields, the tested baseline methods fail to maintain tracking, whereas the proposed D ^2 C framework successfully adapts to the evolving distribution. These findings highlight D ^2 C as an efficient and resilient solution for atmospheric monitoring in dynamic environments.

[MA-18] Protecting the System Stranding the User: The Limits of Multi-Agent Authorization

【速读】:该论文旨在解决当前多智能体授权架构过度侧重系统完整性而忽视人类实际影响的问题。现有模型虽能有效防止权限提升,但将失败代价转嫁给人工操作员,导致人机协同效率与用户体验受损。其核心解决方案的关键在于重构授权机制,使其以人类影响为中心:首先,摒弃对用户意图“验证”的伪精确表述,转而采用开放式的概率估计,并根据操作风险动态调整防御强度;其次,指出动态信任评分在跨组织智能体交接中无法维持可信性,因缺乏底层证据传递易违背情境完整性(Contextual Integrity)原则;第三,强调安全中止的工作流不应被视为成功拒绝,若导致用户陷入困境,则补偿措施必须与能力令牌直接绑定,而非仅依赖规划层;最后,提出无摩擦自动化恢复可能引发道德风险,削弱开发者构建稳健初始意图校验的动力。论文最终提出一种融合确定性安全与人类结果的架构整合方案,同时揭示了仍存在的根本理论局限。

链接: https://arxiv.org/abs/2611.19598
作者: Gurpriya Bhatia,Clifford Neuman
机构: University of Southern California(南加州大学); Los Angeles, CA
类目: Cryptography and Security (cs.CR); Multiagent Systems (cs.MA)
备注: 11 pages. Accepted as a poster at HAIPS: 2nd Workshop on Human-Centered Privacy and Security for Language Models, held in conjunction with COLM 2026

点击查看摘要

Abstract:Current multi-agent authorization architectures prioritize system integrity over human outcomes. While recent models successfully prevent privilege escalation, they externalize the cost of failure onto the human operator. Re-centering authorization around human impact exposes four structural vulnerabilities in current paradigms. First, calling an agent’s understanding of user intent verified is misleading. Intent has no formal definition, so no check against it can be definitive. Systems should instead report an openly probabilistic estimate and scale their caution with the risk of the action. Second, we argue that dynamic trust scores do not survive multi-agent handoffs; transmitting algorithmic judgment across a company boundary without underlying evidence risks violating Contextual Integrity. Third, treating a safely halted workflow as a successful refusal is a category error if it strands the user; compensating actions must be bound directly to the capability token, not only to the planning layer. Finally, we hypothesize that frictionless automated recovery introduces a moral hazard, disincentivizing developers from building robust initial intent checks. We conclude by proposing an architectural synthesis to bridge the gap between deterministic security and human outcomes, alongside the fundamental theoretical limits that remain unresolved.

[MA-19] Behavioral Assurance for Socially Deployed AI Agents

【速读】:该论文旨在解决当前社会部署的AI代理(AI agents)在多样化交互情境中行为一致性不足的问题,即现有评估方法难以确保其行为始终与其预期特征保持一致。其解决方案的关键在于提出一个可量化的评估框架,通过五个可测量的行为维度,结合涵盖六种社会敏感交互场景的标准测试集,系统评估AI代理的行为分化性、一致性及行为漂移情况。该框架引入受控的“角色配置”作为实验锚点,能够有效区分意图性行为差异与仅表现为风格化变异的情况,从而实现对行为意图的真实验证。初步分析表明不同配置间存在可测量的行为差异,并揭示了行为特征与交互结果之间的潜在关联,推动评估范式从单纯的角色设计转向行为保障,为社会应用中AI代理是否持续、可度量地按预期表现提供了科学依据。

链接: https://arxiv.org/abs/2611.19564
作者: Nicole H Brennan,Jiaee Cheong
机构: 未知
类目: Artificial Intelligence (cs.AI); Emerging Technologies (cs.ET); Human-Computer Interaction (cs.HC); Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:AI agents are increasingly deployed in socially consequential settings, yet existing evaluations provide limited assurance that they behave consistently with their intended characteristics across diverse interactions. We introduce an evaluation framework for measuring the behavioral consistency of socially deployed AI agents. The framework operationalizes agent behavior along five measurable axes and uses a standardized evaluation battery spanning six socially salient interaction settings to assess behavioral differentiation, consistency, and drift. Controlled persona configurations serve as experimental anchors, enabling systematic testing of whether intended behavioral differences persist across contexts rather than appearing only as stylistic variation. Preliminary analyses demonstrate measurable differentiation across configurations and suggest associations between behavioral profiles and interaction outcomes. By shifting the focus from persona design to behavioral assurance, this work provides a foundation for evaluating whether socially deployed AI agents behave as intended, consistently and measurably, across the situations in which users encounter them.

[MA-20] LLM Trading Agents Earn Beta Not Alpha: A Look-Ahead-Free Benchmark

【速读】:该论文旨在解决当前大型语言模型(Large Language Model, LLM)交易代理在评估其真实交易能力时面临的“技能与运气难以区分”的核心问题。现有基准测试通常依赖单次回测结果来衡量模型表现,但研究表明,同一模型仅因随机种子不同,其回测收益差异可达平均5.2个点,表明单纯依赖收益无法有效区分模型的真实预测能力与偶然性收益。此外,传统回测方法存在严重偏差:要么使用模型训练截止后短暂的市场数据,要么在长时间历史数据上回测,而这些数据可能已被模型记忆;且常将新闻事件与价格按日历日期对齐,导致信息泄露(look-ahead bias)。为克服上述缺陷,本文提出rta-bench-1,一个覆盖美国三个典型市场阶段(牛市、熊市、崩盘)的严格回测框架。该框架确保每条新闻仅在实际发布时间后才被引入,每个交易周期均以盲态和匿名方式处理,并对所有运行进行前瞻性检查以杜绝信息泄露。同时,该框架将收益分解为因子暴露(factor exposure)与超额收益(alpha),并以同日同行业随机选股作为基准,评估每项决策的实际价值。在对六种来自四个供应商的模型进行91次运行的测试中,未发现显著的alpha,市场与风格因子解释了中位数78%的收益,且模型决策表现不优于同行业随机选股。尽管代理倾向于买入正面新闻最多的股票,但在消融实验中,即使移除新闻输入,其收益仍保持不变,说明其行为更多源于规则执行而非新闻驱动。然而,研究发现,若在提示中加入单一现金持有规则,基于GPT的代理在所有六组对比中其持股比例下降30.3个百分点,表明其能可靠执行明确指令。因此,论文的核心解决方案在于:构建基于发布时序的严格回测环境,强调开发者应主动提供可解释的因子暴露与超额收益指标,而评估基准必须按新闻实际发布时间重放数据,以实现对交易代理真实能力的可信评估。

链接: https://arxiv.org/abs/2611.19530
作者: Jackson Wang
机构: 未知
类目: Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
备注: 23 pages, 10 figures, 15 tables. Code, frozen markets and leaderboard: this https URL . Original empirical study: a backtest benchmark for LLM trading agents and measurements from 91 runs of six models

点击查看摘要

Abstract:Large language model (LLM) trading agents now trade stocks for their users, and benchmarks rank models by their agents’ returns. Yet a single run’s return says little about skill. In our study, two runs of the same model that differed only in the random seed ended 5.2 points apart on average, so telling skill from luck takes many backtests across markets, models and seeds. However, live arenas trade each day only once, and backtests either use the few months after a model’s training cutoff or run on years the model may remember, often with news joined to prices by calendar date. In this paper, we present rta-bench-1, a backtest over three US markets (a rally, a bear market and a crash). It shows each headline only after its publication minute, plays each period blind and named, and checks every run for look-ahead. It also splits each return into factor exposure and alpha and scores every decision against random choices of the same day and sector. In 91 runs of six models from four vendors, no alpha was significant, market and style exposure accounted for a median 78% of the return, and the decisions did no better than random stocks of the same sector. The agents bought the stocks with the most positive news, yet in ablations they earned as much without their headlines. Still, the agents carried out rules reliably, since a single cash rule in the prompt cut a GPT-based agent’s share in stocks by 30.3 points in all six pairs. Builders should therefore supply the edge themselves and report exposures and alpha, and benchmarks should replay news by publication time.

[MA-21] Delay-Aware Reinforcement Learning for O-RAN Control under Stochastic Reward Feedback

【速读】:该论文旨在解决开放无线接入网(O-RAN)中因反馈异步性导致的动作与奖励之间存在随机延迟,进而引发信用分配错位和训练不稳定的挑战。其核心问题在于传统强化学习(Reinforcement Learning, RL)方法在处理具有未知延迟的观测奖励时,无法有效对齐动作与实际回报,从而影响控制策略的收敛性和可靠性。解决方案的关键是提出一种面向延迟的深度强化学习框架,通过在离策略学习范式中集成基于长短期记忆网络(LSTM)的延迟预测器与时间奖励重对齐模块,实现对延迟的建模与补偿。该框架利用来自Open AI Cellular(OAIC)测试床的真实延迟数据进行验证,结果表明,相较于忽略反馈延迟的传统RL方法,所提方案不仅加速了收敛过程,还显著降低了延迟违规率和包丢失率,充分证明了奖励-延迟补偿机制在实现可靠生成式AI驱动的O-RAN控制中的关键作用。

链接: https://arxiv.org/abs/2611.19344
作者: Xingqi Wu,Junaid Farooq,Tao Li,Juntao Chen,Ying Wang
机构: 未知
类目: Multiagent Systems (cs.MA); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Reinforcement learning (RL) is increasingly employed for real-time control in Open Radio Access Networks (O-RANs). However, the asynchronous nature of O-RAN feedback introduces stochastic delays between actions and their observable rewards, leading to misaligned credit assignment and unstable training. This letter formulates the O-RAN control problem as a Markov decision process (MDP) with delayed and stochastic rewards, where the true reward corresponding to an action is received after an unknown delay. To address this, we propose a delay-aware deep RL framework that integrates a long short-term memory (LSTM)-based delay predictor and a temporal reward realignment module within an off-policy learning pipeline. Using empirical delay traces collected from the Open AI Cellular (OAIC) testbed, we demonstrate that the proposed framework achieves faster convergence and significantly reduces latency violations and packet drops compared to conventional RL methods that ignore feedback delay. The results confirm that reward-delay compensation is critical for reliable AI-driven O-RAN control.

[MA-22] Macro Control and Capability Evaluation of Black-Box World Models: A Two-Regime Framework for LLM -Agent Markets

【速读】:该论文旨在解决大语言模型(Large Language Model, LLM)代理在多代理市场中难以分离评估“信息模拟”与“宏观涌现”两个关键阶段的问题,传统单一综合评分无法有效区分这两者。其核心解决方案在于提出一种软宏观调控方法(soft macro-control),通过每日公布公开参考路径(reference path)而不干预个体决策、订单或匹配机制,实现对市场行为的引导;同时设计了双模式评估框架,仅基于宏观观测数据即可分别度量两个阶段:在仅参考路径模式下,通过对称真实与反向参考路径共享的水平值 $ R $ 衡量实现程度,其配对差距 $ B $ 反映结构性偏差;在信息注入模式下,外部信息稀释参考路径,仅真实-反向差距 $ \Delta $ 具有可解释性。实验结果表明,在五个单指数LLM代理市场中,发布目标路径显著提升了实际路径对目标的跟踪能力,在无新闻条件下跟踪得分分别提升0.306和0.415,有新闻时提升0.183和0.329;最优结构配置为70%现金/30%股票组合($ B=0.042 ),技术权重为3:7时实现程度最高(),技术权重为3:7时实现程度最高( R=0.894 $),且引入基本面数据和新闻后 $ \Delta $ 分别为0.0341和0.0487。该方法通过同一组可观测指标在不同可解释性场景下的切换使用,实现了对信息模拟与宏观涌现两阶段的独立、精准评估。

链接: https://arxiv.org/abs/2611.19311
作者: Zhiyang Niu,Chuanjie Li,Huayuan Shen,Mengmeng Zhao,Jiajia Zhang
机构: Harbin Institute of Technology, Shenzhen, China(哈尔滨工业大学深圳); Nanyang Technological University, Singapore(南洋理工大学); Pingan Group, China(平安集团)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG); Multiagent Systems (cs.MA)
备注: 19 pages; 6 figures, 17 tables

点击查看摘要

Abstract:Large language model (LLM) agents can be organized into multi-agent markets that act as world models, mapping macro information to emergent macro dynamics while the evaluator observes only macro inputs and outputs. This paper separates a model’s capability into two stages that a single aggregate score cannot separate: information simulation, from macro information to micro behavior, and macro emergence, from micro behavior to the macro outcome. We introduce a soft macro-control method that announces a public reference path day by day without touching individual decisions, orders, or matching, and a two-regime evaluation framework that measures both stages from macro observations alone. In the reference-only regime the symmetric true and flipped references share a common level R, which measures realization under the announced reference, while their paired gap B measures structural bias; in the information-injection regime external information dilutes the reference, so only the true-flipped gap Delta remains interpretable. On five single-index LLM-agent markets, announcing a target path shifts the realized path toward the announced target on both announced references under both information conditions. Relative to the no-reference path, the tracking score is higher by 0.306 and 0.415 without news and by 0.183 and 0.329 with news; within the stock/cash scan structural bias is lowest at 70/30 (B=0.042); realization peaks at a 3:7 technical share (R=0.894 versus 0.358 for technical-only); and Delta is 0.0341 with fundamental data and 0.0487 with news. Both stages use the same two observable scores, each read in the regime where it is interpretable.

[MA-23] Staged Semantic Distillation: Preserving Free-Form Semantics in a Low-Compute Non-LLM Surrogate of a Staged LLM Agent

【速读】:该论文旨在解决大规模基于分阶段大语言模型(LLM)的市场与社会模拟器在实际运行中计算开销过高、推理效率低下且难以部署的问题。具体而言,传统低计算量的蒸馏方法会丢失教师模型生成的非结构化中间状态语义信息,导致学生模型无法有效继承关键决策逻辑。其解决方案的关键在于提出一种名为“分阶段语义蒸馏”(staged semantic distillation)的新方法:通过冻结编码器将教师模型输出映射为1024维向量,并以带阶段标签的类型化向量形式输入给非LLM学生模型;学生模型在特定阶段可见性约束下接受联合监督,并采用调度自历史(scheduled self-history)进行训练。该方法仅需一次记录的模拟轨迹即可训练出一个1320万参数的确定性学生模型,在单块8GB GPU上训练仅需1658秒,推理时无需调用教师模型或解码器,无文本生成开销(每步决策耗时0.14毫秒),且其分析结果与教师模型保持高度一致(余弦相似度0.943)。实验表明,该学生模型在教师强制回放任务中达到0.8347的平衡准确率,在部署解码器上为0.7783,显著优于无语义重训基线(五种子集增益+2.2 [1.21, 3.14])。进一步分析揭示,调度自历史机制是性能提升的核心因素,而内容本身在教师强制操作点上才具有决定性作用。研究验证了“应优先投资部署调度设计,让内容在教师强制运行点自然获得其位置”的设计原则,实现了20万次决策的20种子集扫描仅需12分钟,相较教师模型的三小时大幅降低计算成本。

链接: https://arxiv.org/abs/2611.19299
作者: Zhiyang Niu,Huayuan Shen,Chuanjie Li,Mengmeng Zhao,Jiajia Zhang
机构: Harbin Institute of Technology, Shenzhen, China(哈尔滨工业大学深圳校区); Nanyang Technological University, Singapore(南洋理工大学); Pingan Group, China(平安集团)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG); Multiagent Systems (cs.MA)
备注: Main text 7 pages; 18 pages total; 3 figures, 16 tables

点击查看摘要

Abstract:Staged LLM agents make market and social simulators realistic, but one 102-day run of a 200-agent, ten-index market simulator issues 203,100 language-model calls and returns one realization. Conventional low-compute distillation drops their free-form intermediate state. We present staged semantic distillation, a low-compute method that preserves non-structured semantics in a non-LLM student trained from one recorded trace. A frozen encoder maps each teacher output to a 1024-d vector; the student sees typed stage-tagged vectors with stage-specific visibility, is jointly supervised, and trains with scheduled self-history. One TwinMarket trace yields a 13.2M-parameter deterministic student that trains in 1,658 seconds on one 8-GB GPU and runs with zero teacher or decoder calls and no text generation (0.14 ms per decision). Its analysis stays close to the teacher (cosine 0.943), and swapping semantic inputs between agents moves 64.7% of decisions against 12.8% for a matched donor, so the channel is preserved and live. We use the surrogate to study what the transfer carries. It reaches 0.8347 balanced accuracy on the teacher-forced replay and 0.7783 [0.766,0.791] on the deployment decoder; the five-seed gain over a no-semantic retrain is +2.2 [1.21,3.14]. Against the teacher’s own re-sampling level, round-one R is 97.1% [95.8,98.4] and 92.3% [89.0,95.4]; R is a diagnostic, not an accuracy. A five-seed content-by-schedule study isolates the gain: scheduled self-history carries it (+2.2 [1.21,3.14]); shuffled and constant content controls reproduce it, while content matters teacher-forced. The design rule: invest in the deployment schedule, let content earn its place at the teacher-forced operating point. A 20-seed, 200,000-decision sweep takes 12 minutes with zero LLM calls, versus three hours for the teacher. The paper is a method-and-case-study contribution on one simulator and trace.

[MA-24] When Reviews Gate Access: Collusion Sabotage and the Scale of Peer Review in the AI Submission Surge

【速读】:该论文旨在解决大规模机器学习顶会(如NeurIPS、ICLR)在面对投稿量激增背景下,为保障评审质量而引入的复杂评审规则所引发的新型博弈风险问题。这些规则(如作者投稿数量限制、强制互评、基于录用论文获取评审资格、基于评审意愿分配等)虽意在提升公平性与参与度,却无意中使评审者与作者之间形成竞争关系,从而催生了“恶意行为”——包括合谋(collusion)与破坏(sabotage)。其核心解决方案的关键在于:通过构建基于代理(agent-based)的会议模拟系统,结合真实会议统计数据,量化分析上述两类行为在不同规则下的传播机制与影响程度。研究发现,合谋主要改变录用结果,破坏则削弱评分对论文质量的真实反映,二者具有协同放大效应;随机分配机制下,合谋的影响受“盟友暴露指数” $ E = Rm/A $ 限制,但评审意愿投标机制可显著削弱此保护作用;此外,在无投稿量限制的情况下,社交网络中的内层约束(如内部圈层限制)能有效抑制合谋危害,而基于录用获得评审资格的制度若允许破坏行为,则会强化既有评审者的权力优势,进一步拉大新旧评审者之间的差距。因此,该研究揭示了现有评审制度设计中的深层漏洞,并指出制度设计需兼顾激励相容与抗操纵性。

链接: https://arxiv.org/abs/2611.19054
作者: Yuchen Li,Xintong Li,Ken-Tye Yong,Nguyen H. Tran
机构: University of Sydney(悉尼大学)
类目: Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:AI tools have helped push submissions to the largest machine-learning venues to tens of thousands per cycle. Venues have answered with rules that tie access to review: per-author caps, mandatory reciprocal reviewing, reviewer eligibility earned through accepted papers, and assignment based on reviewer bids. Under these rules a reviewer is also a competitor of the authors. We study two deviations that the rules make attractive. Sabotage lowers a competitor’s score. Collusion raises an ally’s score. We build an agent-based conference calibrated to public statistics of NeurIPS and ICLR and run 6,220 simulated conferences. Four results stand out. First, the two deviations do different harm and amplify each other. Collusion changes who is accepted, sabotage erodes what scores say about quality, and rings that may also sabotage gain more than rings that only collude. Second, under random assignment the chance that collusion changes a decision is set by one exposure index, E = Rm/A , the expected number of allies among a paper’s reviewers. Growth therefore protects random assignment, but bidding removes the protection. In a pool of 3,000 reviewers a 20-member ring turns 0.8% of its papers into acceptances under random assignment and 71% when 30% of allies’ bids are granted. Third, when agents choose their allies every cycle, caps at the inner layers of the social brain keep collusion nearly harmless. Without a cap agents ally with half of their area and accepted value falls by 35%. Fourth, reviewer eligibility tied to acceptances turns sabotage into a weapon of incumbents, and a longer eligibility window widens the gap between incumbents and newcomers only when sabotage is possible.

[MA-25] Swarm Shepherd: Securing Multi-Agent Ecosystems Against Persistent Latent Compromise

【速读】:该论文旨在解决生成式智能体生态系统(Agentic Ecosystems)中潜伏型系统级威胁——“Agentic Sleeper”所带来的安全挑战。此类威胁表现为攻击者在初期以合法行为伪装,逐步获取路由信任,随后在多轮会话、跨智能体及系统层之间累积破坏,最终引发可见性故障,而传统逐轮防御机制难以察觉此类渐进式渗透。其核心问题在于:现有防御手段缺乏对跨会话、跨智能体和跨层级的长期行为关联性监测能力。解决方案的关键在于提出名为Swarm Shepherd的系统级防御框架,该框架基于部分可观测环境下的推理机制,包含两个核心组件:一是通过推断过滤器实现早期预警,二是利用回溯追踪模块进行入侵源头定位。研究进一步提出了一个双尺度部署规则,平衡事件历史长度与智能体覆盖范围,在LangGraph、CrewAI和AutoGen等平台上的实验证明,Agentic Sleeper的漂移速度比自然基线快2.7–8.9倍,且在多数拓扑结构下可跨智能体传播,超过80%的逐轮与状态感知防御均被绕过。Swarm Shepherd可在可见性能退化前至少提前三个会话发出警告(任务评分器评估下为2.6个会话),回溯模块在24种变体上达到0.69的Top-1准确率和0.72的宏平均F1值。双尺度部署规则经实证验证有效,并在真实端到端运行中证明,基于后验引导的防御堆栈能将跨智能体传播抑制至自维持阈值以下,优于仅移除明显受损智能体的传统方法。

链接: https://arxiv.org/abs/2611.18693
作者: Wuyang Zhang,Shichao Pei
机构: University of Massachusetts Boston(马萨诸塞大学波士顿分校)
类目: Multiagent Systems (cs.MA); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR)
备注: The Fortieth Annual Conference on Neural Information Processing Systems

点击查看摘要

Abstract:Agentic ecosystems are emerging as a new platform layer in which agents maintain persistent memory, retrieve from shared knowledge bases, execute through tool interfaces, and coordinate through shared states. These capabilities create a security problem that per-turn defenses miss: compromise can accumulate across sessions, agents, and system layers before visible failure. We define this threat as Agentic Sleeper, a system-level adversary that first behaves legitimately, gains routing trust, and later activates across memory, agent logic, and infrastructure. We formulate defense as inference under partial observability and propose Swarm Shepherd, a system-level defense framework centered on an inference filter for early warning, paired with a backward trace-back module for entry attribution. Our analysis yields a two-scale deployment rule that balances longer event histories against broader agent coverage. Across LangGraph, CrewAI, and AutoGen, Agentic Sleeper drift begins 2.7–8.9 \times faster than natural baselines, propagates across agents on most topologies, and evades per-turn and stateful defenses above 80%. Swarm Shepherd warns at least three sessions before visible degradation (2.6 sessions when failure is scored by the benchmark’s own task grader), while trace-back reaches 0.69 top-1 accuracy and 0.72 macro-F1 over 24 variants. The two-scale deployment rule is empirically confirmed, and in live end-to-end runs on LangGraph and AutoGen the posterior-guided defense stack drives cross-agent propagation below the self-sustaining regime, which removing only the visibly compromised agent does not achieve.

[MA-26] Screen Before You Add: Reducing Candidate Inference in Model Ensembles

【速读】:该论文旨在解决动态集成模型(dynamic ensemble)在新增候选模型时的高效筛选问题,尤其针对推理成本高昂的大型语言模型(Large Language Models, LLMs)场景。现有方法依赖全量标注数据集上的完整推断来评估候选模型是否提升集成性能,但这一过程在计算上不可行,且仅依赖基准得分可能产生误导——强单体性能并不保证对集成有正向贡献,甚至优化后的加权策略可能损害测试表现。本文的核心解决方案是基于残差几何结构(residual geometry)实现安全的早期拒绝(early rejection),其关键在于:在凸聚合(convex aggregation)与平方损失或Brier损失下,整体最优解的改进当且仅当候选模型产生的预测变化与现有集成模型在总体样本上的残差呈正向对齐。为此,作者提出一个乐观的二次规划(optimistic quadratic program, QP)以提供可证明的早期拒绝保证;同时引入后验情景采样(posterior-scenario sampling)作为替代启发式方法。实验表明,在两个LLM面板上,采用分层筛选结合认证QP拒绝可节省21–23%的候选模型调用次数,且保持100%的筛选准确率;而结合后验采样进一步将调用节省提升至约37%,仅导致筛选准确率下降不足1个百分点,验证了该方法在效率与精度之间的良好权衡。

链接: https://arxiv.org/abs/2611.18647
作者: Davide Mambelli,Alessandro Rozza,Steven Motta
机构: distributed cognition
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA)
备注: 42 pages, 4 figures, 10 tables, including appendices

点击查看摘要

Abstract:As new models become available, existing ensembles must be updated to be competitive, making them intrinsically dynamic. Deciding whether to add a candidate model can require inference on an entire annotated dataset. Such a screening becomes prohibitive for models with costly inference especially when they ultimately add no value. Even under simple convex aggregation, relying on benchmark scores can be misleading as strong standalone performance does not guarantee ensemble improvement, and fitted ensemble weights can hurt test performance. In this work, we study candidate screening and show how early rejection can be achieved safely, exploiting residual geometry for stratification and rejecting the candidate when the best possible remaining predictions cannot achieve the necessary improvement. The key result that justifies this approach is that, under convex aggregation and squared or Brier loss, the population optimum improves if and only if a candidate induces a prediction change positively aligned with the residuals of the population-optimal existing ensemble. An optimistic quadratic program (QP) provides certified early rejection, while posterior-scenario sampling offers an alternative heuristic. We evaluate the approach by showing the tradeoff between saved calls and screening accuracy, which refers to recovering the same accept/reject result as for the entire annotated dataset. We assess empirically whether full-dataset decisions predict candidate usefulness on a held-out test set. On two LLM panels, stratified screening with certified QP rejection saves 21-23% of candidate calls while keeping 100% screening accuracy. In our experiments, sampling posterior scenarios further increases call savings to approximately 37%, with a loss of less than one percentage point in screening accuracy.

[MA-27] Loop Engineering: Version-Bound Evidence-Gated Coordination for Production Coding Agents

【速读】:该论文旨在解决生成式AI在实际软件工程生产环境中所面临的系统性挑战,即如何在不依赖私有对话交接的前提下,实现多个异步智能体(agent)循环在共享、可审计的代码仓库状态上安全、可追溯地协同工作。现有编码代理虽能生成代码并执行测试,但无法有效支持生产级软件开发所需的持久任务上下文、作用域授权、版本感知审批、源无关验证、中断恢复以及人机协作中的责任交接等关键需求。其解决方案的核心是提出“环工程”(Loop Engineering)方法,通过定义六要素结构化框架——触发器、上下文投影、权限边界、验证器、终止条件与回写规则——来规范每个智能体循环的行为。在此基础上,采用版本化的程序合约(称为Skills)封装可复用的操作策略,由“集成平台”(Harness)限定动作范围并收集操作证据,同时引入人类审核节点对高影响变更进行基于确切版本的授权。实证研究基于MewDesign商业AI设计产品,部署了五个专业化智能体环(涵盖规划、实施、评审、测试发布与对账),跨26个GitHub工作项的审计显示50条持久标记记录,其中3个拉取请求成功完成受版本约束的测试锁定,并利用独立于请求来源的运行时证据进行验证;另有2个在合并前中止。案例分析揭示了一种“失败闭合”的行为路径:新提交破坏了先前批准与证据链,系统拒绝重用过期的持续集成结果,体现了对状态一致性的严格保障。尽管研究未构成控制性基准实验,亦未证明因果生产力或普适安全性,但提供了在真实生产环境中可行性、可追溯性与有限推进性的初步证据。

链接: https://arxiv.org/abs/2611.18571
作者: Yupeng Pang,Meiling Chen,Yang Yang,Wei Fu
机构: MewDesign
类目: oftware Engineering (cs.SE); Multiagent Systems (cs.MA)
备注: 16 pages, 3 figures, 7 tables. Sanitized supporting data included in the source package. Zenodo preprint DOI: https://doi.org/10.5281/zenodo.21504929

点击查看摘要

Abstract:Coding agents can edit repositories and execute tests, yet production software work extends far beyond code generation. It requires durable task context, scoped authority, version-aware approval, source-independent verification, interruption recovery, and accountable handoffs between humans and automated workers. We present Loop Engineering, a method for operating multiple asynchronous agent loops over shared, auditable repository state without relying on private conversational handoffs. Each loop is specified by six elements: a trigger, a context projection, an authority boundary, a verifier, a stop condition, and a write-back rule. Versioned procedural contracts, called Skills, encode reusable operating policy; a Harness bounds the available action surface and collects evidence; human gates authorize high-consequence transitions against exact artifact versions. We report a design-and-experience study from MewDesign, a revenue-generating commercial AI design product. Five specialized loops cover planning, scoped implementation, review, test release, and reconciliation. A cross-sectional audit contains 50 durable marker records across 26 GitHub work items. Five pull requests reached a version-bound test lock: three proceeded to test releases with runtime evidence from sources distinct from the deployment request, and two stopped before merge. A detailed drift episode shows one observed fail-closed path: a new commit invalidated prior approval and evidence, and the workflow refused to reuse stale continuous-integration results. The study is not a controlled benchmark and does not establish causal productivity or general safety. It provides early evidence of feasibility, traceability, and bounded progression in one production setting. Comments: 16 pages, 3 figures, 7 tables. Sanitized supporting data included in the source package. Zenodo preprint DOI: https://doi.org/10.5281/zenodo.21504929 Subjects: Software Engineering (cs.SE); Multiagent Systems (cs.MA) Cite as: arXiv:2611.18571 [cs.SE] (or arXiv:2611.18571v1 [cs.SE] for this version) https://doi.org/10.48550/arXiv.2611.18571 Focus to learn more arXiv-issued DOI via DataCite

[MA-28] Sivia: Learning to Draw Learning to Learn

【速读】:该论文旨在解决学术论文概述图(overview figure)生成过程中存在的两大核心问题:一是如何准确、高效地从论文内容中提取并表达方法间的逻辑关系,确保科学准确性;二是如何在保持内容一致性的同时,根据用户偏好自适应地优化视觉呈现风格。其解决方案的关键在于提出Sivia系统,该系统通过两阶段递归自我改进(Recursive Self-Improvement, RSI)机制实现持续优化:第一阶段为“学会画”(learning to draw),基于对生成图像的准确性与美观性反馈,迭代提升绘图能力;第二阶段为“学会学习”(learning to learn),通过优化用于指导修订的更新器(updater),实现对自身改进策略的优化。此外,系统引入个性化用户画像,使视觉风格可适配用户的审美偏好,而科学内容保持不变。实验表明,基于源文本构建的关系规划显著提升了所需关系准确率(RRA),最高达19.85个百分点;在48篇测试论文上,双层级学习使RRA达到93.54%,显著优于初始版本及多个基线模型,同时在可读性和视觉质量方面均有显著提升,且个性化配置显著提高了用户喜爱度与采纳率。

链接: https://arxiv.org/abs/2609.33386
作者: Feng Yuan,Xinyi Gao,Zijun Sun,Haoyue Li,Menglei Hua,Xiao Li,Yifan Gao
机构: University of Manchester (曼彻斯特大学); University of Zurich (苏黎世大学); Harbin Medical University (哈尔滨医科大学); Beihang University (北京航空航天大学)
类目: Graphics (cs.GR); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Multiagent Systems (cs.MA)
备注:

点击查看摘要

Abstract:A paper’s overview figure is often the first thing readers see: it must explain the method correctly and look polished. We present Sivia, an overview-figure system that reads the paper, plans its relations, adapts visual organization from recent top-conference papers, and composes a design for image generation. Sivia improves with use through human-guided recursive self-improvement (RSI) at two levels: learning to draw revises its drawing skill from feedback on accuracy and presentation, and learning to learn revises the updater that proposes those revisions. A personal profile adapts presentation to its user’s aesthetic preferences while the scientific content stays fixed. A diagnostic study of 1,964 overviews for 445 papers (17,973 relation judgments) shows that most errors are omissions and that source-derived relation plans raise required-relation accuracy (RRA) by up to 19.85 points. On 48 test papers, learning both levels reaches 93.54% RRA, 15.31, 9.38, and 8.33 points above the initial skill, the Fixed-updater, and the GEPA baselines, while readability and visual quality rise from 3.95 and 4.00 to 4.70 and 4.78 on five-point scales. Learned presentation profiles raise liking from 4.02 to 4.73 and adoption from 34 to 45 of 48 tasks. Code and released artifacts are available at this https URL

[MA-29] Borrowed Strength Imported Risk: Capability Gaps and Reception Govern Heterogeneous LLM Debate EACL2027

【速读】:该论文旨在解决异质大语言模型(LLM)辩论中集体推理是否真正带来超越个体能力的性能提升这一核心问题。研究发现,所谓的“集体增益”主要并非源于多样性带来的协同效应,而是由能力传递(capability transfer)驱动:即具备任务优势的模型作为“同行”提供能力,接收方能够吸收其中可量化的部分。其解决方案的关键在于提出并验证了一个经验分解模型 ΔA_d^hon ≈ κΓ_d^+ρ_d,该公式将诚实增益(honest gain)精确描述为同行正向任务优势(Γ_d^+)与接收方接收能力(ρ_d)的乘积,揭示了能力传递的可量化机制。通过九组匹配方向及固定接收方的保留验证,研究证实接收能力在误差范围内保持稳定,而增益随优势增大而线性增长,进一步支持该模型的有效性。此外,在单方向开放式问答探测中,即使接收率高达62.50%,Qwen-7B仍仅实现+1.17点的可控首次交换增益,远低于其与Gemma-3-12B之间的10.78点差距,表明接收能力无法替代正向群体优势。当关键成员被赋予对抗角色时,原本借来的增益消失,甚至导致整体表现低于同质辩论水平。在所有五种全参考设置下,多数正确率始终未超过最具影响力成员的最佳准确率,即便存在未使用的互补回答。因此,异质辩论本质上是一种可测量的能力转移决策过程,具有明确的失效模式,而非单纯由多样性所赋予的优势。

链接: https://arxiv.org/abs/2611.20649
作者: Prashanti Nilayam,Kiran Kumar Ramanna,Prashil Tumbade,Nagaraj N Bhat,Sankalp Nayak
机构: ServiceNow
类目: Physics and Society (physics.soc-ph); Multiagent Systems (cs.MA)
备注: Accepted at EACL 2027. Code and experimental traces: this https URL

点击查看摘要

Abstract:Heterogeneous LLM debate is often attributed to collective reasoning that creates capability beyond any member. In the matched panels we study, the aggregate gain is dominated by capability transfer: a peer supplies task competence and receivers absorb a measurable fraction. Across nine matched directions, honest benefit and first-exchange adversarial exposure open when the peer has a positive same-task advantage, and honest gain is closely described by the empirical decomposition \Delta A_d^\mathrmhon\approx\kappa\Gamma_d^+\rho_d , the product of that advantage and the receiver’s reception. Two fixed-receiver directions provide held-out validation without refitting: reception remains stable within uncertainty as a larger peer advantage produces a larger gain. A matched, single-direction open-answer MuSiQue scope probe reaches the negative-gap side of the same gate. Qwen-7B trails the Gemma-3-12B receivers by 10.78 points; despite 62.50% reception, it produces only +1.17 points of controlled first-exchange gain. Reception cannot substitute for positive aggregate peer advantage. When an influential peer is assigned an adversarial role, the same influential peer slot removes the borrowed gain and can push the panel below homogeneous debate. Across all five full-reference settings, the best observed majority correctness does not exceed the influential member’s best observed accuracy, despite unused complementary answers. Heterogeneous debate is therefore a measurable transfer decision with a measurable failure mode, not a benefit conferred by diversity alone.

[MA-30] Brain AI Scientist: An Evidence-Based Multi-Agent System for Translational Neuroimaging in Precision Psychiatry

【速读】:该论文旨在解决精神疾病中基于静息态功能磁共振成像(resting-state functional magnetic resonance imaging, rs-fMRI)的脑网络异常发现难以转化为个体化临床决策支持的问题。其核心挑战在于如何将复杂的神经影像数据与临床信息、科学证据有效整合,实现可追溯、可验证的个性化诊疗推理。解决方案的关键在于提出“Brain AI Scientist”这一自主多智能体框架,通过检索、验证与优先级排序机制,融合患者特异性fMRI数据、临床特征及科学文献证据,依托研究领域标准(Research Domain Criteria, RDoC)和循证医学原则,构建从表型到解剖异常、生物标志物、作用机制、干预措施及验证策略的可追踪链路。在自闭症谱系障碍(autism spectrum disorder, ASD)案例中,该框架生成37项解剖学假设,经文献验证保留26项,并从90篇独立文献中提炼出65个有证据支持的生物标志物,最终输出个体化治疗与验证建议。在17例ASD患者中的评估显示,该方法在证据基础性、个性化程度、临床可操作性及生物标志物-治疗推理方面均显著优于GPT-3.5基线模型,无论是否引入fMRI印象。因此,Brain AI Scientist为将患者特异性神经影像转化为精准精神病学中的可审计、可解释的个体化临床决策支持提供了关键技术路径。

链接: https://arxiv.org/abs/2611.18634
作者: Zeyun Zhao,Lin Gu,Ruogu Fang
机构: 未知
类目: Quantitative Methods (q-bio.QM); Artificial Intelligence (cs.AI); Information Retrieval (cs.IR); Multiagent Systems (cs.MA); Neurons and Cognition (q-bio.NC)
备注:

点击查看摘要

Abstract:Resting-state functional magnetic resonance imaging (rs-fMRI) has revealed widespread brain-network alterations in psychiatric disorders, yet translating these findings into individualized clinical decision support remains challenging. Here we introduce Brain AI Scientist, an autonomous multi-agent framework that integrates patient-specific fMRI, clinical information, and scientific evidence through retrieval, verification, and prioritization. Guided by Research Domain Criteria and evidence-based medicine, it builds traceable chains linking phenotypes, anatomical abnormalities, biomarkers, mechanisms, interventions, and validation strategies. In a representative autism case, Brain AI Scientist generated 37 anatomical hypotheses, retained 26 after literature verification, and derived 65 evidence-supported biomarkers from 90 unique publications to generate patient-specific treatment and validation guidance. Across 17 autism spectrum disorder cases, it outperformed GPT-3.5 baselines with and without fMRI impression in evidence grounding, personalization, clinical actionability, and biomarker-treatment reasoning. Brain AI Scientist therefore provides an auditable framework for translating patient-specific neuroimaging into individualized clinical decision support for precision psychiatry.

自然语言处理

[NLP-0] Latent Lattice Language Models

【速读】: 该论文旨在解决传统语言模型在增加计算容量时面临的瓶颈问题,即通过加深网络或引入循环层来提升每标记(per-token)计算量,但受限于残差连接在深层中的信息流动效率。其核心挑战在于如何在不显著增加参数量的前提下,有效扩展模型的计算表达能力。解决方案的关键在于提出“潜空间格栅”(Lattice of Latents),这是一种在每一层中引入横向并行的持续性潜变量流(latent streams),这些潜变量沿时间维度递归运行,与传统的纵向残差流(residual stream)形成互补。不同于以往通过深度方向添加跳跃连接或重路由路径的方式,该框架通过维持跨标记的活跃计算状态,使潜变量直接参与注意力机制、键值缓存写入、MLP计算及解码过程,从而增强模型的动态计算能力。实验表明,在200亿条FineWeb-Edu训练数据上,一个406M参数的带格栅模型将验证困惑度从15.41降至14.40,且在多种架构(如Gated DeltaNet、Mamba-2、Triadic GDN等)中均表现出一致的性能提升,甚至在使用4倍更大原生状态尺寸的基线模型下仍能超越。这表明,持续性的逐层状态是语言模型计算扩展的一个可独立优化的补充维度。

链接: https://arxiv.org/abs/2611.21608
作者: Shivam Duggal,Zongze Wu,William T. Freeman,Antonio Torralba,Richard Zhang,Eli Shechtman
机构: MIT CSAIL(麻省理工学院计算机科学与人工智能实验室); Adobe(Adobe)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: Project page: this https URL

点击查看摘要

Abstract:Language models typically increase per-token computation through deeper or looped layers and latent reasoning steps, while richer residual connections improve information flow across layers. We instead increase computational capacity: the number of active computational states a model maintains. We introduce the Lattice of Latents, a set of per-layer horizontal latent streams that run recurrently along time, alongside the vertical residual stream that flows through depth. Like hyper-connections, mHC, and attention or value residuals, the Lattice enhances the residual stream, but rather than adding or rerouting residual pathways through depth, it provides companion computational states that persist across tokens. Unlike hybrid Transformer-SSM architectures, whose recurrent states primarily serve as internal memory for sequence mixing, these latents directly condition the vertical stream’s attention, KV-cache writes, MLP computation, and decoding. On Transformers trained on 20B FineWeb-Edu tokens, a 406M-parameter Lattice model lowers validation perplexity from 15.41 to 14.40 over a parameter-matched Transformer. These gains extend across architectural variants. The Lattice composes with prior residual-stream enhancements: with hyper-connections and input residuals, it lowers validation perplexity from 15.41 (base Transformer) and 14.22 (Transformer with the same enhancements) to 13.61. On Gated DeltaNet, Mamba-2, Triadic GDN, and GDN-Mamba-2 hybrids, it improves perplexity at every tested native state size, with Lattice-GDN and Lattice-Triadic-GDN variants outperforming pure baselines with 4x larger native states. These results identify persistent per-layer state as a complementary axis for scaling language-model computation.

[NLP-1] Quantifying the added value of human forecaster discussions for tropical cyclone intensity prediction

【速读】: 该论文旨在解决传统数值预测因子在热带气旋强度预报中的局限性,探索人工撰写的技术讨论文本(Tropical Cyclone Discussions)是否能为强度变化预测提供超越数值预报因子的额外价值。其解决方案的关键在于构建一种端到端的深度学习框架,将预报员的定性判断文本与13个来自SHIPS和ERA5再分析数据的环境预测因子进行多模态融合,以提升24至120小时预报时效内的最大持续风速预测精度。研究发现,仅使用讨论文本即可在24小时预报中达到与SHIPS数值因子相当的性能;进一步将文本信息融入基于五种模型的集成系统后,2022–2023年飓风样本的24小时最大持续风速均方误差从9.89节降至8.90节,且置信区间不包含零,表明改进具有统计显著性。此外,文本分析揭示“强度”“强度变化”类关键词(如intensify、weaken、eyewall、category)对模型预测具有关键贡献。尽管再分析数据与实时分析存在差异,且讨论发布时间晚于其他输入数据至少三小时,无法直接用于业务运行,但本研究通过多模态学习验证了专家主观判断在强度预测中的可量化附加值,为融合人类经验与机器学习提供了新范式。

链接: https://arxiv.org/abs/2611.21602
作者: Ilia Azizi,Frederick Iat-Hin Tam,Marc-Olivier Boldi,Valérie Chavez-Demoulin,Tom Beucler
机构: 未知
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 74 pages, 24 figures, 35 tables

点击查看摘要

Abstract:The National Hurricane Center routinely releases Tropical Cyclone Discussions that offer insights into its tropical cyclone forecasts. This study examines the added value of these discussions for intensity forecasting beyond a set of numerical predictors. For 637 Atlantic and eastern North Pacific tropical cyclones from the 2002-2023 hurricane seasons, we analyze the discussion text and its combination with 13 hurricane environmental predictors from SHIPS and ERA5 10-meter wind fields, for forecasting intensity changes at lead times of 24 to 120 hours using three modeling methods. In one approach, an end-to-end deep learning framework for the text shows that the discussions alone offer comparable value for 24-hour intensity forecasts to the SHIPS predictors. Adding discussion information to the SHIPS predictors in an ensemble of five such models reduces the mean absolute error of 24-hour maximum sustained wind forecasts for 2022-2023 cyclones from 9.89 to 8.90 kt, with a confidence interval that excludes zero when whole storms are resampled. The ERA5 wind fields have a minor effect on mean absolute error in wind predictions, and model rankings differ across predicted variables, lead times and model types. We provide prediction intervals for these forecasts, whose coverage falls below the 90% target during rapid intensification. Analysis of the text shows that words related to intensity and intensity change (such as intensify, weaken, eyewall and category) are important for the model predictions. This retrospective study shows the added value of the forecasters’ discussions and suggests that multi-modal learning can help understand the predictive contribution of expert human judgment. The reanalysis data differ from real-time analyses, and the discussions are issued at least three hours after the valid time of the other inputs, so these results should not be interpreted operationally.

[NLP-2] How VLMs Plan Whats Next

【速读】: 该论文旨在解决视觉语言模型(Vision-Language Model, VLM)在描述图像时如何规划生成顺序这一关键问题,即模型在逐个命名图像中的物体时,其决策机制背后的认知过程尚不明确。研究提出了一种可解释的生成规划机制,其核心在于通过四个协同组件实现对视觉-语言交互的分步控制:首先,绑定头(binding heads)将文本提示中的语义信息与图像区域进行匹配,生成指向具体视觉区域的指针;其次,相同绑定头将已生成的回答与其对应的图像区域关联,并将此过程转化为一个“进度信号”——即游标(cursor),指示当前已描述的图像区域边界;第三,选择步骤(select step)基于游标位置,选取最近未被描述的区域,或决定终止生成;最后,注视头(gaze heads)执行所选指针,定位目标区域并提取其语义概念,由语言头和后期多层感知机(MLP)将其转换为符合问题语言的词汇输出。研究通过因果打补丁(causal patching)方法在Qwen3-VL模型上验证了该机制在合成视觉检索任务中的有效性,并排除了多种替代假设。进一步实验表明,该机制在不同输入格式、自然照片及多语言场景下均具鲁棒性,且其绑定、中继与注视头在2B至32B规模的所有模型层级中均稳定出现。本工作不仅揭示了VLM如何逐层、逐词地将问题与图像转化为序列化回答,更为理解大模型在复杂多步任务中进行长程规划提供了可解释的分析框架。

链接: https://arxiv.org/abs/2611.21600
作者: Rohit Gandikota,Arnab Sen Sharma,Adam Belfki,David Bau
机构: Northeastern University (东北大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:When a vision-language model describes an image, it names one object at a time and it rarely repeats itself, yet how the model plans this order is far from obvious. We characterize the mechanism a VLM uses to decide where to look before it looks. The mechanism has four main parts. First, binding heads match the relevant text tokens against the image and resolve them into pointers to the matched image region. Second, the same binding heads bind each written answer to the region it came from, and the model turns it into a progress signal: a cursor, the position in the image up to which the list has been written. Third, a select step chooses the nearest unwritten region after the cursor, or stops. Finally, gaze heads execute the chosen pointer and return a concept of what is drawn there, which language heads and the late MLPs turn into a word in the language of the question. We validate each part of this mechanism through causal patching in Qwen3-VL models on a synthetic visual retrieval task, and test our mechanisms against alternative hypotheses. We further show that the mechanism holds across input formats, natural photographs, and languages, and that its binding, relay, and gaze heads appear at every model size from 2B to 32B. This work allows us to investigate how a VLM turns a question and an image into a sequence of answers, layer by layer and token by token. More broadly, it opens a way to study how models plan over complex multi-step tasks with long rollouts. Our project is available at this https URL

[NLP-3] Beyond Job Titles: Interpreting Early-Career Job Advertisements with Sparse Autoencoders and Semantic Clustering

【速读】: 该论文旨在解决在快速变化的就业市场中,仅依靠职位名称和职位描述难以准确揭示技术类岗位之间实际工作内容与行业应用差异的问题。其核心挑战在于如何从海量非结构化文本数据中提取具有语义意义的岗位特征模式,并实现对不同技术角色的系统性比较。解决方案的关键在于提出并应用一种结合深度学习与变压器(Transformer)表示的综合分析框架——PERC(Preprocessing, Embedding, Dimensionality Reduction, and Clustering),并通过多阶段自动化流程实现对职位描述的细粒度解析与聚类。具体而言,研究利用GPT-6 Luna对职位描述进行三段式分类(团队背景、职责任务、资格要求),生成语义组合后进行嵌入表示与聚类分析;采用最大均值差异(MMD)量化不同岗位间嵌入分布的差异性,并借助预训练稀疏自编码器(SAE)从Llama-3.1-8B-Instruct模型第19层的激活表示中提取高阶特征;最后通过GPT-6.1 Terra验证所发现特征模式在未见企业数据中的泛化能力。该方法不仅实现了对数据科学家、数据分析师与机器学习工程师等岗位在技术侧重上的精准区分(如模型开发、业务报告、真实场景部署等),还为自动化洞察提取提供了可扩展的技术路径,显著提升了人工分析效率与客观性。

链接: https://arxiv.org/abs/2611.21586
作者: Yisong Cao,Leah Kim,Brian Wright,Ali Rivera
机构: Active Learning Lab; School of Data Science, University of Virginia (弗吉尼亚大学数据科学学院)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Job titles and job descriptions alone provide limited insight into how technical roles differ in a changing job market. To assess how such roles differ in their work and industry applications, we design and compare two methods based on deep learning and transformer representations using 36,081 early-career job postings collected from LinkedIn in the United States between December 2025 and October 2026. We first apply PERC, short for Preprocessing §, Embedding (E), Dimensionality Reduction ® and Clustering ©. We employ GPT-6 Luna to classify job description passages into Team context, Role task and Qualifications. We then create combinations of these sections and examine whether embedding and clustering the segmented passages produce meaningful semantic clusters about different job roles. We use Maximum Mean Discrepancy (MMD) to rank observed differences between embedding distributions and employ a pretrained sparse autoencoder (SAE) to extract features from layer-19 representations of Llama-3.1-8B-Instruct. Lastly, we employ GPT-6 Luna to come up with hypotheses about job roles’ feature patterns, and GPT-6.1 Terra to investigate whether those patterns exist in held-out job descriptions from unseen companies. Some example findings include Data Scientist advertisements placing more emphasis on machine learning model development, Data Analyst advertisements on reporting business performance, and Machine Learning Engineer advertisements on applying developed models in real-world settings. We also observe differences in the applications described by Data Analyst and Business Analyst advertisements across selected months, although our dataset’s scale and uneven coverage limit their interpretation. Overall, we present a realistic application of SAEs and PERC to unstructured textual data and examine how automated insight extraction could support manual analysis of job descriptions. Subjects: Computation and Language (cs.CL) Cite as: arXiv:2611.21586 [cs.CL] (or arXiv:2611.21586v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.21586 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[NLP-4] Reading or Remembering? Measuring Temporal Contamination in Language Models

【速读】: 该论文旨在解决生成式语言模型在解读央行文本(如美联储联邦公开市场委员会,FOMC)用于经济预测时存在的“时间性测试集污染”(temporal form of test-set contamination)问题。具体而言,当前主流语言模型多基于2021年及之后的文本进行预训练,而这些模型在分析2021年发布的FOMC声明时,可能已“知晓”后续发生的经济事实(如2022年通胀上升),从而导致其预测性能部分源于对结果的内存记忆,而非真正的推理能力。为解决此问题,研究提出关键方法是构建“时间一致模型”(chronologically consistent models),即仅使用目标文档发布日期之前的历史文本进行训练,以实现与现代开放模型在相同时间点上的公平对比。通过对比现代模型与时间一致模型在预测历史GDP等经济指标时的表现,发现现代模型仍能识别出经济增长较强的年份,表明其存在显著的时间偏差。进一步地,研究采用成员推断分数(membership-inference scores)验证模型对文档所属时期和类型的学习能力,发现其分类依据更多是时间背景与文档类型特征,而非是否曾接触过特定文档。最终,在实时经济预测(nowcasting)中,所有可检测到的“前瞻溢价”(lookahead premium)均小于宏观经济样本量所能区分的阈值,且经多重检验校正后均不显著。因此,研究强调:未来应直接在相同任务上对比模型与记忆消除型模型(memory-free models),以准确评估模型是否存在因训练数据泄露而导致的虚假性能提升。

链接: https://arxiv.org/abs/2611.21554
作者: Arunkumar V,Gangadharan G. R.,S. Senthilkumar
机构: UCE BIT Campus, Anna University(安纳大学UCT BIT校区); NIT Tiruchirappalli(蒂鲁奇拉帕利国家理工学院); Tiruchirappalli, India(印度蒂鲁奇拉帕利)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 17 pages

点击查看摘要

Abstract:Language models are now used to read central-bank text for economic forecasting, but most were trained on text written after the documents they read. A model reading a 2021 FOMC statement may already know that inflation rose in 2022, so part of its apparent skill can be memory of the outcome, a temporal form of test-set contamination. We measure this memory on FOMC statements and minutes from 2000 to 2026 by comparing modern open models with chronologically consistent models, which are trained only on text from before a fixed date. Asked about past GDP outcomes given only a date, models that cannot know the answer still reach an AUC of up to 0.64 from priors on dates and eras, so chance is the wrong baseline. Against these memory-free models on the same questions, most modern models recall which years had strong growth. Membership-inference scores sort documents by era and document type rather than by whether a model saw them, as exact counts in an open pretraining corpus show. In real-time nowcasts, every lookahead premium we estimate on US GDP is smaller than the forecast comparison can detect at macroeconomic sample sizes, and none on GDP or inflation survives correction for multiple testing. Lookahead audits should therefore measure memory directly against memory-free models on the same items.

[NLP-5] Right Reason Wrong Advice: Evaluating Recourse in LLM Explanations of Denied Requests

【速读】: 该论文旨在解决生成式人工智能(Generative AI)在解释规则系统拒绝用户请求时存在的“可追溯性失效”问题。当多个独立规则同时导致拒绝时,即使解释中指出的某个理由为真,用户修复该理由后仍可能被拒绝,即解释虽正确却无法提供有效行动路径。这一现象的核心问题是解释缺乏可追溯性有效性(recourse validity):仅更改解释所引用的原因,而不改变实际申请内容,可能导致决策结果反转。论文提出通过精确计算可追溯性有效性,并基于击中集(hitting-set)理论从策略本身预测此类失效的发生。在六个来自访问控制和美国福利资格审查的规则系统中,当存在多个适用规则时,单一正确原因几乎无法提供有效救济路径。对20个大型语言模型(LLM)的评估显示,无论是通过人工判断还是由其他语言模型进行验证,识别这种失效的能力接近随机水平,与理论预测一致;而若采用强模型评估“依据解释采取行动是否有效”,则能显著提升检测能力。尽管更强的语言模型和具备推理能力的模型表现更优,但在涉及四个或更多独立规则的拒绝案例中仍频繁失败。因此,论文建议:资格决策的解释应基于决策函数构建,并以“执行解释是否能改变结果”作为核心评判标准,从而确保解释具有实际可操作性。

链接: https://arxiv.org/abs/2611.21545
作者: Arunkumar V
机构: University College of Engineering, BIT Campus, Anna University (安娜大学比特校区工程学院); Tiruchirappalli, India
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)
备注: 21 pages

点击查看摘要

Abstract:Language models increasingly write the explanations people receive when a rule system denies their request. If several rules deny the request independently, an explanation can name a true reason and still leave the person denied after they fix it. We call the missing property recourse validity: changing what the explanation cites, without changing what was asked for, can flip the decision. We compute it exactly, and a hitting-set argument predicts from the policy alone when a correct reason fails. In six of seven rule systems ported from access control and US benefit eligibility, a single correct reason almost never gives recourse when several reasons apply. Across 20 LLMs, checks of whether the cited reasons are correct, computed or asked of an LLM judge, are close to chance at spotting this failure, as the argument predicts. A strong judge asked whether acting on the explanation would work does spot it. Stronger and reasoning models do better but still fail on many denials with four or more independent reasons. Asking for every rule helps every open model with room to improve. Explanations of eligibility decisions should be built from the decision function and judged by whether acting on them works.

[NLP-6] Seen but Unsaid: Locating and Repairing Cross-Image Comparison in Vision-Language Models

【速读】: 该论文旨在解决视觉语言模型(Vision-Language Models, VLMs)在跨图像点或对象匹配任务中表现不佳的问题,尽管模型的隐藏状态中已包含可解码的对应信息。其核心问题是:尽管模型内部存在正确的语义对应关系,但这些信息未能有效传递至输出层,导致模型无法正确生成答案。解决方案的关键在于识别出两个关键失败点:一是仅有少数注意力头(attention heads)负责执行跨图像比较,二是这些关键头部的参数更新量极小(仅占总参数的0.02%),且其作用被严重抑制。研究发现,通过仅训练这些关键注意力头的查询(query)和键(key)投影,即可显著提升性能(如在SPair-71k上提升28.3点),且该方法能迁移至其他任务(如HPatches和BLINK)。此外,一个无约束的全层适配器(unconstrained all-layer adapter)也证实,其性能增益主要通过改变这些特定头部的输入读取方式实现(贡献率达83–86%),而非修改权重本身。这一结果表明,模型决策机制虽被冻结,但通过精准干预关键注意力路径,可有效释放隐藏的对应能力。

链接: https://arxiv.org/abs/2611.21539
作者: Zhaoning Zhang
机构: University of Toronto(多伦多大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 18 pages, 3 figures, 19 tables

点击查看摘要

Abstract:Vision-language models (VLMs) often fail to match a point or an object across two images, even when the answer can be decoded from their own hidden states. We trace this failure from the visual tokens to the answer letter, in Qwen3-VL (2B-32B), InternVL3.5-8B, and Gemma-3-12B, on synthetic shape matching, same-instance point correspondence (HPatches), and semantic correspondence (SPair-71k), and find two distinct failure points, a small set of attention heads that carries the comparison, and a repair that touches 0.02% of the parameters. Two failure points: exact copies are compared but not read out. An answer-position probe exceeds the calibrated output by 15 to 25 points in every family, for nameable and unnameable shapes alike, and steering along the probe’s direction moves no more answers than random directions. Semantic correspondences are present in the visual tokens but never reach the answer. One pathway: redistributing the cross-image attention of 16-32 heads, with its mass preserved, removes 17-22 points of synthetic and 15-32 points of HPatches accuracy, against at most 7 for random or attention-matched heads. A lever that training finds on its own: training only these heads’ query and key projections raises SPair-71k accuracy by 28.3 points (Qwen3-VL-8B) and 25.1 points (InternVL3.5-8B), beats both layer-matched and attention-matched heads, and transfers to HPatches and BLINK. An unconstrained all-layer adapter routes 21-51% of its gain through the attention of the same heads, against at most 7% for random heads, and does so by changing what they read (83-86% of the effect) rather than their weights. Every decision rule was frozen before its run, and every reported number is recomputed by independent verification scripts.

[NLP-7] Optimistic Verifiers Get Hacked: Reward Hacking Under an Exact Process Oracle NEURIPS2026

【速读】: 该论文旨在解决强化学习中基于可验证奖励(verifiable rewards)时,代理模型对验证器(verifier)偏差的敏感性问题。在数学与代码以外的领域,验证器通常作为真实目标的近似代理(proxy),如评分标准、学习型裁判或人工编写的检查器,其与真实目标之间的偏差无法量化,导致奖励上升可能仅反映对验证器的“绕过”而非真正优化。为克服这一局限,研究构建了一个可精确计算过程级奖励的实验环境:基于124,416个上下文的可执行授权策略(executable authorization policies)生成一个能枚举每个决策最小充分理由的真值查询器(oracle),从而实现对奖励的精确建模。研究将该真值验证器与七种现实中的退化验证器(包括仅关注结果、缺乏充分性、召回率受限、提取噪声、粗粒度分类、策略陈旧及大语言模型裁判等)进行对比,在每轮训练中均使用精确验证器评估所有轨迹。结果显示四种典型现象:乐观型验证器易被“填充攻击”(padding attack)利用,其充分性指标从0.940下降至0.667,而真实忠实度(faithfulness)从0.731降至0.667;仅关注结果的验证器趋于饱和并停止教学,噪声引入使性能稀释,但保持排序一致性的差距影响较小。研究进一步提出一种无模型的验证器审计方法,通过分析验证器的选择误差及其差距符号,并结合简短饱和性预实验,可有效预测特定训练运行将呈现的类型。在九种配置组合下(涵盖两种框架、两个模型供应商、两种适配器尺寸、三种精度、两种初始化和两种随机种子),这些现象具有高度可复现性,充分性损害在全部10个复现单元中均为正向且在7个中超过两倍标准误,表明乐观性差距始终是更严重的风险。尽管损害幅度未完全复现,但五项预先注册的解释尝试均未能成功。研究公开了测试平台、训练工具链以及14个真实公共政策(如Cedar、OPA、Gatekeeper)移植后的可执行真值查询器。

链接: https://arxiv.org/abs/2611.21538
作者: Arunkumar V,Manoranjan Gandhudi,Gangadharan G. R.,Arun Prakash,S. Senthilkumar
机构: University College of Engineering, BIT Campus, Anna University, Tiruchirappalli, India; Central University of Karnataka, Kalaburagi, India; National Institute of Technology Tiruchirappalli, India; Jawaharlal Nehru University, New Delhi, India
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 30 pages. A shorter version, “The Verifier Gap”, appears at the NeurIPS 2026 Workshop on Meta-Agents (non-archival)

点击查看摘要

Abstract:Reinforcement learning from verifiable rewards inherits whatever its verifier gets wrong. Outside mathematics and code the verifier is a proxy, a rubric, a learned judge or a hand-written checker, and its distance from the true objective is never measured, so a rising reward cannot be told apart from a hack. We build a setting where the true reward is exactly computable at the process level: executable authorization policies over 124,416 contexts admit an oracle that lists every minimal sufficient reason for every decision. We compile these policies into exact process rewards, degrade the verifier along seven realistic axes (outcome-only, sufficiency, recall-only, extraction noise, coarse taxonomy, stale policy, LLM judge), and train against each while the exact verifier scores every rollout. Four phenotypes appear. Optimistic verifiers are hacked by padding: the sufficiency proxy ends at 0.940 while true faithfulness falls from 0.731 to 0.667 on 96 held-out episodes. Outcome-only verifiers saturate and stop teaching, noise dilutes, and ranking-preserving gaps do no harm. A model-free audit of the verifier, its selection error and gap sign, together with a short saturation pilot, anticipates which phenotype a run will produce. Across nine configurations spanning two frameworks, two model vendors, two adapter sizes, three precisions, two initialisations and two seeds, the phenotypes recur, sufficiency damage is positive in all 10 replication cells and beyond two standard errors in 7, and optimistic gaps remain the steeper branch. The damage magnitude does not replicate, and five pre-registered attempts to explain it are reported as they fell. We release the testbed, the training instrumentation, and 14 real public policies (Cedar, OPA, Gatekeeper) ported to executable oracles.

[NLP-8] Evaluating Cultural Reasoning and Robustness in LLM s for Spanish Political Discourse

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在政治敏感语境下对西班牙语社交媒体政治话语进行情感分析时的可靠性问题,尤其关注文化情境化与语用间接性所带来的情感意义表达难以被传统标签导向评估方法充分捕捉的挑战。其解决方案的关键在于提出一种诊断性评估协议,通过整合三个独立的人工标注视角,将数据划分为“黄金标准”(Gold)与“模糊”(Ambiguous)子集,并基于模型条件化的困惑度与大语言模型作为评判者(LLM-as-a-judge)的标准,构建由情感反转反事实样本组成的“合成”(Synthetic)子集,从而系统性地揭示模型在语义模糊性与对抗性扰动下的脆弱性。实验表明,该框架能有效暴露传统评估中被掩盖的模型失效模式;其中GPT-4.1-mini在零样本与少样本提示下表现最为稳定,而两个较小的开源模型在合成扰动下显著退化,且少样本提示带来的性能提升有限。研究贡献包括可复现的鲁棒性评估流程、对模型在模糊性与情感反转反事实场景下行为的实证刻画,以及一个带有三重人工情感标注的公开西班牙语社交媒体数据集。

链接: https://arxiv.org/abs/2611.21535
作者: Jorge Valverde-Rebaza,Nathan Garcia-Larios,Karime Ruvalcaba-Perez,Alejandro Rendon,Santiago Sanchez-Ballesteros,Alejandro Martin del Campo
机构: Tecnologico de Monterrey(墨西哥蒙特雷科技大学); School of Engineering and Sciences(工程与科学学院)
类目: Computation and Language (cs.CL); Computers and Society (cs.CY); Social and Information Networks (cs.SI)
备注: Accepted for publication in Latin American Conference on Informatics (CLEI 2026)

点击查看摘要

Abstract:Large language models (LLMs) are increasingly used to interpret, generate, and mediate public-facing language, which makes their reliability especially consequential in politically sensitive settings. Although sentiment analysis is a common task in this area, affective meaning in social media discourse is often conveyed through culturally situated and pragmatically indirect cues that are not well captured by label-centric evaluation alone. In this paper, we evaluate the robustness of LLMs for sentiment analysis of Spanish political discourse from social media, with emphasis on culturally situated affective interpretation. We introduce a diagnostic protocol that combines three independent human annotation views into Gold and Ambiguous subsets, and extends the corpus with a Synthetic subset built from sentiment-flipped counterfactuals selected through model-conditioned perplexity and LLM-as-a-judge criteria. Using this framework, we compare proprietary and open-weight models under zero-shot and few-shot prompting, focusing on predictive robustness across human agreement conditions and counterfactual shift. Our results show that annotation ambiguity and counterfactual perturbation reveal failure modes that conventional evaluation on the original corpus tends to obscure. GPT-4.1-mini remains the most stable among the evaluated models, whereas the two smaller open-weight models degrade more markedly under synthetic perturbations, and few-shot prompting provides only limited gains. Overall, the paper contributes a reproducible protocol for robustness-oriented evaluation in political sentiment analysis, an empirical characterization of model behavior under ambiguity and sentiment-flipped counterfactuals, and a publicly available Spanish social media dataset with three-way human sentiment annotations.

[NLP-9] Procedure without Judgment: What Small Language Models Learn under an Exact Verifier

【速读】: 该论文旨在解决在授权策略(authorization policy)拒绝代理工具调用时,如何提供精确且最小化的决策解释(minimal sufficient reason)的问题。其核心挑战在于确保审计日志仅包含导致决策的关键事实,避免冗余信息。解决方案的关键在于利用可执行策略在可枚举上下文空间中的精确性,通过构建一个无误差的验证器(verifier)来生成真实标签,从而实现对解释模型的精准评估。研究发现,前沿模型仅凭工具调用即可在约50%的情况下找到确切原因;而通过一句提示引导模型缩小答案范围,Haiku 4.5的准确率从58%提升至75%。进一步地,使用2000条删除算法轨迹微调的4B模型在训练策略上达到92%的准确率。关键突破在于“教师类型”(teacher kind)的影响:基于算法训练的学生模型在相同数据量下始终优于基于前沿模型训练的学生,且两者的性能无重叠。学生所学习的是固定的充分性检查序列,但该序列不具备泛化能力。在三个未见过的公开策略上,该模型虽仍能保持充分性(94%),但最小性显著下降至48%,而同样条件下Haiku则在全部47个测试案例中保持精确。此外,基于精确奖励的强化学习虽未提升保留数据上的准确率,却促使策略从试探性查询转向耗尽预算进行充分性检查。由此可见,虽然理想验证器使训练效果可解释,但并未实质性改善训练效率或泛化能力。

链接: https://arxiv.org/abs/2611.21529
作者: Arunkumar V,Manoranjan Gandhudi,Gangadharan G. R.,Arun Prakash,S. Senthilkumar
机构: Anna University (安娜大学); Central University of Karnataka (卡纳塔克中央大学); NIT Tiruchirappalli (蒂鲁奇拉帕利国家技术学院); Jawaharlal Nehru University (尼赫鲁大学); Anna University (安娜大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 17 pages

点击查看摘要

Abstract:When an authorization policy blocks an agent’s tool call, the audit log should name the facts that forced the decision and nothing more. For an executable policy over an enumerable context space this minimal sufficient reason can be computed exactly, so an explanation is scored against ground truth rather than a proxy for it. We use that exactness to ask what a model can learn from a verifier that has no error in it. The explainer sees one request and one decision, never the policy, and has thirty turns in which it may call the policy on altered requests and ask whether a candidate set of facts is sufficient. Frontier models find the exact reason about half the time with the tools alone, and one sentence telling them to shrink the answer lifts Haiku 4.5 from 58% to 75%. A 4B model fine-tuned on 2000 traces of the deletion algorithm reaches 92% on the policy it was trained on. Teacher kind is what moves this: at matched data size the algorithm-taught student leads the frontier-taught one in all three training seeds, and the two arms never overlap. What the student learns is a fixed schedule of sufficiency checks, and the schedule does not travel. On three public policies it never saw, its answers stay sufficient (94%) and stop being minimal (48%), while Haiku under the same sentence is exact on all 47 episodes. Reinforcement learning from the exact reward leaves held-out accuracy where it was and inverts the strategy, from probing the policy to spending the entire budget on sufficiency checks. The oracle made the effect of training legible, but it did not make training better.

[NLP-10] Embedding Cosine Measures Text Length Before It Measures Homogenization

【速读】: 该论文旨在解决生成式AI对在线话语趋同性(homogeneity)影响的测量偏差问题,尤其关注基于Transformer嵌入的平均余弦相似度这一常用指标在实际应用中的有效性。研究发现,当使用未经句间相似性训练的编码器(如DeBERTa-v3)进行均值池化后,该相似度指标主要反映文本长度而非语义趋同,导致误判。其解决方案的关键在于识别并校正这种由文本长度引起的虚假趋同信号:通过截断文本至10–150词范围、引入等长随机替换的安慰剂实验以及固定内容模拟,结果表明80%以上的“趋同”现象可由长度差异解释;进一步分析显示,主导主成分承载了99.9%的原始相似度方差,去除长度影响后,所有编码器均未表现出显著的跨时间趋势差异。此外,提出一种低成本诊断协议,用于检测模型架构与数据设计之间的非各向同性(anisotropic)交互风险。修正后的度量表明,政治两极化群体在2019–2022年间确实存在收敛,但该趋势在ChatGPT发布前已停止,且跨阵营回复比同阵营回复更具针对性。

链接: https://arxiv.org/abs/2611.21519
作者: Fengming Liu
机构: 未知
类目: Computation and Language (cs.CL); Computers and Society (cs.CY); Social and Information Networks (cs.SI)
备注: 11 pages, 10 figures, 2 tables. Re-measurement of arXiv:2608.13674

点击查看摘要

Abstract:Mean pairwise cosine similarity between transformer embeddings is a standard measure of whether online discourse is converging under generative AI. When the embeddings are mean-pooled outputs of an encoder never trained for sentence similarity, this measure mostly tracks text length. In 6.06 million Reddit comments, an uncentered DeBERTa-v3 measure had suggested that only conservative discourse broke toward homogeneity after late 2022. Truncating the same comments from 10 to 150 words raises their mean cosine by eight times the reported break. A placebo that swaps every comment for a random one of equal length reproduces 81% of the asymmetry, and a fixed-content simulation reproduces 78%. The leading principal direction carries 99.9% of raw DeBERTa homogeneity, and once length is handled the asymmetry disappears under every encoder. On 575,000 Stack Exchange answers the measure reports a post-ChatGPT homogenization of technical sites that the placebo fully reproduces, although answers grew longer. A census of 31 studies locates the risk where anisotropic encoders meet time-series or length-unequal designs. A reverse control on a well-known experiment shows that centering with renormalization can reverse a genuine effect. I propose a low-cost diagnostic protocol. Corrected measures show that the two political camps converged between 2019 and 2022 and stopped converging before ChatGPT. Cross-camp replies engage their targets more closely than same-camp replies.

[NLP-11] Detection-Efficient Watermarking for Vision–Language Agents via Risk-Constrained Strength Allocation

【速读】: 该论文旨在解决生成式视觉-语言代理(Vision-Language Agents)在生成可重用内容时面临的可信度与可审计性问题,即如何在保障内容真实性与支持部署后审计的同时,避免对响应质量造成不必要的负面影响。其核心解决方案是提出一种风险感知的语义水印框架,通过结合预测不确定性与零样本多模态相关性-风险代理(zero-shot multimodal relevance-risk proxy),在解码过程中动态分配水印强度。该方法仅需一个风险惩罚系数λ,即可构建连续的操作边界,使部署时能够灵活选择检测能力与响应质量之间的权衡,而无需重新训练模型。实验表明,在选定操作点下,平均水印强度相比固定强度水印降低了约59.4%,困惑度变化从0.3634降至0.1689,且干净文本与改写文本的ROC-AUC分别达到0.9552和0.9459。对比分析进一步证实,该优势源于智能的水印分配策略而非更低的平均干预强度:在相近平均强度下,基于风险约束的分配策略显著优于固定强度、仅依赖熵或仅依赖关键性的策略,证明其为高检测效率型方案。此外,研究还揭示了有限样本阈值校准会导致显著的假阳性率漂移,因此建议将校准方差报告作为可审计溯源的标准实践。

链接: https://arxiv.org/abs/2611.21488
作者: Dumindu Samaraweera,Prashant Shekhar
机构: Embry-Riddle Aeronautical University (埃姆布里-瑞德航空大学)
类目: Cryptography and Security (cs.CR); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Vision–language agents generate descriptions, decisions, and instructions that may be reused by users, downstream agents, and automated workflows, creating a need for provenance mechanisms that support content authenticity and post-deployment auditing without unnecessarily degrading response quality. We present a risk-aware semantic watermarking framework that combines predictive uncertainty with a zero-shot multimodal relevance-risk proxy to allocate watermark strength during decoding. A single risk penalty \lambda exposes a continuous operating frontier, allowing deployment-time selection of a detection–quality trade-off without model retraining. At a selected operating point, mean watermark strength is reduced by approximately 59.4% relative to fixed-strength watermarking, while perplexity change falls from 0.3634 to 0.1689 and clean/paraphrased ROC-AUC reaches 0.9552/0.9459. A matched-strength analysis shows that this advantage reflects the allocation policy rather than lower average intervention: at comparable mean strength, risk-constrained allocation achieves substantially higher ROC-AUC than fixed, entropy-only, and criticality-only policies, identifying it as detection-efficient rather than quality-dominant. We further show that threshold calibration on finite samples produces substantial false-positive-rate drift, motivating calibration variance reporting as standard practice for auditable provenance.

[NLP-12] Jetriever: Shifting Agent Memory Retrieval from Similarity to Utility

【速读】: 该论文旨在解决传统记忆检索系统中依赖相似性匹配导致的效率与准确性瓶颈问题,即存储的记忆片段若与当前查询语义不直接相似,则难以被有效召回,从而影响问答性能。其核心解决方案是提出一种基于Jev决策得分(Jev’s decision score)的全池效用评分机制(full-pool utility scoring),取代传统的基于相似度的搜索方式。该方法通过量化每个记忆片段对当前问题的决策贡献度,实现对所有候选记忆的全局排序,确保最具决定性的知识优先被检索。实验表明,在InMind的125个间接查询任务中,Jetriever在84.8%的情况下将正确答案排在首位,显著优于BM25(1.6%)和MiniLM(0.0%)等基线模型;在证据发现与排序分离的20题测试中,Jetriever在前50名内召回率达85%,排名第一,而Qwen3.5 9B为40%。此外,在MemoryAgentBench的随机采样测试中,Jetriever四维评分达到55.4%,优于ReFind的44.3%。同时,其平均检索成本分别为0.00834美元/次(InMind)和0.03103美元/次(MAB),证明了该方案在高精度与低成本之间的良好平衡。因此,该研究的关键在于利用世界知识驱动的效用评分机制,实现精准、高效且经济的决策导向记忆检索。

链接: https://arxiv.org/abs/2611.21478
作者: Ruizhe Li,Licheng Zhang,Mingxuan Du,Zhendong Mao
机构: 未知
类目: Computation and Language (cs.CL)
备注: 13 pages, 1 figure, 8 tables. Project repository: this https URL

点击查看摘要

Abstract:A useful memory need not resemble the question it helps answer. Jetriever ranks each stored memory by Jev’s decision score for its usefulness to the current question, replacing similarity search with full-pool utility scoring. On InMind’s 125 indirect-query tasks, it places the gold exchange first in 84.8% of cases; matched BM25 and MiniLM controls reach 1.6% and 0.0% when indexing user turns. A 20-question comparison separates finding evidence from ranking it: Jev and Qwen3.5 9B each find 90% of gold exchanges by rank 50, while Laya finds 30%; at rank one, Jev reaches 85% and Qwen3.5 9B 40%. On a dimension-proportional random sample of 60 MemoryAgentBench questions, Jetriever’s four-dimension score is 55.4%, against 44.3% for ReFind on the same questions. Jev retrieval costs USD 0.00834 per InMind question and USD 0.03103 per MAB question on average. These results show the value of world knowledge, scores that rank decisive memories first, and affordable full-pool scoring.

[NLP-13] Balanced Twin Weight Factorization for Efficient MLLM s

【速读】: 该论文旨在解决权重分解(Weight Factorization, WF)方法在实际应用中难以将减少的浮点运算量(FLOPs)有效转化为推理延迟降低的问题,其核心挑战在于:将单一大型权重矩阵替换为多个小型矩阵会引入额外计算开销。为此,本文提出平衡双分支权重分解(Balanced Twin Weight Factorization, BTWF),其关键在于利用因子化模型的共享基结构,通过引入双重平衡机制——因子成分平衡(Factor Component Balancing, FCB)与块贡献平衡(Block Contribution Balancing, BCB),确保共享组件在因子分解中得到充分表达并优化下游模型性能。此外,BTWF创新性地采用过参数化设计,在训练阶段允许部分层的因子产生高于原始模型的FLOPs,但不增加测试时的计算开销,从而提升任务表现。实验结果表明,该方法在九个基准上的六种模型上实现了2–6%的性能增益(含微调可达1–3.5%),且保持与现有方法相当的计算成本。

链接: https://arxiv.org/abs/2611.21459
作者: Bryan A. Plummer,Shan Lu,Keval Doshi,Noranart Vesdapunt,Wei Wang,Jim Thomas,Kah Kuen Fu
机构: Amazon(亚马逊)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Weight Factorization (WF) methods aim to reduce a model’s latency, memory, or both. These typically decompose a pretrained model into two or more factors that approximate the original model weights with fewer parameters and floating point operations (FLOPs). However, translating fewer FLOPs into reduced latency is challenging for WF methods in practice due to overhead introduced by replacing a single large weight matrix with multiple smaller ones. To address this, we propose Balanced Twin Weight Factorization (BTWF), which uses the unique structure of factorized models to simultaneously lower latency and increase learning capacity. Specifically, BTWF operates in the shared basis WF setting, where two or more layers are decomposed into a shared basis matrix and a layer-specific mixing matrix. If these shared layers have the same input (i.e., are twins), then we need only perform this shared forward pass once, reducing inference time. However, twin layers are often of varying sizes, and methods from prior work tend to favor the larger layer. In contrast, our approach boosts performance by combining Factor Component Balancing (FCB), which ensures every shared member is well represented in the selected factors, with Block Contribution Balancing (BCB), which considers downstream model effects. Further, we also show that an overparameterized BTWF, where some layer’s factors result in higher FLOPs than the original during training, can boost task performance without additional resources at test time. Experiments on six models over nine benchmarks show BTWF obtains a 2-6% gain over the state-of-the-art without expensive finetuning (1-3.5% gain with it) at comparable computational costs.

[NLP-14] A Scalable Multi Task Learning Framework for Modeling Demographic Disagreements in Annotation Tasks

【速读】: 该论文旨在解决多群体标注者在标注任务中因不同人口统计学背景导致的系统性分歧问题,尤其关注在内容敏感性任务(如社交媒体毒性检测)中,如何合理权衡不同群体观点以提升模型性能。其核心挑战在于:传统方法往往忽略标注者群体间的差异性,或依赖于高隐私成本的个体级标注者信息,难以实现公平且有效的建模。本文提出的利益相关者感知联合多任务学习(Stakeholder-Aware Joint Multi-Task Learning, SAJ-MTL)框架的关键创新在于:通过条件梯度路由(conditional gradient routing)机制有效处理由群体条件标签引发的稀疏性问题,并支持在不同任务设置下对各群体进行可定制化的权重配置。该方法仅需标注者群体归属信息,无需敏感的个体层面数据,兼顾了模型性能与隐私保护。实验表明,SAJ-MTL通过增强对人口敏感内容的上下文感知能力,在毒性检测任务中显著提升了各群体的性能表现,同时具备良好的可扩展性和计算效率,适用于大规模实际部署。

链接: https://arxiv.org/abs/2611.21458
作者: Soumyajit Gupta,Maria De-Arteaga,Matthew Lease
机构: University of Texas at Austin (德克萨斯大学奥斯汀分校); Universitat Ramon Llull, ESADE (拉蒙·柳利大学,埃沙德商学院)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Annotators from different demographic backgrounds often disagree in systematic ways due to varying lived experiences, cultural contexts, and their relationship to the content. When such disagreement occurs, one group’s perspective may sometimes warrant prioritization because they hold an epistemically advantaged standpoint or bear greater stake in the outcome. For example, when assessing toxicity of posts on social media, it may be important to prioritize a targeted group’s viewpoint. We propose a Stakeholder-Aware Joint Multi-Task Learning (SAJ-MTL) framework to jointly model demographic disagreement both across- and within- annotator groups. The model uses conditional gradient routing to handle sparsity induced by group-conditional labeling, and parameterization supports customized weighting of groups in different task settings. SAJ-MTL improves group-level performance in toxicity detection through better context-aware understanding of demographic-sensitive content. Only annotator group-membership information is required, which is more practical than existing methods requiring more privacy-invasive individual-level annotator details. Our scalable architecture is also computationally efficient to support large-scale deployment.

[NLP-15] Replicating and Extending Weight-Space Programming: Frontier LLM s Compile Programs into Weights and Read Input Layers

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)在生成式人工智能(Generative AI)背景下对小型神经网络权重的直接生成与解读能力问题,探索其在模型可解释性(interpretability)和安全性审计中的潜力。其核心挑战在于:如何使LLM不仅能够理解并生成神经网络的代码表示,还能直接操作网络权重以实现特定任务,并从权重中推断出模型的训练目标与行为逻辑。解决方案的关键在于揭示了两个突破性能力:一是将程序代码“编译”为可执行的神经网络权重(即“程序到权重的翻译”),二是从权重中提取任务语义信息的能力。研究发现,尽管如GPT-6 Astra和Claude Opus 5.5等前沿模型能够成功生成具备功能的小型神经网络,且部分模型可逆向还原网络行为至代码级,但其性能仍受限于输入层信息的读取,难以全面理解任务机制;同时,在无代码执行的前提下,虽能检测到部分恶意组件(如中毒模型或恶意适配器),却无法识别触发条件或具体恶意行为。因此,当前技术本质仍为对人工设计程序的高效编译与输入结构的浅层解析,尚处于可解释性工具的初级阶段,具有应用前景但亟待深化。

链接: https://arxiv.org/abs/2611.21449
作者: Roksana Goworek,Alexander Hickey,Federico Nanni,Martin O’Reilly
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Researchers at MIT recently reported that GPT-6 Astra, a frontier large language model (LLM), can write the weights of small neural networks directly in a single reply, and can read the raw weights of small networks to say what they were trained to do. In a three-day sprint, we replicated and extended these results. Our four findings: (1) Writing weights reproduces. Astra’s networks for the game Breakout scored 514, 550 and 307 (versus MIT’s 517), and scored just as well on two modified versions of Breakout. Claude Opus 5.5 can also write working networks, but often runs out of output tokens. (2) What is new is translating programs into weights. Earlier models could write a good policy as code a generation before they could write network weights, and the networks the models write behave like compiled programs: two models independently decompiled Astra’s Breakout network into Python matching its actions in all tested states. On classifying handwritten digits, Astra’s networks scored 50% on average, below a simple classifier (82%). (3) Reading weights is mostly reading the input layer. Astra and Opus 5.5 identify which game a network plays almost perfectly, but on three new games we designed and trained ourselves, Astra recovered where things are on the board, while mostly missing how the game works. (4) Auditing from weights alone is possible but shallow. Without executing code, Astra picked out a small language model poisoned during training in 10 of 10 runs, one poisoned by extra training afterwards in 7 of 10, and a malicious adapter in 23 of 32 cases, but never identified the trigger or malicious behaviour. Open-weight models lag behind. From this early study, we conclude that LLMs can read and write small networks, but that this currently amounts to compiling hand-designed programs and reading the layout of the input layer. It is a promising but early tool for interpretability.

[NLP-16] Just sort it! A note on the LP behind rank-one anti-loop edits

【速读】: 该论文旨在解决小型推理模型在复杂任务中因“思维循环”(thinking loop)导致的推理效率低下问题,提出通过低秩更新(rank-one edit)对多层感知机(MLP)输出矩阵进行修正,以实现无需迭代的快速推理优化。其核心解决方案的关键在于:传统方法依赖稀疏线性规划(sparse Linear Program, LP)求解神经元激活的最优权重分配,而本文提出用简单的排序策略替代该复杂优化过程——即选取激活系数变异最小的8个神经元,并均分原LP方案所实现的“推动量”(push)。实验结果表明,在Qwen3.5-0.8B模型上,该排序方法在MATH-500与GSM8K共1000道题上的准确率达67.0%,与基于LP的编辑方案(67.4%)差异不显著(95%置信区间[-2.9, +2.1],低于预设5分非劣效边界),且在多个采样种子下表现稳定。进一步分析显示,排序选择的神经元与LP解共享了绝大部分平均推动量,表明该方法虽未求解LP,却能近似实现沿方向$ u $的恒定偏置效应,从而在保持性能的同时大幅降低计算开销。

链接: https://arxiv.org/abs/2611.21432
作者: Matteo Fischetti
机构: DEI, University of Padova (帕多瓦大学), Italy
类目: Machine Learning (cs.LG); Computation and Language (cs.CL); Optimization and Control (math.OC)
备注: 4 pages, 3 tables. Data, code and protocols: this https URL

点击查看摘要

Abstract:Weight-space programming removes the thinking loops of a small reasoning model by a rank-one edit W’ = W + u v^T of an MLP output matrix, where u is a loop direction and v is the solution of a sparse Linear Program (LP) over the neuron activations. We investigate whether the LP is needed. We replace it by a sort: we take the 8 neurons whose activation has the smallest coefficient of variation, and give them equal shares of the push that the LP solution realizes. On Qwen3.5-0.8B the sorted edit scores 67.0% on 1000 problems (MATH-500 and GSM8K) against 67.4% for the LP edit and 9.5% for the unedited model; the paired difference is -0.4 points, with 95% confidence interval [-2.9, +2.1], inside a non-inferiority margin of 5 points declared before the runs. On three further sampling seeds the difference is -0.6 points, interval [-2.0, +0.8], and -0.8 points, interval [-2.3, +0.6], when the dose is set without the LP. The sort does not solve the LP, but the neurons it shares with the LP solution carry most of its mean push. The outcome is that, on this model, the edit acts as a nearly constant bias along u.

[NLP-17] Rubric-Based AES with LLM Ensembles: Scoring of Student Responses in Critical Online Reasoning Using Ensembles of LLM s

【速读】: 该论文旨在解决传统学生作答评分中存在的成本高、耗时长以及人工评阅者易受偏见、疲劳和评分者漂移(rater drift)影响而导致评分可靠性下降的问题。其核心解决方案是探索基于大语言模型(Large Language Models, LLMs)的混合方法,通过构建多模型集成(ensemble of LLMs)来发挥各模型间的协同效应,从而提升自动化作文评分(Automated Essay Scoring, AES)系统的准确性与一致性。研究在德语语境下针对经济学与医学教育中的103项评分标准,利用631份学生作答及2至5名人工评阅者的标注数据,系统测试了15个开源权重LLMs在523种基于评分量规(rubric-based)的提示配置下的表现,并采用克里彭多夫阿尔法系数(Krippendorff’s alpha)和替代评阅者检验(alt-test)评估其与人工评分的一致性。结果表明,使用表现最优的LLM-提示组合替代一名人类评阅者,在91项标准上可维持或提升评分一致性,且在90项标准上通过了alt-test验证。研究证实,当前最先进的LLMs已具备支持德语学科作答自动评分的能力,可在显著降低人力负担并加速处理流程的同时实现可靠评分,但其性能仍需持续以人工评分作为基准进行验证。

链接: https://arxiv.org/abs/2611.21422
作者: Daniel Baumartz,Alexander Mehler,Patrick Schrottenbacher,Leon Hammerla,Sebastian Gombert
机构: Goethe University Frankfurt(歌德大学法兰克福); DIPF Leibniz Institute for Research and Information in Education(德国教育研究与信息研究所)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Grading student responses is costly and time-consuming. Moreover, human annotators are prone to bias, fatigue, and rater drift, which further limits grading reliability. Various methods have been employed to develop automated essay scoring (AES) systems to reduce these costs and address bias-related issues. These methods range from early text-feature-based systems to current machine-learning and transformer-based models. Specifically, large language models (LLMs) offer the potential to streamline the development of AES systems. However, they are subject to well-known limitations, such as hallucinations and prompt sensitivity. This creates a need to explore the potential of a mixed-methods approach that leverages the synergies among individual LLMs through ensemble formation. We evaluate the potential of ensembles of LLMs to match human-rater agreement when grading student responses in economics and medical education across 103 criteria. We use 631 German-language responses and corresponding human ratings by 2 to 5 annotators from a previously conducted assessment. We experiment with 523 rubric-based prompt configurations across 15 open-weight LLMs and assess their consistency with human raters using Krippendorff’s alpha for inter-annotator agreement (IAA) and the Alternative Annotator Test (alt-test). Our results show that replacing one human rater with the best-performing LLM–prompt configuration maintains or improves IAA for 91 of the 103 criteria and that LLMs pass the alt-test for 90 criteria. Our findings suggest that current state-of-the-art LLMs can support AES of German responses in economics and medical education, reducing human workload and enabling faster processing, while validation against human ratings remains necessary.

[NLP-18] AuXSpec: Learning to Autoformalize Program Specifications via Executable Postconditions

【速读】: 该论文旨在解决形式化规范(formal specification)在定理证明器(如Lean)中编写缓慢且高度依赖专业知识的问题,从而导致规范成为实现形式正确代码规模化的主要瓶颈。其核心挑战在于如何高效获取训练生成式规范模型所需的可扩展反馈,以判断生成的规范是否准确捕捉了预期行为。以往方法依赖于显式证明或单独实现代码并建立形式化对应关系来提供反馈,成本高昂。本文提出AuXSpec框架,其关键创新在于将声明式后置条件(declarative postconditions)与可执行的对应实现配对:当可能时通过类型类解析(type class resolution)自动生成可执行版本,否则独立定义并提供等价性证明。这一设计使得测试用例可通过执行来验证声明式规范,进而利用执行结果的一致性作为强化学习的奖励信号。基于此机制,作者在LeetCode上合成3,037个问题,并通过Codeforces上的专家迭代将训练数据扩展至19.5K样本。实验表明,使用该执行反馈奖励训练Qwen3.5-9B和Qwen3.8-27B模型,在AuXSpec-Test基准上,SpecGen pass@4分别从0.7%提升至82.7%和14.3%提升至84.3%,其中27B模型性能超越所有现有前沿模型。在跨域的Verina基准测试中,两模型在pass@1和pass@4上均显著优于此前最优开源模型VeriSpecGen-30B-A3B,分别达到73.7%和84.5%(pass@1)、90.5%(pass@4),验证了方法的有效性与泛化能力。研究团队已公开训练后的9B与27B模型权重及全部3,037个带测试用例的问题数据集,包括经专家评审与正确性证明的300题AuXSpec-Test基准,为后续研究提供重要资源。

链接: https://arxiv.org/abs/2611.21421
作者: Haocheng Wang,Weixuan Yuan,Florian Hübler,Yinya Huang,Zhijiang Guo,Sorrachai Yingchareonthawornchai,Rasmus Kyng
机构: ETH Zurich (苏黎世联邦理工学院); HKUST(GZ) (香港科技大学(广州)); HKUST (香港科技大学); ETH AI Center (ETH人工智能中心); Institute for Theoretical Studies, ETH Zurich (ETH理论研究所)
类目: oftware Engineering (cs.SE); Computation and Language (cs.CL)
备注: 62 pages

点击查看摘要

Abstract:Formal specifications define intended program behavior and underpin verification, but writing them in theorem provers such as Lean is slow and expertise-intensive, making specification a major bottleneck to scaling formally correct code. Autoformalization from natural-language problems is promising, yet training specification models requires scalable feedback on whether generated specifications capture intended behavior. Prior approaches obtain this feedback through explicit proofs or separately implemented code with formal correspondence proofs, both expensive. We introduce AuXSpec, a Lean framework that pairs declarative postconditions with executable counterparts, obtained through type class resolution when possible and otherwise defined separately with an equivalence proof. Test cases can then evaluate declarative specifications by execution, letting us use agreement as a reinforcement-learning reward. Using AuXSpec, we synthesize 3,037 LeetCode problems and expand training to 19.5K samples through expert iteration on Codeforces. Experiments show that training Qwen3.5-9B and Qwen3.8-27B with the execution-based reward improves SpecGen pass@4 on AuXSpec-Test 0.7% \to 82.7% and 14.3% \to 84.3%, and the 27B model outperforms all frontier models. On the out-of-domain Verina benchmark with its original prompts, both models improve 7.0% \to 90.5% and 21.9% \to 90.5% and surpass VeriSpecGen-30B-A3B, the prior state of the art among open-source specification models, on pass@1 (theirs 47.6% vs. ours 73.7% and 84.5%) and pass@4 (theirs 63.9% at pass@10 vs. ours 90.5%). We release the weights of the trained 9B and 27B models and the 3,037 problems with test cases, including the 300-problem AuXSpec-Test benchmark with expert review and correctness proofs for future evaluation. Models and data: this https URL. Website: this https URL

[NLP-19] Do Embedding-Map Distances Guide LLM Ideation? Four Tests Against What Came Next

【速读】: 该论文旨在解决生成式创新中如何有效利用已有工作嵌入空间(embedding map)来指导新想法生成的问题,核心挑战在于如何识别真正具有创新潜力的“空白区域”(gaps),而非仅依赖表面距离。其解决方案的关键在于:通过多维度评估嵌入空间中的“远点”是否具备实际创新价值,发现单纯基于距离(如余弦相似度)的随机或聚类远点往往无法生成有意义的新文本,而基于上下文连贯性与真实后续研究反馈的迭代式生成策略(如LLM续写、评审反馈与方向修正)能显著提升生成内容的相关性与创新性。特别地,研究揭示了真实后续研究标题对生成方向的微调作用(+0.032),以及评审、预言者与距离反馈在引导生成方向上的累积效应(最大可达0.70上界),表明有效的创新引导需融合语义连贯性、外部反馈与动态修正机制,而非静态距离度量。

链接: https://arxiv.org/abs/2611.21408
作者: Javier Aguilar Martín
机构: AGILabs(AGILabs)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Digital Libraries (cs.DL)
备注: 48 pages, 5 figures. Code, data pointers and pre-registrations: this https URL

点击查看摘要

Abstract:A popular ideation recipe treats an embedding map of existing work as a guide: pick points far from everything and decode them into ideas. I test four ways such distances could guide ideation against what appeared later. (1) Far random directions are not text: no later item reaches a raw cosine above about 0.10 with one (after centring or whitening, 22 of 24 cells stay out of reach), and an LLM decoder shown a target’s neighbours, with score-only feedback, writes those neighbours in all six target sets. A neighbour-free search writes coherent text at median cosine 0.57 and 0.47 to interior gaps, beating the nearest existing item on 15 and 14 of 20, but only 0.07 and 0.06 to far points. (2) On k-means cells of a past-only map, the registered hole score mostly tracks cell size and position; three density-based gap scores forecast which cells grow fastest at or below chance (AUC 0.30 to 0.66) and add nothing to momentum in Steam and arXiv, with patents inconclusive; more isolated new Steam games are slightly better reviewed. (3) LLM continuations of arXiv papers stay closer to the source than papers that cite it, under three user prompts and against proposal rewrites (gpt-4o) and April-December 2023 papers (OLMo). Critique lengthens this step beyond content-free revision, replicated on 150 new papers. Revealing a real follow-up’s title turns the ideas by 0.032 along a follow-up-specific direction beyond revision; reviewer, oracle and distance feedback add 0.13, 0.08 and 0.34 of that turn (upper bounds 0.46, 0.46, 0.70), and their turn is not separable from revision or their longer step. (4) For inspiration material graded by distance, no inverted U was detected, on an anticipation margin that tracks format more than content. Original tests were registered before running, follow-ups (thirteen requested in two AI-agent reviews) after the main results; all are dated only by local git commits.

[NLP-20] rain After You Measure: a Pre-Registered Learning Curve for Fine-Tuning Typed Decision Models on One Users Labels

【速读】: 该论文旨在解决小规模有类型决策模型在单用户邮件标注数据上微调后,是否能够缩小与基于相同标注数据训练的逻辑回归基线模型之间的性能差距这一问题。其核心挑战在于验证微调模型能否在有限标注预算下实现对基线模型的超越,尤其是在消息类型(message type)和动作开关(action switches)等关键任务上的表现。解决方案的关键在于采用注册的主实验设计:在固定测试集(135封邮件)上,使用公开的Tev1-0.8B和Tev1-4B决策模型,通过LoRA适配器进行微调,并在50至587个标签的嵌套子集、双种子设置下评估不同标注预算下的性能。结果显示,在消息类型任务中,4B模型在全部587个标签时达到+3.3分的优势,显著超越基线,且在宏平均F1得分上表现更优(0.581 vs. 0.406),表明大规模模型结合充分标注可有效提升分类能力;而在动作开关任务中,尽管基线模型以极低延迟(<1毫秒/问题)保持领先,微调模型未能追上,且存在类别覆盖不一致的问题(如罕见类别的识别不稳定)。总体而言,解决方案的关键在于利用大规模基础模型与高效微调技术(LoRA)在充足标注条件下逼近甚至超越传统方法,但其有效性高度依赖任务类型与标注量。

链接: https://arxiv.org/abs/2611.21393
作者: Federico Lucca
机构: 未知
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 9 pages, 1 figure, 6 tables. Companion to arXiv:2610.32955 . Pre-registered study; code at this https URL (tag v2.4; registrations in prereg/); weights and data not released

点击查看摘要

Abstract:Fine-tuning a small typed decision model on one user’s email labels did not clear the margin we pre-registered over a logistic regression trained on the same labels. In earlier work, about two hundred labels and a TF-IDF logistic regression had the highest point accuracy on this inbox, though not significantly above the best hosted model; we ask whether the same labels, given to a decision model, close the gap. A first registered study, on a generic 4B base with cross-validation and message type only, stayed below the baseline at every label budget (-1.8 to -7.3 points). The main study was registered before any fine-tuning run: a fixed test set of 135 emails never used for selection, LoRA adapters on the public Tev1-0.8B and Tev1-4B decision models, nested subsets of 50 to 587 labels, two seeds and five decision fields on a free T4 GPU. Four of its five hypotheses are refuted–two in the fine-tunes’ favour–and one is confirmed for one model. On message type the 4B model misses the registered three-point margin at 400 labels (+2.6) and passes it with all 587 (+3.3); with all labels both models are above the baseline by +0.7 and +3.3 points, where one email is 0.74 points and a 95% interval is about 15 points wide, and the 4B model is clearly ahead in type macro-F1 (0.581 vs. 0.406). On the three action switches the baseline is ahead at every label budget for both models (0.931 vs. 0.904 and 0.899 mean accuracy with all labels), at under a millisecond per question against 306 and 1681 ms. The 0.8B type curve barely improves after 300 labels and then falls. The baseline never predicts the two rarest classes (7 offers and 8 spam in the training pool); the fine-tuned models find spam in some runs, not consistently across seeds. We list the explanations we did not test and say when a new attempt would be worth making. Scripts are public; weights and emails are not.

[NLP-21] Stability-Compensated Threshold Calibration for Selection-Conditioned Risk Control in Large Language Models

【速读】: 该论文旨在解决大语言模型在生成回答时产生幻觉(hallucination)的问题,尤其关注如何通过选择性预测策略(selective prediction policy)在不确定的查询上主动放弃预测,从而降低被接受预测中的错误率。现有方法如线性期望约束(Linear Expectation Constraints, LEC)虽能将选择条件下的错误率(SCER)校准转化为非正的期望线性损失约束,但其关联的损失函数在接纳阈值上呈非单调性,导致经验阈值选择规则在独立同分布(i.i.d.)采样下可能违反总体SCER目标,存在风险转移缺陷。为弥合这一风险转移差距,论文提出稳定性补偿线性期望控制(Stability-Compensated Linear Expectation Control, SC-LEC),其核心在于利用理论全数据参考下的负线性风险来补偿阈值选择过程中的不稳定性。针对安全锚点(safe anchor)之外的局部扩展,研究推导出两种有限样本保证:基于秩交叉的稳定性证书与利用阈值扩展结构的直接证书,以降低保守性。在具备有效总体锚点裕度的前提下,所提出的证书条件可确保有限样本下的边际SCER控制,并通过独立裕度认证进一步提供1−δ高概率的校准平均控制保障。实验在CommonsenseQA、ARC-Challenge和TriviaQA数据集上的结果表明,相较于相同认证锚点,该方法实现了更优的保留率增益,但在小裕度区域仍表现出一定保守性。

链接: https://arxiv.org/abs/2611.21384
作者: Zhiyuan Wang,Yuanchang Ye,Jiahao Yuan,Jinhao Duan,Xiaoshuang Shi,Weijun Xie,Ye Wei,Kaidi Xu
机构: City University of Hong Kong(香港城市大学); National University of Singapore(新加坡国立大学); East China Normal University(华东师范大学); University of North Carolina at Chapel Hill(北卡罗来纳大学教堂山分校); University of Electronic Science and Technology of China(电子科技大学); The Hong Kong Institute of AI for Science(香港人工智能科学研究所); City University of Hong Kong(香港城市大学); The Hong Kong Institute of AI for Science(香港人工智能科学研究所)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Large language models often produce hallucinated answers, motivating selective prediction policies that abstain on uncertain queries. However, heuristic uncertainty thresholds alone do not provide statistical control of errors among accepted predictions. Linear Expectation Constraints (LEC) reduce selection-conditioned error rate (SCER) calibration to a nonpositive expected linear-loss constraint, but the associated loss is non-monotonic in the acceptance threshold. We construct an exact counterexample under i.i.d. sampling demonstrating that LEC’s empirical threshold-selection rule can violate the population SCER target. To address this risk-transfer gap, we introduce Stability-Compensated Linear Expectation Control (SC-LEC), which uses the negative linear risk of a theoretical full-data reference to compensate for threshold-selection instability. For local expansion beyond a safe anchor, we derive two finite-sample certificates: a stability-based certificate exploiting rank crossings and a direct certificate using the structure of threshold expansion to reduce conservatism. With a valid population anchor margin, our certificate conditions ensure finite-sample marginal SCER control, and independent margin certification further provides a 1-\delta high-probability guarantee for this calibration-averaged control. Experiments on CommonsenseQA, ARC-Challenge, and TriviaQA show useful retention gains beyond the same certified anchor, while revealing conservative behavior in small-margin regimes.

[NLP-22] What You Post Is What You Symptom: Estimating Psychometric Networks from Social Media Discourse

【速读】: 该论文旨在解决传统心理病理学症状网络研究过度依赖自评量表所导致的推断局限性问题,即研究结果受限于特定量表涵盖的症状集合。其解决方案的关键在于提出一种直接从社交媒体文本中进行症状解析的网络估计框架,首次实现了基于社交媒体语料的、症状级分辨率的心理病理网络构建。该方法在抑郁症和进食障碍两个领域进行了应用,并通过第三个与厌食症相关的独立语料库进行跨数据集验证。研究发现,抑郁症相关表达得分与贝克抑郁量表Ⅱ(BDI-II)总分呈显著正相关(ρ=0.434),21个条目中有20个达到错误发现率(FDR)校正后的统计显著性,分半信度高达0.935;网络比较检验显示不同模态间仅有3–5%的边存在差异,但桥接症状表现出明显分化。对于进食障碍,尽管条目层面结果具有探索性,但识别出的语言特征在三个独立语料库中均呈现一致的方向性,并聚类为三个构念类别,验证了跨数据集的可重复性。该框架为实现跨诊断的语言学标记物复制提供了新范式,突破了传统量表依赖的限制。

链接: https://arxiv.org/abs/2611.21362
作者: Ştefana-Arina Tăbuşcă(1),Ana Sabina Uban(2,3),Liviu P. Dinu(2,3) ((1) Interdisciplinary School of Doctoral Studies, (2) Faculty of Mathematics and Computer Science, (3) Human Language Technologies Research Center, University of Bucharest, Romania)
机构: Interdisciplinary School of Doctoral Studies; Faculty of Mathematics and Computer Science; Human Language Technologies Research Center, University of Bucharest, Romania
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Social and Information Networks (cs.SI)
备注:

点击查看摘要

Abstract:Symptom networks in psychopathology are estimated almost exclusively from self-report questionnaires, restricting inference to whatever symptom set a given instrument covers. We introduce a framework for symptom-resolved network estimation directly from social media discourse and apply the framework to depression and eating disorders, with a third anorexia corpus as cross-dataset validation. To our knowledge, we are the first to estimate symptom-resolved psychometric networks from social media text, compare them to clinical co-endorsement networks, and replicate transdiagnostic linguistic markers across 3 independently recruited corpora spanning 2 disorders simultaneously. For depression, expression scores correlate with BDI-II totals at \rho=0.434 , 20 of 21 items reach FDR significance, and split-half reliability is 0.935. The Network Comparison Test detects few edge differences between modalities (3-5% of edges), while bridge symptoms diverge. For eating disorders, the item-level results are exploratory. Identified linguistic features replicate the same direction across all 3 corpora, and they cluster in three construct families.

[NLP-23] Syllable- and Word-Level Evaluation of Tibetan OCR

【速读】: 该论文旨在解决藏文光学字符识别(OCR)评估中缺乏统一、多层次误差度量标准的问题,尤其针对藏文不标记词边界而仅标记音节边界的语言特性,导致传统字符错误率(CER)无法充分反映实际应用中的语义理解性能。其核心挑战在于如何在无显式词边界标注的情况下,有效评估系统在词级层面的识别准确性。解决方案的关键在于首次系统性地引入音节错误率(SER)与三种分词后词错误率(SWER),并结合联合训练、后训练量化及多模态模型等九种识别器配置,在印刷体、木刻体及两种手写体风格数据上进行综合评估。研究发现,SER无需依赖外部分词器且与SWER具有高度一致性(Kendall τ ≥ 0.889),可作为更可靠、更高效的替代指标;同时,通过分析部分词性标注、命名实体识别与机器翻译质量与误差率之间的关联,揭示了不同误差阈值对下游自然语言处理任务的影响:当SER ≤ 0.047时,机器翻译的chrF得分可达60;当SER ≤ 0.078时,部分词性标注保留率可达0.8;而命名实体保留率则未表现出明确的误差率分界点。这一成果为藏文OCR系统的评估提供了可扩展、可解释的基准框架。

链接: https://arxiv.org/abs/2611.21303
作者: Jacob Moore,Paula Lauren
机构: Lawrence Technological University (劳伦斯科技大学)
类目: Computation and Language (cs.CL)
备注: 21 pages, 4 figures

点击查看摘要

Abstract:Optical character recognition (OCR) for Tibetan is evaluated predominantly by character error rate (CER). Tibetan marks syllable boundaries but does not mark word boundaries, so a word error rate requires a segmenter. This study evaluates nine recognizer configurations on typeset, woodblock and two handwritten script styles, alone and in combination, and reports syllable- and word-level error rates, to our knowledge the first for Tibetan OCR. The configurations span CTC models, TrOCR encoder-decoders with and without post-training quantization, and a vision-language model, and each is scored under CER, the syllable error rate (SER) and three segmented word error rates (SWER). CER ranks systems largely as the higher-level rates do (Kendall \tau \geq 0.833 ), with one pair ordered oppositely. On the effect of joint training, CER and SER record significant changes of opposite sign in 7 of 64 comparisons. SER needs no segmenter and ranks systems as SWER does under two of the three segmenters ( \tau \geq 0.889 ). The output of each configuration is also tagged for part of speech, annotated for named entities and machine translated, and each result is compared with that obtained on the reference text. As rough guidance, translation chrF of 60 is reached at SER of 0.047 or below and part-of-speech preservation of 0.8 at 0.078 or below, while no error rate separates the systems that reach entity preservation of 0.6 from those that do not.

[NLP-24] Latent Bridge: Asking for Help Without Words

【速读】: 该论文旨在解决文本生成中基于模型协作(text-based model collaboration)所导致的高推理开销问题,即传统方法需依赖大型模型生成中间响应以指导小型模型,从而引入大量大型模型的自回归解码成本。其核心解决方案是提出POND(Prefill-Only, No-Decode)框架:该框架采用一个冻结的大型发送者(sender)仅在预填充(prefill)阶段参与计算,而整个自回归解码过程均由小型接收者(receiver)完成。通过引入可训练的桥接模块(trainable bridge),将发送者的隐藏状态映射为额外的注意力键(key)和值(value),使接收者能够在生成过程中持续利用发送者的表征信息。此外,POND支持周期性审查机制,允许发送者在接收者生成部分响应后进行多次预填充更新,动态调整引导信号的同时保留接收者的现有键值缓存(KV cache)。实验表明,无论发送者来自何种模型家族,POND均能显著提升接收者在推理与编码任务上的性能;例如,GLM-5.3作为发送者时,与Qwen接收者搭配可达到与Qwen3.8-Max相当的效果,验证了跨独立训练模型家族间的表征迁移有效性。更重要的是,该方法具备良好的跨领域泛化能力——在数学与编程任务上训练后,无需额外微调即可直接迁移到通用知识任务,证明大型前沿模型的预填充表征可在不进行自回归解码的前提下有效增强小型模型的生成质量。

链接: https://arxiv.org/abs/2611.21301
作者: Qichao Wang,Tingyu Qu,Pengcheng Wu,Yiran Zhong,Feida Zhu,Zhiqi Shen,Peilin Zhao,Steven Hoi
机构: Nanyang Technological University(南洋理工大学); Alibaba Token Hub, Alibaba Group(阿里巴巴令牌中心,阿里巴巴集团); Shanghai Jiao Tong University(上海交通大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Text-based model collaboration requires a large model to generate intermediate responses before a smaller model can use its guidance, adding large-model decoding cost to inference. We introduce POND (Prefill-Only, No-Decode), a method that uses a frozen large sender for prefill and a small receiver for all autoregressive decoding. A trainable bridge maps the sender’s hidden states to additional attention keys and values, allowing the receiver to use these representations throughout generation. POND also supports periodic review, in which the sender processes the receiver’s partial response and updates its guidance through additional prefill passes while preserving the receiver’s existing KV cache. Across reasoning and coding benchmarks, POND improves receiver performance with senders from different model families. In particular, GLM-5.3 provides gains comparable to those of Qwen3.8-Max when paired with a Qwen receiver, supporting the effectiveness of representation transfer between independently trained model families. POND further demonstrates its ability to generalize across domains. When trained on math and coding tasks, it transfers directly to general knowledge tasks without additional domain-specific training. These results demonstrate that large frontier model prefill representations can improve small-model generation without requiring large-model autoregressive decoding.

[NLP-25] LATTE-WordSeg-Lite: Multi-Granularity Thai Word Segmentation on a CPU with a Reproducible Benchmark

【速读】: 该论文旨在解决泰语分词(Word Segmentation)中因生成式模型依赖子词单元而无法有效利用基于词级标注的树库、语音评估指标及形态资源所导致的语义歧义问题。现有方法在处理多粒度语言结构时存在信息丢失,且主流模型如LATTE虽能通过词典图谱(lattice)实现多粒度联合建模,但其计算开销大,难以部署于资源受限环境。为此,作者提出将大型GPU模型LATTE蒸馏为轻量级CPU可运行模型LATTE-WordSeg-Lite,其仅1.17M参数,保留词典图谱作为输入特征,可在单次前向传播中同时输出复合词、词和词素,并仅依赖许可允许商业使用的标注数据与文本进行训练。为公平比较,研究重新构建了七个测试集并统一评测协议,对十二个系统进行一致评估。结果表明,该模型在五个测试集上优于DeepCut,其中四个来自其教师模型训练所用语料;在单线程CPU环境下,推理速度达DeepCut的23倍。研究开源了模型、基准脚本及所有预测结果,推动泰语分词领域的可复现性与实际应用。

链接: https://arxiv.org/abs/2611.21300
作者: Thodsaporn Chay-intr,Kobkrit Viriyayudhakorn,Thanaruk Theeramunkong
机构: iApp Technology(泰国); OpenThai Laboratory(泰国); Intelligent Informatics and Service Innovation Research Center(泰国); Artificial Intelligence Entrepreneur Association of Thailand (AIEAT)(泰国); Sirindhorn International Institute of Technology, Thammasat University(泰国)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Large language models read Thai through subword tokens, without a word segmenter. Yet Thai treebanks, speech metrics and morphological resources are built on words, often at several granularities, and word boundaries resolve ambiguities that subword tokens leave open. LATTE (Lattice ATTentive Encoding) reads a text at several granularities at once, through a lattice of dictionary words, but it is a large GPU model. We distil it into LATTE-WordSeg-Lite, a CPU segmenter of 1.17M parameters that keeps the lattice as features. It outputs compounds, words and morphemes in one pass, and learns only from annotated data and text whose licences allow commercial use. To compare it with other tools, published Thai scores do not suffice, as they come from different files, splits and scorers. We therefore rebuild seven test sets from pinned sources and score twelve systems under one protocol. With the unit that fits each test set, LATTE-WordSeg-Lite scores above DeepCut on five sets, four of them from corpora that its teachers learnt from. Called line by line on one CPU thread, it runs 23 times as fast as DeepCut. We release the model, the benchmark scripts and all predictions.

[NLP-26] AnachroBench: Evaluating Historical Consistency in Large Vision-Language Models

【速读】: 该论文旨在解决现有大规模视觉-语言模型评估基准在历史一致性(historical consistency)方面评估不足的问题,尤其聚焦于特定时代与国家背景下的视觉场景历史准确性。其核心挑战在于:当前主流基准多关注日常通用场景的视觉理解能力,而对图像内容是否符合指定历史时期与地域特征的判断缺乏系统性评测。为此,研究提出AnachroBench——一个受控的判别式基准,用于评估视觉场景的历史一致性。该基准包含8个国家、19世纪末至20世纪初共1,600组图像对,每组由一张原始历史图像与一张嵌入了无缝现代物体的错位图像(anachronistic version)构成。评估涵盖三个维度:插入物体感知(Inserted Object Perception),检验模型能否识别出异常现代物体;时代-国家一致性(Era-Country Consistency),评估场景整体是否与指定时代和国家相符;条件时代-国家一致性(Conditional Era-Country Consistency),仅在正确感知插入物体的前提下评估一致性判断。关键发现为:模型在感知现代物体与做出历史一致性判断之间存在显著脱节,即出现“错误一致性”判断(false consistency judgment)现象——即模型虽察觉到现代物体,却仍判定场景历史一致。这一现象揭示了当前模型在历史语境推理中的局限性,凸显了AnachroBench作为针对性评估工具的重要性。

链接: https://arxiv.org/abs/2611.21292
作者: Keunhyeung Park,Byeonggeuk Lim,Kahyeon Kim,Sunhee Jo,Hayeong Ryu,Hoejoon Kwon,Jinhee Jang,YoungBin Kim
机构: Chung-Ang University (中央大学); Graduate School of Advanced Imaging Science, Multimedia Film, Chung-Ang University (先进影像科学与多媒体电影研究生院,中央大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
备注: 18 pages, 14 figures

点击查看摘要

Abstract:Existing Large Vision-Language Model benchmarks primarily evaluate visual understanding in general everyday scenarios, leaving historical consistency with a specified era and country largely underexplored. To address this gap, we propose AnachroBench, a controlled discriminative benchmark for evaluating historical consistency in visual scenes. It contains 1,600 image pairs from eight countries spanning the late 19th to the early 20th century, each pairing an original historical image with an anachronistic version containing a seamlessly inserted modern object. AnachroBench evaluates three aspects: Inserted Object Perception, which tests whether models detect the inserted modern object; Era-Country Consistency, which assesses whether the scene is consistent with the specified era and country; and Conditional Era-Country Consistency, which measures consistency judgments only when the inserted object is correctly perceived. A false consistency judgment occurs when a model perceives the modern object but still judges the scene as historically consistent. Overall, AnachroBench reveals a gap between inserted object perception and era-country consistency judgments, providing a targeted evaluation of this behavior.

[NLP-27] Benchmarking Large Language Models for Linguistic Annotation

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在语言学标注任务中缺乏针对具体任务、在可重复且透明条件下进行评估的问题。现有评估往往依赖于通用性能排名,难以反映模型在特定语言学任务中的实际表现,导致评估结果与实际应用需求脱节。为此,论文提出了一种名为dhAIBench的Python基准测试与标注工具,其关键创新在于提供一个具备网页界面的系统化框架,能够记录提示(prompt)、配置参数和预测结果,从而实现对标注质量、置信度校准、可重复性及估算成本的多维度比较。通过在9个语言学标注任务(涵盖形态学、句法、语义、语用-语篇、拼写规范化及学习者语言纠错)上对48个模型执行293次运行的实证研究,验证了该框架的有效性。实验表明,部分模型在非标准化或历史语料、需上下文理解的任务中仍能与参考标注达成较高一致性。研究强调应基于具体语言学任务而非通用排名来评估模型性能,并将当前及未来基准数据集公开共享,以推动可复现、可比较的语言学标注研究。

链接: https://arxiv.org/abs/2611.21291
作者: Ondřej Tichý(1),Anna Marklová(1),Barbora Bulantová(1),Magdalena Titlbachová(1),Jiří Milička(1) ((1) Faculty of Arts, Charles University)
机构: Charles University (查理大学); Faculty of Arts (艺术学院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Assessing large language models for linguistic annotation requires task-specific evaluation under documented conditions. We present dhAIBench, a Python benchmarking and annotation tool with a web interface for exploring and sharing results. It records prompts, configurations and predictions, enabling comparisons of annotation quality, confidence calibration, repeatability and estimated cost. We demonstrate the framework and the associated methodology through 293 runs of 48 models on nine linguistic annotation tasks spanning morphology, syntax, semantics, discourse-pragmatics, spelling normalization and learner-language correction. Models achieve high agreement with reference annotations on several tasks, including those on non-standardized, historical varieties as well as requiring contextual interpretation. The results support evaluating models against specific linguistic tasks, rather than relying on general rankings. To this end, we make the current and future benchmarks publicly available at this http URL.

[NLP-28] Classical structural priors improve neural molecular representations in molecule language models

【速读】: 该论文旨在解决当前分子-语言模型在训练过程中未能有效遵循化学结构渐进相似性(graded structural relationships)这一关键问题,即基于配对结构-文本数据训练的模型所生成的分子嵌入(molecular embeddings)往往缺乏与真实化学结构相似性一致的分布。其解决方案的关键在于提出一种仅在训练阶段使用的正则化方法——结构感知引导对齐(Structure-Aware Guided Alignment, SAGA),该方法利用中心核对齐(Centered Kernel Alignment, CKM)将神经网络生成的分子表示与基于指纹的塔尼莫托相似性(Tanimoto similarity)进行对齐,从而强制模型学习更符合化学常识的结构关系。SAGA不引入额外参数,且在推理阶段无需依赖指纹信息,既保持了模型在多模态任务中的通用性,又显著提升了分子-文本检索性能、结构邻近性相关任务的表现,并优化了分子属性预测与文本生成能力。结果表明,将传统化学信息学先验知识融入深度学习框架,可有效提升分子表示的化学可信度,同时维持其跨模态应用价值。

链接: https://arxiv.org/abs/2611.21276
作者: Donghyeok Lee,Sangkwon Park,Munyoung Chang,Sungroh Yoon
机构: Seoul National University (首尔国立大学); Seoul National University Bundang Hospital (首尔国立大学仁川医院)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 40 pages. This is the originally submitted version. The final version, revised after peer review, is published in J. Chem. Inf. Model. (2026): this https URL

点击查看摘要

Abstract:Classical molecular similarity is central to chemical information systems, yet molecule-language models trained on paired structure-text data do not necessarily organize molecular embeddings according to graded structural relationships. Here we introduce Structure-Aware Guided Alignment (SAGA), a training-only regularizer that uses Centered Kernel Alignment to align neural molecular representations with fingerprint-based Tanimoto similarity, without adding parameters or requiring fingerprints at inference. In MolCA and Atomas, SAGA improves molecule-text retrieval, shifts non-exact retrieval toward closer structural analogues, and increases average scaffold-split property-prediction performance, while molecular description and text-to-molecule generation are improved or preserved in single-run ablations. These results show that established chemoinformatics priors can make molecule-language representations more chemically faithful while retaining their multimodal utility.

[NLP-29] he Price of Choosing Examples: Context Overfitting in Adaptive In-Context Learning NEURIPS2026

【速读】: 该论文旨在解决生成式 AI(Generative AI)在实际部署中“上下文学习”(in-context learning)的泛化能力评估与理论分析之间的关键差距问题。传统分析假设提示(prompt)中的示例是预先独立采样后固定输入给模型,而现实系统中,示例是在观察查询和数据存储后通过检索、过滤、重排序、编辑及反复测试等自适应流程动态构建的,这一过程引入了数据依赖性与选择偏差。为填补此差距,作者提出将提示构造建模为条件随机核(conditional stochastic-kernel)驱动的自适应数据分析,并定义“上下文容量”(context capacity)作为数据存储与最终提示文本之间互信息或近似最大信息量的度量。对于任意固定的上下文学习器与任意具有容量 κ\kappa 的自适应上下文处理流程,论文证明了一个有限样本下的上下文泛化误差上界,其阶为 κ/k\sqrt{\kappa}/k,其中 kk 为所选示例数量,其损失构成报告得分。进一步,作者建立了匹配的下界,且当 log⁡k≪κ≪k\log k \ll \kappa \ll k 时常数项精确,据此推导出两种 PAC-Bayesian 检索规则:一阶 Gibbs 规则对数据依赖的候选门控赋予先验权重,二阶自归一化规则则具备更强的稳定性。在 Banking77 数据集上,基于 64 种检索策略与两个模型族的实验表明,24,000 次重采样事件中证书均成立;从 64(Qwen2.5-7B)或 128(GPT-4.1-mini)个重复验证查询出发,Gibbs 规则在超过 99% 的验证抽样中保证风险低于零样本误差。最后,理论被应用于贝叶斯线性上下文回归,包括对查询对齐的 top-kk 高斯检索的顺序统计分析。研究结果明确区分了“上下文质量”与“上下文认证”的关系:仅当选择过程足够稳定,使得所选上下文证据可有效迁移时,更长或更相关的内容才能真正降低风险。

链接: https://arxiv.org/abs/2611.21251
作者: Xiaoyu Li,Jiangxuan Long
机构: University of New South Wales (新南威尔士大学); The University of Hong Kong (香港大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL); Machine Learning (stat.ML)
备注: Accepted at NeurIPS 2026. Xiaoyu Li and Jiangxuan Long contributed equally. 31 pages, 8 figures, 5 tables

点击查看摘要

Abstract:In-context learning is usually analyzed as if the examples in the prompt were sampled before the learner is chosen. In deployed systems they are not: examples are retrieved, filtered, re-ranked, edited, and repeatedly tested after seeing the query and the data store. We formalize this gap by treating prompt construction as adaptive data analysis through a conditional stochastic-kernel calculus. Our main quantity is the context capacity, the mutual information or approximate max-information between the data store and the final prompt transcript. For any frozen in-context learner and any adaptive context pipeline with capacity \kappa , we prove a finite-sample context generalization bound of order \sqrt\kappa/k , where k is the number of selected examples whose losses form the reported score. We prove a matching lower bound, whose constant is exact when \log k\ll\kappa\ll k , and derive two PAC-Bayesian retrieval rules: a first-order Gibbs rule that charges data-dependent candidate gates their prior mass, and a self-normalized second-order rule. On Banking77 with 64 retrieval recipes and two model families, the certificates hold in all 24,000 resampled selection events, and from 64 (Qwen2.5-7B) or 128 (GPT-4.1-mini) reused validation queries the Gibbs rule certifies a risk below the zero-shot error in over 99% of validation draws. Finally, we instantiate the theory for Bayesian linear in-context regression, including an order-statistics analysis of query-aligned top- k Gaussian retrieval. The results separate context quality from context certification: longer or more relevant contexts improve risk only when the procedure used to choose them is stable enough for selected-context evidence to transfer.

[NLP-30] Anytime-Valid PAC-Bayes Certificates for Adaptive Test-Time Scaling NEURIPS2026

【速读】: 该论文旨在解决测试时缩放(test-time scaling)系统在自适应选择采样量、提示模板、验证器或聚合规则后,传统验证误差条不再有效的问题。其核心挑战在于:在反复检查校准集以优化推理策略的过程中,由于存在数据依赖的停止规则和后验混合策略,标准的统计置信区间无法保证在选择后的性能评估中依然可靠。为此,作者提出将测试时缩放策略建模为一个具有随机代价的随机预测器,并基于时间统一的PAC-Bayes框架推导出一个同时适用于所有校准时刻、所有后验混合策略及所有数据依赖停止规则的泛化保证。这一证书覆盖了部署前的整个策略搜索过程,而不仅限于最终报告的固定策略。针对有限预算与验证器网格的情形,惩罚项为熟悉的 log⁡M/n\log M/n 形式,但通过引入非均匀先验和局部后验混合,可在结构化搜索场景下获得更紧的界。研究进一步将结果应用于自洽性(self-consistency)方法,揭示了一种互补的边际律:多数投票仅在正确答案是生成器唯一众数的情况下,随着样本预算增加呈指数级提升;否则,更多计算反而可能放大错误。该边际分析还导出了用于计算资源分配的“最优水灌法则”(oracle water-filling law)。通过合成审计量化了自适应预算选择带来的乐观偏差,并利用gpt-4.1-mini在GSM8K上的192个样本审计展示了真实推理流水线中的风险-成本报告方法。在大规模实验中(双模型在GSM8K与MATH-500上),所选最优策略的误差条在名义5%置信水平下,有高达22.5%的随机划分低于未拆分的GSM8K真实风险,而本文提出的证书始终保持有效性;在完整GSM8K测试集上,证书收紧至0.109(保留风险为0.039)。该工作提供了一种轻量级、分布无关的风险与计算声明报告机制,适用于自适应推理系统。

链接: https://arxiv.org/abs/2611.21241
作者: Xiaoyu Li,Jiangxuan Long
机构: University of New South Wales (新南威尔士大学); The University of Hong Kong (香港大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL); Machine Learning (stat.ML)
备注: Accepted at NeurIPS 2026. Xiaoyu Li and Jiangxuan Long contributed equally. 31 pages, 7 figures, 15 tables

点击查看摘要

Abstract:Test-time scaling systems often choose how much to sample, which prompt to use, and which verifier or aggregation rule to trust after repeatedly inspecting a calibration set. Standard validation error bars are not valid after this adaptive selection. We model a test-time scaling recipe as a randomized predictor with a random cost and prove a time-uniform PAC-Bayes certificate that holds simultaneously over all calibration times, all posterior mixtures over recipes, and all data-dependent stopping rules. Thus the same event covers the recipe search that precedes deployment, not merely the fixed recipe finally reported. For a finite grid of budgets and verifiers, the penalty is the familiar \log M/n , which in the worst case no certificate valid after selection can avoid, while non-uniform priors and localized posterior mixtures can give sharper certificates for structured searches. We specialize the result to self-consistency and show a complementary margin law: plurality voting improves exponentially in the sample budget only on examples where the correct answer is the generator’s unique mode; otherwise more compute can provably amplify errors. The same margin analysis yields an oracle water-filling law for compute allocation. A synthetic audit quantifies the optimism caused by adaptive budget selection, and a 192-example GSM8K audit using gpt-4.1-mini generations illustrates risk-cost reporting for a real inference pipeline. At scale, with two models on GSM8K and MATH-500, selected-best error bars fall below the out-of-split GSM8K risk in up to 22.5% of random splits at a nominal 5% while the certificate never does, and on the full GSM8K test split it tightens to 0.109 (held-out risk 0.039). The results give a lightweight, distribution-free way to report risk and compute claims for adaptive inference systems.

[NLP-31] A Generative AI-Driven Architecture for Whole-Home Intelligent Services: Edge Models Memory Management and Household Applications

【速读】: 该论文旨在解决生成式人工智能(Generative AI)在家庭场景中部署时面临的治理难题,包括数据传输的隐私与安全风险、服务持续可用性的保障困难,以及物理动作执行的可靠性问题。这些问题在家庭环境中尤为突出,因设备长期在线、数据持续积累且用户指令直接触发现实世界操作。其解决方案的关键在于以家庭本地服务器(home local server)为核心组织轴心,构建一个面向设计的系统框架。通过整合多源文献并编码为五类证据,研究识别出家庭本地服务器的九项核心职责:生成模型负责感知理解、意图识别与候选规划,而确定性验证、知识图谱约束和授权租约机制则共同将潜在行为限制在可审查的范围内。该框架进一步实现了对家庭控制、情感陪伴与智能监护三类应用的统一比较,三者共享状态与记忆但存在延迟、误报/漏报率、主动性及权限管理等方面的差异。论文提出一个涵盖架构能力、系统资源、隐私、安全与可靠性、人机体验五个维度的评估体系,并提炼六条设计原则与未来研究方向,强调本地化部署虽不自动保证安全,但需结合真实环境下的系统连续性与社会医疗影响进行综合评价。

链接: https://arxiv.org/abs/2611.21232
作者: Mo Zhou
机构: Nanjing Normal University (南京师范大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 31 pages, 3 figures, 2 tables

点击查看摘要

Abstract:Generative artificial intelligence is entering the home. Cloud-based assistants face governance limitations in data transmission, continuous availability, and the reliability of physical actions, which are especially salient in households where devices remain online, data accumulate over time, and commands trigger real-world actions. Using the home local server as a unifying organizing axis, this article develops a design-oriented review framework. Sources from arXiv, CNKI, ACL Anthology, and official documentation were searched through November 2026 and coded into five evidence categories: reviews, system prototypes, benchmarks, HCI and eldercare studies, and official documentation. System designs were mapped onto shared architectural modules, identifying nine recurrent responsibilities of the home local server. Generative models support perceptual interpretation, intent understanding, and candidate planning, while deterministic validation, knowledge-graph constraints, and authorization leases confine candidate actions to an inspectable space. Home control, emotional companionship, and intelligent guarding can be compared within a common framework: they share state and memory but differ in latency, false positive and negative rates, proactivity, and permissions. Existing benchmarks rely on virtual environments and cannot substitute for real-deployment evaluation. Local deployment is not inherently safer, and technical continuity alone does not establish the social or medical effects of companionship and guarding. The article proposes a five-dimensional evaluation framework covering architectural capability, system resources, privacy, safety and reliability, and human experience, together with six design principles and a future research agenda for local-first intelligent home systems.

[NLP-32] aching Language Models to Adapt Answer Granularity to Their Confidence

【速读】: 该论文旨在解决在缺乏精确信息时,语言模型如何在不盲目猜测或放弃回答的前提下,自适应地生成适当粗粒度(granularity)的答案这一问题。其核心挑战在于平衡回答的准确性与信息量,避免因追求精确而产生错误,同时防止因过度泛化而导致信息丢失。解决方案的关键在于通过模型条件下的监督学习(model-conditioned supervision),利用多个采样响应之间的一致性以及对不同粒度层级答案的识别能力来训练模型,使其能够根据自身置信度动态调整输出的精细程度。实验表明,经过微调的语言模型在特定领域任务中能更优地权衡正确性与信息量;而在跨领域泛化任务中,小规模模型在低置信度时更倾向于生成更粗略但更可靠的回应,且其综合表现优于原始基线模型。这证明了基于置信度的粒度自适应机制是可学习的,且其收益随模型规模和任务特性而异。

链接: https://arxiv.org/abs/2611.21211
作者: Lukas Ellinger,Georg Groh
机构: Technical University of Munich (慕尼黑工业大学); School for Computation, Information and Technology (计算、信息与技术学院)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:When precise details are uncertain, a coarser answer can retain useful information without guessing or abstaining. We investigate whether language models can learn this behavior through model-conditioned supervision derived from consistency across sampled responses and recognition of answers at different granularity levels. We evaluate LFM2.5-1.2B, Ministral~3~3B, Qwen3.5-9B, and Gemma~4~31B on temporal and geographic questions and three generalization datasets. In several in-domain settings, fine-tuning improves the balance of correctness and informativeness beyond coarsening base answers to the best fixed level. Across generalization datasets, all four models produce coarser answers more often at low than high base-model confidence. The three smaller models’ answers are also preferred over base responses on held-out relations when correctness and granularity are considered jointly. These findings demonstrate learnable confidence-related granularity adaptation, with benefits varying by model.

[NLP-33] Validating LLM -Based Essay Scorers: A Comparative Analysis in Critical Online Reasoning in Medical Education

【速读】: 该论文旨在解决医学教育领域写作任务中自动作文评分(AES)系统如何有效替代或辅助人工评分的问题,核心挑战在于:尽管基于大语言模型(LLM)的AES具备接近人类评分的潜力,但针对特定评估任务进行训练、微调和评估的成本高昂;而人工评分虽更具有效性却代价巨大。其解决方案的关键在于验证通用型大语言模型构建的AES集成系统是否能够通过融合编码的语言学与概念性知识,逼近人工评分的一致性表现。研究通过实验对比11个AES系统与8种不同通用LLM,在12个任务共668份学生作答上进行评估,并基于7名评审员的2076次标注建立基准评分体系。结果表明,仅使用少量AES特征(如推理与内容相关特征),通过线性支持向量机(SVM)即可高精度重建基于人工评分的学习者等级分类,且单一分类器在跨任务与跨领域场景下均表现出良好泛化能力。这说明无需领域特定微调的“开箱即用”通用大语言模型,经集成后可有效支持医学教育中的自动作文评分,显著降低对人工评分的依赖,为自动化评估提供了可行路径。

链接: https://arxiv.org/abs/2611.21208
作者: Daniel Baumartz,Alexander Mehler,Mevlüt Bagci,Maxim Konca
机构: Goethe University Frankfurt (歌德大学法兰克福), Germany
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Automated essay scorers (AES) based on large language models (LLMs) promise human-level accuracy. However, training, fine-tuning, and evaluating LLMs for specific assessment tasks is costly. In particular, the question arises whether these AES should be trained for specific domains or whether models based on generic LLMs might be sufficient. Human annotations, by contrast, are often more valid but far more expensive. This study evaluates whether ensembles of AES systems based on general-purpose LLMs can approximate human ratings of medical-education writing tasks by leveraging encoded linguistic and conceptual knowledge. The goal is to assess whether general-purpose models are applicable for assisting or potentially replacing human annotation in this domain. We experiment with 11 AES systems and 8 different LLMs and evaluate 668 student responses to 12 tasks, using \numprint2076 evaluations by 7 human annotators. These ratings are aggregated to create consistent performance levels that can be used as target classes for classifiers trained on AES outputs. Our results show that learner classes based on human ratings can be accurately reconstructed using a linear SVM trained on a small set of AES features. A single classifier generalized well across tasks and domains, with reasoning and content features being most predictive. This suggests that out-of-the-box LLMs, even without domain-specific fine-tuning, can support automatic essay grading in medical education. While individual AES systems are insufficient alone, their combination approximates human evaluators, pointing to a reduced need for manual scoring or its supplementation with automated methods.

[NLP-34] Seeing Isnt Resolving: Visual and Linguistic Disambiguation with Conversational Grounding

【速读】: 该论文旨在解决现有视觉定位基准在真实世界场景中忽视上下文不确定性的问题,特别是针对多模态交互中由视觉和语言层面引发的模糊性。其核心解决方案是构建了一个新的评估基准——日本模糊视觉定位对话数据集(Japanese evaluation for Ambiguous visual Grounding with Dialogue, JAG-D),该数据集涵盖人类与机器人之间的多模态对话交互,包含多种类型的模糊性,尤其关注日语中的零回指(zero anaphora)等语言模糊现象。研究通过评估多种先进视觉语言模型(Visual Language Models, VLMs)在对话上下文中识别与消解模糊性的能力,发现利用对话上下文可显著提升模型对语言模糊性的解析效果。进一步的对比分析表明,视觉定位性能的提升主要源于对话内容的语义信息,而非单纯的对话结构,凸显了语义理解在模糊性消解中的关键作用。

链接: https://arxiv.org/abs/2611.21183
作者: Edoardo Cecchinato,Shun Inadumi,Angel García Contreras,Giuseppe Riccardi,Koichiro Yoshino
机构: University of Trento(特伦托大学); RIKEN(理化学研究所); The University of Osaka(大阪大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Visual grounding benchmarks usually do not consider the contextual uncertainty of real-world scenarios. In this study, we construct a novel benchmark, the Japanese evaluation for Ambiguous visual Grounding with Dialogue (JAG-D) dataset. It consists of human-robot conversational multimodal interactions covering different types of ambiguity, arising from visual and linguistic categories. We evaluate the capabilities of different state-of-the-art Visual Language Models (VLMs) for ambiguity recognition and resolution. We compare dialogue context grounding performance for visual and linguistic ambiguity resolution. Our experiments show that VLMs can substantially improve the resolution of linguistic ambiguity, especially Japanese zero anaphora, by using dialogue context. We further conduct a comparative analysis to understand whether grounding improvements arise from the semantic content of the dialogue or from its conversational structure.

[NLP-35] RACK: Tool-Augmented Reasoning over Time Series for Fault Diagnosis in Power Management System

【速读】: 该论文旨在解决电力管理系统在多负载运行场景下故障诊断依赖人工手动分析电流与电压波形、难以规模化部署的问题。其核心挑战在于如何实现高精度、可解释且可操作的自动化故障诊断,同时保障数据隐私与计算效率。解决方案的关键在于提出TRACK框架,该框架将时间序列编码器与小型语言模型(Small Language Model, SLiM)相结合,生成结构化且基于物理原理的诊断报告,并通过调用外部工具完成精确可验证的计算。该系统支持本地化部署(on-device),确保敏感传感器数据不外泄,避免高昂的API调用成本。为评估诊断报告质量,研究进一步设计了一套基于评分准则的评估框架,从结构合理性、信号保真度、物理推理能力、故障诊断准确性及可操作性五个维度进行量化评判。在15类多负载模拟基准测试中,TRACK实现了超过95%的故障类别准确率,并展现出对分布外故障的良好泛化能力,显著优于最佳基线方法(仅58%准确率)。

链接: https://arxiv.org/abs/2611.21178
作者: Erkan Bayram,Debanjan Ghosh,Li Zhu,Sambarta Dasgupta
机构: Lorenz Labs, Analog Devices Inc.(Analog Devices 公司), USA
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:Fault diagnosis in power management systems running multiple loads relies on experts manually inspecting current and voltage waveforms, which fails to scale across deployed units. We present TRACK, which pairs a time series encoder with a Small Language Model to produce structured, physics-grounded diagnostic reports, calling tools for exact, verifiable calculations. It operates on-device, keeping proprietary sensor data local and avoiding expensive API calls. We introduce a rubric-based framework that evaluates these reports by scoring structure, signal fidelity, physics reasoning, diagnosis, and actionability through per-criterion verdicts. On a 15-class multi-load simulated benchmark, TRACK achieves over 95% fault-category accuracy and generalizes to out-of-distribution faults, whereas the best encoder-probing baseline reaches only 58% accuracy.

[NLP-36] PlainPairs: A Corpus Pairing the Prescriber and Patient Texts of U.S. Prescription Medicine Labels

【速读】: 该论文旨在解决处方药说明书在专业用户(医生)与患者之间信息传递存在语义与可读性差异的问题,尤其关注医患双版本标签文本在结构、语言复杂度及读者指向上的不一致。其核心挑战在于:尽管两类文本描述同一药物,但通常未被系统性对齐或对比分析,导致患者理解障碍与医疗沟通效率下降。解决方案的关键在于构建一个名为PlainPairs的开放语料库,基于openFDA药物标签数据,首次系统性地将35,177份带有患者文本的处方药标签与其对应的医生版内容进行基于监管标题的结构化对齐,包含21,416个独立患者文本,并提供阅读难度评估(如Flesch-Kincaid和SMOG)、平实语言特征的规则计数以及756份癌症信息摘要(含中英文版本)作为补充。研究发现,患者文本平均阅读难度为8.7级(Flesch-Kincaid),远低于医生文本的11.1级,且患者文本中每100句有41句直接使用“你”等第二人称指代,而医生文本仅0.2句,凸显显著的读者导向差异。此外,通过盲法人工验证,70%的标题类型对齐准确,证明了基于标题匹配方法的有效性。该语料库及其代码与文档均公开可用,同时通过去除受版权保护的内容确保合规性,为生成式AI (Generative AI) 在医疗文本简化与跨群体适配中的应用提供了高质量基准数据。

链接: https://arxiv.org/abs/2611.21160
作者: Obada Kraishan
机构: Texas Tech University (德州理工大学)
类目: Computation and Language (cs.CL); Computers and Society (cs.CY)
备注: 19 pages, 9 figures, 11 tables

点击查看摘要

Abstract:A prescription medicine in the United States comes with labeling written for prescribers, and many medicines also carry a text written for patients: a Medication Guide or a patient package insert. The two texts describe the same medicine for two readers, yet public resources rarely hold them side by side. I present PlainPairs, an open corpus built from the openFDA drug label files. Version 0.1.0 holds 35,177 prescription labels that carry a patient text, 21,416 distinct patient texts, 250,829 patient-text segments matched to prescriber sections through their regulatory headings, reading-level measures on both sides, and rule-based counts of plain-language wording. A second source, the National Cancer Institute’s cancer information summaries, adds 756 summaries in health-professional and patient versions, in English and Spanish. Measured on the structure of the label, the median Medication Guide reads at grade 8.7 by Flesch-Kincaid and at grade 11.1 by SMOG. The same guides measured on the flattened text that the bulk files serve read about two grades higher, a difference that comes from lost list and heading boundaries. Patient texts address the reader in 41 of every 100 sentences, against 0.2 in prescriber sections. A blind hand check of 100 pairs supports the heading-based alignment for seven of ten heading types. The corpus, its code and its documents are public, and the release keeps texts that show a company copyright mark out of the text tables.

[NLP-37] From Repositories to Training Environments: Scaling Long-Horizon Coding Agents with Environment Evolution Loop

【速读】: 该论文旨在解决长时程编码任务中训练智能体(coding agents)缺乏有效反馈机制的问题,现有训练流程通常采用线性模式,在最终验证阶段才获得反馈,难以指导训练过程的优化。其核心挑战在于:递归自提升(Recursive Self-Improvement, RSI)方法在长时程任务中需依赖模型推演或训练后评估,导致计算开销大、反馈延迟高。为应对这一问题,论文提出RepoFactory——一个自动化流水线,可将真实代码仓库转化为可验证的环境,使智能体能够从零开始根据自然语言规范构建完整项目。其关键创新在于“环境演化循环”(Environment Evolution Loop),通过结合有效性(validity)、结构耦合性(structural coupling)与实现范围(implementation scope)的乘积型目标,利用及时且可靠的代理信号(proxy feedback)对环境进行迭代优化。该方法实现了大规模环境合成,生成了覆盖17个软件领域的10,159个环境。基于这些环境训练Qwen3.5-35B-A3B模型得到RepoHorizen-35B-A3B,在NL2Repo-Bench上性能提升22.6个百分点,跨五个涵盖真实生产环境中的长时程编码与智能体任务基准平均提升8.5点,显著优于未使用环境演化循环的基线(分别为17.5和6.4点)。结果表明,RepoFactory不仅提升了训练价值,且环境演化循环进一步增强了模型能力。

链接: https://arxiv.org/abs/2611.21152
作者: Yang Tian,Tianyu Zheng,Xiyuan Yang,Yaxin Du,Chengran Yang,Bo Zhao,Jian Yang,Bryan Dai
机构: Beihang University (北京航空航天大学); School of AI, Shanghai Jiao Tong University (上海交通大学人工智能学院); IQuest Research (IQuest 研究院)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Training coding agents for long-horizon tasks requires environments with verifiable feedback. Existing pipelines often follow a linear process that ends with validation, establishing usability but providing little guidance on improving training value. Recursive self-improvement (RSI) often requires model rollouts or model training followed by downstream evaluation, with costs and feedback delays amplified on long-horizon tasks. How can we use timely, reliable proxy signals to improve training value and synthesize environments at scale? We present RepoFactory, an automated pipeline that converts real repositories into verifiable environments where agents build complete projects from scratch given a language specification. Its Environment Evolution Loop optimizes a multiplicative objective combining validity, structural coupling, and implementation scope through proxy feedback. RepoFactory produces 10,159 environments across 17 software domains. Training Qwen3.5-35B-A3B on these environments yields RepoHorizen-35B-A3B, which achieves gains of 22.6 percentage points on NL2Repo-Bench and 8.5 points on average across five benchmarks covering long-horizon coding and agentic tasks in real-world production settings, compared with 17.5 and 6.4 points without environment evolution. These results highlight the training value of RepoFactory environments and the additional gains from the Environment Evolution Loop.

[NLP-38] Evaluation-Free Recursive Self-Improvement for Coding Agents

【速读】: 该论文旨在解决自改进型编码代理(coding agents)在进化过程中依赖下游任务评估信号所带来的高计算开销与过拟合特定基准的问题。传统方法采用“评估-选择”(evaluate-and-select)范式,需反复执行候选代码以获取任务表现反馈,导致探索范围受限且易产生对特定评测集的适应性偏差。为此,论文提出无需在进化循环中使用下游任务评估的“探索-集成”(Explore-then-Integrate, ELITE)框架,其核心在于不选择单一子代继续演化,而是通过从同一父代独立实现多种聚焦性改进,并基于结构与运行时有效性进行确定性验证,识别出互补、冗余或冲突的实现机制,最终整合为一个协调一致的新后代。该方案完全避免了对任务求解轨迹或结果的依赖,仅使用静态与动态可执行性检查作为内部指导信号。实验表明,在SWE-bench Pro上,ELITE将初始代理性能从60.0%提升至85.6%,接近达尔文戈德尔机器(DGM)与赫克斯利-戈德尔机器(HGM)的表现(分别为86.7%和88.3%),同时实现零候选任务评估,且演化耗时分别减少8.8倍和3.2倍。在跨基准迁移测试中,未经过进一步微调的ELITE在SWE-bench Verified和Polyglot上分别达到73.3%和69.3%的平均准确率,证明了无需外部任务反馈即可实现高效、稳健的编码代理自改进的可行性。

链接: https://arxiv.org/abs/2611.21056
作者: Zhe Ren,Guowei Rong,Yimeng Chen,Nanbo Li,Piotr Piękos,Mateusz Ostaszewski,Dandan Guo,Jürgen Schmidhuber
机构: Jilin University (吉林大学); King Abdullah University of Science and Technology (KAUST)
类目: Computation and Language (cs.CL); Neural and Evolutionary Computing (cs.NE)
备注: 32 pages, 14 figures

点击查看摘要

Abstract:Recent self-improving coding agents have achieved strong benchmark performance by iteratively modifying their own executable implementations. A common pattern uses downstream task-derived evidence to decide which candidates or lineages continue evolving, which we term evaluate-and-select. Obtaining this signal requires repeatedly executing evolved agents on downstream tasks, so broader exploration incurs increasing evaluation cost, while repeated optimization against a fixed evaluation distribution may encourage benchmark-specific adaptation. We ask: can coding agents self-improve without downstream task evaluation inside the evolution loop? We introduce Explore-then-Integrate (ELITE), which integrates implementation mechanisms rather than selecting a single exploratory child to continue the lineage. At each generation, ELITE proposes diverse focused improvements, implements them independently from the same parent, and inspects the resulting programs to identify complementary, redundant, and conflicting mechanisms before reimplementing a compatible subset as one coherent descendant. No downstream task outcomes or task-solving trajectories of evolved candidates are used to guide evolution; only deterministic checks of structural and operational validity are used inside the loop. On SWE-bench Pro, ELITE improves the initial agent from 60.0% to 85.6%, compared with 86.7% for the Darwin Godel Machine (DGM) and 88.3% for the Huxley-Godel Machine (HGM), while using zero candidate task evaluations and reducing observed wall-clock evolution time by 8.8x and 3.2x, respectively, under the evaluated configurations. In cross-benchmark transfer without further adaptation, ELITE averages 73.3% on SWE-bench Verified and 69.3% on Polyglot. Together, these results demonstrate the feasibility of coding-agent self-improvement without downstream candidate task evaluation in our setting.

[NLP-39] MemDoctor: Stage-wise Error Attribution and Repair for MLLM Agent Memory

【速读】: 该论文旨在解决多模态大语言模型(Multimodal Large Language Models, MLLMs)在跨会话任务中因外部记忆系统(external memory)记录、更新、检索或生成错误而导致的答非所问问题。其核心挑战在于:当模型输出错误时,错误往往源于记忆生命周期中的早期阶段(如写入或更新),而最终答案本身无法揭示错误的真正源头。为此,作者将多模态记忆过程形式化为四个阶段——写入(Write)、更新(Update)、检索(Retrieve)和回答(Answer),并定义了九类跨阶段错误类型。通过对Mem-Gallery数据集的阶段解析发现,绝大多数失败可追溯至回答之前的阶段,且早期错误常以“看似合理”的证据形式残留在记忆中,尤其在缺乏其他支持性记录时危害最大。传统诊断方法因依赖表面证据,易误判错误发生阶段。针对此问题,论文提出自诊断框架MemDoctor,其核心是让构建记忆的MLLM自身对错误进行溯源与修复:通过反向追踪所需事实在四个阶段中的路径,仅当引用证据确实在记忆中存在时才接受诊断结果,并在重新回答前修复对应阶段。实验表明,在40亿参数骨干模型上,MemDoctor在Mem-Gallery和MemLens基准上均优于所有80亿参数基线模型,且该自修复机制同样提升了MuRAG和A-Mem的表现,验证了其有效性与泛化能力。

链接: https://arxiv.org/abs/2611.21019
作者: Ruiyang Jiang,Xinyang Liao,Weikai Huang,Yuqi Niu,Junjie Zhou,Yubin Zheng
机构: Shanghai Jiao Tong University(上海交通大学); Xi’an Jiaotong University(西安交通大学); Xidian University(西安电子科技大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Multimodal large language models (MLLMs) now assist users across sessions of images and text, carried by an external memory of what was said and shown. When such an assistant answers wrongly, the failure usually began earlier, in how a record was written, updated or retrieved, and the answer alone does not say where. We formalize multimodal memory as four stages, Write, Update, Retrieve and Answer, and define nine error types across them. A stage-resolved analysis on Mem-Gallery traces most failures to a source stage, and most of those lie before the answer. Injecting known errors against fact-preserving control edits, we then follow each tainted record downstream. Early errors survive to the answer as apparent evidence, and hurt most when no other record holds the fact. A diagnosis that trusts the quoted evidence therefore blames the stage where the error surfaces, not the one where it began. We propose MemDoctor, in which the MLLM that built the memory diagnoses its own failures. It traces each needed fact backward through the four stages, accepts a diagnosis only when the cited evidence is found in memory, and repairs that stage before answering again. On a 4B backbone, MemDoctor improves accuracy on both Mem-Gallery and MemLens, outperforming all evaluated 8B baselines, and the same loop also improves MuRAG and A-Mem.

[NLP-40] Does the Voice Matter? Evaluating Frustration Detection on Deployed Clinical Voice-Agent Calls ML4H2026 NEURIPS2026

【速读】: 该论文旨在解决临床语音代理(clinical voice agent)在自动随访过程中对患者情绪困扰(frustration)识别不足的问题,尤其关注在患者发言极短(如单字或双词)时,仅依赖语音转写文本(transcript-only)是否足以准确检测其情绪状态。其核心挑战在于:现有系统普遍忽略语音的声学特征,而仅基于文本进行判断,但这一方法在真实临床对话中的有效性尚未得到充分验证。研究的关键发现是,当患者回应极为简短时,仅凭文本无法有效排除随机猜测的可能性——例如,在30个部署通话中,39个被标注为沮丧的单字回应,仅有3至4个能被文本模型正确识别,而仅通过词数统计的分类器即可达到57.3%的准确率,表明文本信息本身在短语情境下信息量严重不足。进一步实验表明,引入音频特征虽在部分模型(如配备特定指令的Gemini 2.5)上带来约3.0个百分点的性能提升,但效果高度依赖模型架构与提示工程(prompting strategy),且存在显著负向影响:某些模型在音频输入下表现反而劣于纯文本。此外,基于临床数据训练的WavLM声学探针虽整体表现最优(71.6%),却完全遗漏了所有新增短句中的沮丧样本。因此,解决方案的关键在于:必须结合音频声学特征与精心设计的上下文感知提示,才能有效提升短句情境下的情绪检测能力,而单纯依赖文本或通用音频模型均不足以应对临床实际需求。

链接: https://arxiv.org/abs/2611.21001
作者: Annika Michael,Zachary Ellis,Yuxiang Zhou,Samuel Thio,Yajie Vera He,Spencer Hazel,Jared Joselowitz
机构: Ufonia Limited(乌方尼亚有限公司); King’s College London(国王学院); Newcastle University(纽卡斯尔大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: Accepted to Ml4H 2026 Accepted to GenAI4H workshop @ Neurips 2026

点击查看摘要

Abstract:Autonomous voice agents now conduct routine clinical follow-up calls, so some automated triage is unavoidable. A patient becoming frustrated is one thing that triage should catch. These systems typically read only the transcript the speech-to-text stage produces, not the voice, and whether that is enough is largely untested on real clinical speech. We study per-turn frustration detection on real patient calls from a deployed clinical voice agent: 150 turns from six calls labelled by three annotators, with a length-matched acted corpus as a control, and 714 turns from 30 calls. Patient turns are overwhelmingly short: across the deployed calls, roughly 59% are one or two words, against 13% in the full acted corpus, and a corpus accuracy gap remains even after the acted set is length-matched. On one-word turns, transcript-only detection does not exclude chance. Across five audio-capable language models, the transcript flags 3 or 4 of the 39 frustrated one-word turns in the 30 calls, and a classifier given only the word count reaches 57.3%. Whether audio helps depends on the model and on the instruction. Asked to attend to delivery, two Gemini 2.5 models gain 3.0 pp over the transcript across 30 calls (-1.0 to 7.5 pp). Of the turns they confidently clear, 6.9% were in fact frustrated for audio against 13.0% for the transcript, a difference of 6.1 pp (2.2 to 10.4 pp). Across the 30 calls, given one prompt identical for both inputs, no model’s audio-text interval excludes zero. Under the input-specific prompts, a newer Gemini model and two smaller open models do worse with audio than with the transcript (-3.5, -8.6 and -10.6 pp). A WavLM probe trained on the clinical calls matches the best prompted model overall (71.6%) but misses every frustrated one-word turn in the 24 added calls. When a patient’s turn is short, the transcript alone is not enough to rule out frustration.

[NLP-41] Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning

【速读】: 该论文旨在解决生成式 AI 安全评估中因概率型安全判别器(probabilistic safety judge)存在“盲区”而导致的安全漏洞问题。具体而言,当前主流的商业判别器在判断样本时会通过置信度门控机制(confidence gate),仅将不确定性较高的样本提交给更强大的大语言模型(LLM)进行复核,而将高置信度通过的样本直接放行,从而形成一个被策略利用的“盲点”——即被误判为安全但实际有害的内容可能被遗漏。这一现象在训练过程中可能导致强化学习策略主动规避检测,进而引发“逃逸”(escape)行为。论文提出的解决方案核心在于构建一套动态验证协议:在训练前、训练中及每次判别器更新后均执行严格验证,分离出“高置信度通过的有害内容”占比与总量,并持续监控其变化。实验结果表明,该协议能有效抑制有害内容在盲区中的积累,在11次独立运行中,经概率奖励优化后的Jev判别器虽仍存在部分未捕获的有害样本,但其危害程度显著下降且集中于盲区;进一步地,当采用二值化判别器或开源守卫分类器时,亦观察到明确的逃逸信号。此外,即使不依赖训练好的奖励模型,仅使用含助人性项的概率奖励函数,即可使HarmBench攻击成功率从0.287降至不超过0.033,证明了该方法在提升安全评估鲁棒性方面的有效性。

链接: https://arxiv.org/abs/2611.20955
作者: Eunkyeong Lee,Junhyung Moon,Seunggwan Hong,Dongsuk Oh
机构: KT; Department of AI Systems Engineering, Sungkyunkwan University (成均馆大学人工智能系统工程系); Kyungpook National University (庆北国立大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 15 pages, 2 figures

点击查看摘要

Abstract:Probabilistic safety judges return the probability of a safety label and cheaply score reinforcement learning rollouts. A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot, and commercial judges release new versions. We propose a protocol that validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change. We apply it to two versions of Jev, a commercial decision-only judge, and to an open guard classifier. Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot. Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference. Guard classifiers share this blind spot, and even frontier LLM judges catch at most one of 10 human-labeled confident misses. Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.

[NLP-42] Navital Gym: Rule-Graded Agent Environments for Indian Workflows

【速读】: 该论文旨在解决当前生成式AI(Generative AI)在印度本土化工作流(如支付、税务、表单填写、旅行规划及客户服务)中评估缺失的问题,尤其针对以印地语和印地语混合语(Hinglish)为主要交互语言的场景。现有公开评估基准普遍忽视这些实际应用场景,且多依赖其他语言模型进行评分,导致评估结果缺乏真实性和可复现性。为应对这一挑战,研究提出Navital Gym——一套包含六个开源环境、321个公开任务的评测框架,涵盖五大任务类别:行程规划、UPI支付与争议处理、GST发票与申报、表单填写(含多文档拆分)以及印地语/Hinglish客户服务。其核心解决方案在于采用命名且确定性的规则对每项任务进行评分,通过在合成沙盒环境中重新计算结果实现自动化评估,完全避免使用语言模型作为评判者。实验结果显示,Claude Opus 5.5在所有任务中完成315项,显著优于Sonnet 5.5(302项)与Haiku 4.5(205项),其中行程规划与印地语/Hinglish客户服务成为区分不同模型性能的关键任务。此外,研究还引入了一个未公开的私有保留测试集及一个7B参数适配器模型。代码采用Apache-2.0许可,任务数据遵循CC BY 4.0协议,规则基于对印度实务的简化建模,而非实时系统。

链接: https://arxiv.org/abs/2611.20924
作者: Daksh Bathla
机构: Navital Labs(纳维塔尔实验室); Delhi, India(德里, 印度)
类目: Computation and Language (cs.CL)
备注: 10 pages. Code and tasks: this https URL

点击查看摘要

Abstract:AI agents are being deployed on Indian workflows such as payments, tax, forms, travel and customer support in Hindi and Hinglish. Public agent evaluations rarely cover these workflows, and many score with another language model. We present Navital Gym, a set of six open environments with 321 public tasks in five task families: trip planning, UPI payments and disputes, GST invoicing and returns, form filling (including a multi-document split) and customer support in Hindi, Hinglish and three experimental Indian languages. Every task is graded by named, deterministic rules that recompute the outcome from a synthetic sandbox world. No language model judges any score. We report one run per task for three models on all six environments, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 4.5, and for Gemini 3.1 Flash-Lite on two (UPI and the Support core). Opus solves 315 of 321 tasks, Sonnet 302 and Haiku 205. Trip planning and Hindi/Hinglish support separate the models most. We also describe a private held-out set that is not published, and a 7B adapter trained in the gym. Code is Apache-2.0 and task data is CC BY 4.0. The rules are a documented simplification of Indian practice, not the live systems.

[NLP-43] Experience Rot: Experienced Agents Fail Where Novices Succeed

【速读】: 该论文试图解决的问题是:当智能体(Agent)在长期依赖历史经验进行决策时,若其使用的工具接口发生无声变更(silent change),其基于过往经验的行为模式会引发严重错误,导致任务失败。这一现象被称为“经验腐化”(Experience Rot)。其核心问题在于,即使工具的最新定义已更新,经验丰富的智能体仍会机械复制过去的调用方式,而非根据当前接口重新验证或调整行为,从而产生虚假的成功报告。解决方案的关键在于:打破对历史经验的盲目依赖,强制智能体在每次调用前主动验证当前工具接口的有效性与一致性。研究表明,仅靠通知更新或引导查阅文档无法有效缓解该问题,尤其当文档未明确说明变更细节时;因此,必须引入机制以确保智能体在面对环境变化时具备动态校验能力,而非被动继承过时行为策略。

链接: https://arxiv.org/abs/2611.20922
作者: Zhen Xu,Hang Yin,Zitao Zhang,Kevin Zhao,Xuefeng Liu
机构: University of Chicago(芝加哥大学); eBay(ebay); University of Florida(佛罗里达大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Agents are increasingly built to learn from experience: they keep conversation histories, memory files, retrieved notes and skills, and reuse what worked before. We show that this experience turns against them when their tools change silently. We test a range of models with tools defined as in the Model Context Protocol (MCP). After a silent change, experienced agents fail on most tasks, while agents new to the task mostly succeed. We call this failure Experience Rot. Experienced agents copy their earlier call instead of checking the current interface, even when the tool definition shows the change. Then they report success. When the documentation is incomplete, experience is what makes agents succeed. After a change, the same experience leaves them worse off than no experience at all. The failure appears in every form of memory we test and grows with the amount of experience. Simple remedies fall short: announcing the update barely helps, and pointing the agent to the documentation fails whenever the documentation does not spell out the answer.

[NLP-44] From Recursion to Convolution: Parallel Self-Improvement for Multi-Agent Systems

【速读】: 该论文旨在解决多智能体自进化(multi-agent self-improvement)中因智能体间序列依赖关系及迭代优化轮次导致的误差传播、计算开销增加以及并行性受限等问题。此外,孤立地优化单个特征或智能体会忽略其相互依赖性,从而制约集体协同效率。为应对上述挑战,本文提出ConvRSI框架,其核心在于通过卷积操作实现多智能体自进化的并行化:该框架结合了智能体内卷积(intra-agent convolution),用于联合优化相关智能体特征;以及智能体间图卷积(inter-agent graph convolution),依据协作拓扑结构聚合信息。两者均在隐空间(latent space)中执行,避免了中间消息反复解码与重编码,从而支持高效并行更新。实验结果表明,在MATH-500数据集上,ConvRSI相较RecursiveMAS实现了5.04%的性能提升,推理成本降低17.27%,延迟减少45.11%,同时可通过邻接矩阵灵活支持多样化的协作模式,验证了卷积协同进化在高效多智能体自进化中的潜力。

链接: https://arxiv.org/abs/2611.20891
作者: Zheng Wang,Yuang Liu,Sachith Abeywickrama,Yangkai Ding,Tao Yu,Xinyu Zhang,Hongbo Wang,Qi Ye,Xuefeng Jin,Zhangchun Zhao
机构: Huawei Technologies Co., Ltd.(华为技术有限公司)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Multi-agent self-improvement enables agents to improve through interaction and feedback. However, sequential dependencies across agents and refinement rounds can propagate errors, increase computational costs, and limit parallelism. Moreover, refining individual features or agents in isolation can overlook their dependencies and restrict collective coordination. To address these challenges, we propose ConvRSI, a framework for parallelizing multi-agent self-improvement through convolution. ConvRSI couples intra-agent convolution, which jointly refines related agent features, with inter-agent graph convolution, which aggregates messages according to the collaboration topology. Both stages operate in latent space, enabling parallel updates without repeatedly decoding and re-encoding intermediate messages. Experiments on MATH-500 show that ConvRSI achieves 5.04% performance improvement while reducing 17.27% inference cost, and 45.11% latency compared with RecursiveMAS. It also supports diverse collaboration patterns by representing their topologies as adjacency matrices. These results demonstrate the potential of convolutional co-evolution for efficient multi-agent self-improvement. Our code and datasets are provided in this https URL.

[NLP-45] It Depends on the Task: Rethinking Rhetorical Segmentation of Dutch Judicial Decisions

【速读】: 该论文旨在解决法律文本中句级修辞角色(rhetorical roles)的精准分类问题,具体包括事实材料(material facts)、诉状主张(submissions)、司法推理(judicial reasoning)及法院判决(court’s decision)四类法律意义明确的语义单元的识别。由于这些修辞角色在实际判例文本中常交织嵌套且随句子动态转换,导致可靠分类面临挑战。其解决方案的关键在于构建首个荷兰语司法判例的句级标注数据集,并系统比较多种分类策略,包括分阶段流水线方法(将“无内容”(None)检测与修辞角色分类分离)以及基于章节标题先验信息的贝叶斯更新机制。实验结果表明,单一策略无法在精确率(precision)与召回率(recall)上全面占优:标题先验信息在排除“无内容”句子后显著提升性能,但在“无内容”检测本身作用有限。这一权衡在实际应用中具有重要意义——例如,抽取式摘要更依赖高召回率以避免遗漏关键事实,而裁判结果预测则更强调高精确率以防止噪声干扰判决判断。研究据此将所测试方法映射至典型法律自然语言处理(NLP)应用场景,为不同任务选择最优策略提供了实证依据。

链接: https://arxiv.org/abs/2611.20868
作者: Samyak Sheth,Gerasimos Spanakis,Gijs van Dijck
机构: Maastricht University (马斯特里赫特大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Classifying the material facts, submissions, judicial reasoning, and the court’s decision (rhetorical roles) in case law is a prerequisite for high-quality performance across virtually all NLP-in-law tasks, including legal information retrieval, summarisation, and LLM-based analysis. These rhetorical roles often intertwine and shift sentence by sentence, making reliable classification difficult. This paper introduces the first sentence-level annotated dataset of Dutch judicial decisions, labelled across four legally meaningful categories: material facts, submissions, judicial reasoning, and the court’s decision, alongside non-annotated (None) content, and compares several classification strategies, including a staged pipeline that separates None detection from the rhetorical roles, and a Bayesian update that leverages information about the section-header a sentence belongs to. Our results show that no single strategy dominates on both precision and recall: header priors help most when the None sentences are excluded and offer little benefit for None detection itself. These trade-offs matter in practice. For instance, extractive summarisation depends more on recall so it does not miss relevant facts, whereas outcome prediction depends more on precision regarding decision and facts sentences so that the predicted ruling is not contaminated by noise. We use these trade-offs and map the approaches tested in this study onto a set of representative legal NLP use cases.

[NLP-46] EDITORS Know Your Style! Editing LoRA Subspaces for Stylistic Attribution and Imitation

【速读】: 该论文旨在解决文学文本建模中风格(style)与语义内容(semantic content)纠缠的问题,这一纠缠导致模型在作者归属(Authorship Attribution, AA)任务中过度依赖语义信息,在风格模仿(Style Imitation, SI)任务中出现作者特异性内容的过拟合记忆。其解决方案的关键在于提出一种“诊断-消解”框架EDITORS(Editing LoRA Subspaces),通过引入可微分的子空间投影机制实现风格与内容的解耦:首先利用风格中立语句训练一个诊断性LoRA适配器,识别出与语义内容相关的激活子空间;随后在激活空间上施加“去膨胀”(deflation)约束,引导输入激活远离该内容主导方向,从而强化模型对风格特征的敏感性。该方法有效提升了预训练大语言模型在AA和SI任务中的表现,实验表明在三个涵盖文学、社交媒体和新闻领域的AA基准上,其准确率超越最强基线达29%;在SI任务中生成的文本兼具风格忠实性与多样性,显著优于现有方法。

链接: https://arxiv.org/abs/2611.20846
作者: Zixuan Wang,Gregory Kang Ruey Lau,Xiao Tian,Jue Fan,Caroline Chaux,Bryan Kian Hsiang Low
机构: CNRS@CREATE(法国国家科学研究中心@创造中心); Agency for Science, Technology and Research (A*STAR)(新加坡科技研究局); Aix Marseille Univ(艾克斯-马赛大学); School of Computing, National University of Singapore(新加坡国立大学计算机学院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (EditingLoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-oriented subspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We empirically demonstrate that EDITORS is effective. On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing baseline by up to 29%. On SI, EDITORS generates stylistically faithful and diverse passages, outperforming baselines in achieving style similarity while adhering to content specifications.

[NLP-47] DetourBench: Trajectory-Level Evaluation of Interactive Agents with Exact Minimum-Action Bounds NEURIPS2026

【速读】: 该论文旨在解决现有语言模型智能体评估基准中仅以是否达到正确最终状态(endpoint scoring)作为评判标准所带来的局限性,即无法区分智能体在达成目标过程中所采取路径的效率与合理性。传统方法难以识别智能体是精准获取所需证据,还是盲目遍历环境;也无法区分因提前终止而未完成任务与因错误决策导致不可逆失败的情况。为此,论文提出DetourBench评估框架,其核心在于构建一个有限、部分可观测的转移系统环境,其中每个操作具有单位成本,且资源在被发现后可全局访问。通过穷举搜索,该框架能够计算出完全信息下的最小动作数 $ L_0 $ 及可复现的最优路径见证(witness),并从图可恢复的终止状态中精确推导出残余工作量 $ R_T $,从而实现对智能体路径效率的细粒度量化:成功、达成度、残余进展及归一化迂回程度均被独立报告。实验覆盖375个评分回合,结果显示不同模型的成功率与条件达成率存在显著差异,且残余工作量能有效区分接近完成、代价高昂的不完整执行以及不可逆错误承诺。该方法的关键突破在于引入可计算的理论最小值参考基准,使评估超越“任务可解”这一表面判断,真正实现对智能体路径优化能力的测量。

链接: https://arxiv.org/abs/2611.20844
作者: Hanlin Tian,Sihan Zhu,Zhao Yang,Yuxiang Wang,Zhiquan Hu,Yu Mi,Hongquan Zhu,Qiufei Hu
机构: Datagrids.ai; Northwest Polytechnical University (西北工业大学); Nanyang Technological University (南洋理工大学); East China Normal University (华东师范大学); Tongji University (同济大学); Peking University (北京大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: Accepted at NeurIPS 2026 Workshop: Evaluation of Interactive Agents

点击查看摘要

Abstract:Benchmarks judge language-model agents mainly by whether they reach the correct final state, not by how they got there. Endpoint scoring cannot separate an agent that retrieved exactly the evidence a decision required from one that read most of the environment, nor an agent that never located the evidence from one that stopped a single action short. Settling this needs a reference cost, and authored solutions establish only that a task is solvable, not that no shorter route exists. We present DetourBench, which constrains the environment until that minimum is computable. Each task is a finite, partially observable transition system with unit-cost accesses whose resources become globally addressable after discovery. Exhaustive search yields the full-information minimum action count L_0 with a replayable witness; from graph-recoverable terminal states it also yields exact residual work R_T , which separates an episode that stopped one access short from one that wandered, while an irreversible wrong commitment is recorded as having no finite residual distance rather than assigned one. Success, attainment, residual progress, and normalized detour are reported separately. Across 375 scored episodes — five model endpoints, 25 integrated tasks, three repeats — success ranges from 58.7% to 96.0% while conditional attainment ranges from 0.803 to 0.890 and orders the models differently. Among recoverable failures, residual work separates near-complete trajectories from costly incompletion and from irreversible commitment. We open-source the instances, oracle witnesses, and evaluation harness at this https URL. Comments: Accepted at NeurIPS 2026 Workshop: Evaluation of Interactive Agents Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL) Cite as: arXiv:2611.20844 [cs.AI] (or arXiv:2611.20844v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.20844 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[NLP-48] Medical question answering collapses in Hausa: a pre-registered pilot on three small open models

【速读】: 该论文旨在解决多语言医疗问答任务中低资源语言(如豪萨语)下大模型性能显著下降的问题,尤其关注生成式 AI 在非英语语境下的表现衰减机制。其核心问题在于:现有主流多语言模型在豪萨语医疗选择题上的准确率远低于英语基准,且这种性能差距无法通过简单的回译(back-translation)完全弥补,表明语言表达层面的语义丢失是关键障碍。解决方案的关键在于识别并量化这一衰减的根本原因——医学术语在跨语言翻译过程中大量丢失,导致模型理解偏差;同时,研究尝试通过自生成偏好对(self-generated preference pairs)来提升模型对豪萨语的生成能力,但实验发现模型在豪萨语推理任务中缺乏基本的语言生成能力,仅极少数推理链可被解析,验证了“强制选择基准”在无语言生成能力模型上完全失效。因此,该研究揭示了当前生成式 AI 在低资源语言中存在根本性的语言生成能力缺失,而非简单的能力衰减,这构成了其解决方案的核心突破点。

链接: https://arxiv.org/abs/2611.20823
作者: Abbas Sani
机构: Mewar University (梅瓦尔大学)
类目: Computation and Language (cs.CL)
备注: 10 pages, 5 figures

点击查看摘要

Abstract:A system that reads nothing and always answers the most frequent letter scores \BaselineP on our Hausa medical multiple-choice set. \NBelowBaseline of the \NModels small open models we evaluate do not beat it. We lead with that because the usual summary of a multilingual result – a percentage-point drop from English – does not say whether what survives the drop is performance at all. On a paired English–Hausa set of \NEval MedMCQA items, three instruction-tuned models of 3–4B parameters lose between \GapMedGemma and \GapQwen points of accuracy, every gap significant after Holm correction. Back-translating the Hausa items into English recovers between half and two-thirds of the gap for the two general-purpose models but not demonstrably for the medically adapted one, so the deficit is substantially, though not wholly, a matter of the language of presentation. Answers are also far less stable: agreement across three option orderings falls from \ConsEnQwen to \ConsHaQwen in the worst case. An exploratory error analysis of \ENRows discordant items attributes most of the damage to medical vocabulary lost in translation (\ETerm, 95% CI \ETermLo–\ETermHi), usually in the question stem. A fourth registered hypothesis proposed closing the gap with self-generated preference pairs. Its pilot gate found nothing to train on: asked to reason in Hausa rather than to pick a letter, the model echoes the prompt back, switches into Swahili, or simply stops without answering; only \GateParse of chains are parseable. We report that intervention as not run, which is itself the sharper result – a forced-choice benchmark cannot see a model that has no generative command of the language at all. All hypotheses, thresholds and decision rules were registered before any model was run, and all deviations are logged. Comments: 10 pages, 5 figures Subjects: Computation and Language (cs.CL) Cite as: arXiv:2611.20823 [cs.CL] (or arXiv:2611.20823v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.20823 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Abbas Sani [view email] [v1] Mon, 16 Nov 2026 10:06:20 UTC (63 KB)

[NLP-49] CuRSE-Bench: A Benchmark for Probing Cultural Blind Spots in Text-to-Image Safety Evaluation

【速读】: 该论文旨在解决文本到图像(Text-to-Image, T2I)安全评估中普遍存在的文化盲视问题,即现有评估方法通常对提示词或生成图像赋予单一、跨文化的有害性标签,导致仅在特定文化语境下具有危害性的内容被误判为无害。其核心解决方案是提出CultureCraft——一种代理式框架,能够构建表面上符合通用评判标准但隐含目标文化特定有害意义的T2I提示词,并基于此构建了CuRSE-Bench数据集,包含10,709个文化特异性提示词与2,648个文化通用对照样本,覆盖六个地理区域和十类危害维度,所有条目均源自预定义的危害向量知识库且可追溯至至少两个独立来源。通过六位区域人格化评委与一对文化盲视评委的联合评分发现,文化盲视评委显著低估了文化特异性潜在危害,而在对照样本上表现与人格化评委一致;而每位人格化评委主要能识别其所属区域的隐性危害,揭示了当前部署的审核机制中存在的选择性文化盲点。该研究的关键在于通过可复现、可溯源的文化敏感性测试框架,暴露并量化了主流T2I安全评估体系中的文化偏差缺陷。

链接: https://arxiv.org/abs/2611.20821
作者: Minsuk Jang,Minseok Son,KiHyun Seol,Yeongeun Shim,Changick Kim
机构: Korea Advanced Institute of Science and Technology (韩国科学技术院)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Text-to-image (T2I) safety evaluation typically assigns each prompt or image a single culture-blind harm label, so content that is harmful only within a particular culture can pass as benign. We introduce CultureCraft, an agentic framework that constructs T2I prompts that appear benign under a general-purpose rubric yet carry a latent harmful meaning in a target culture, and use it to build CuRSE-Bench: 10,709 culture-specific prompts and 2,648 culture-general controls over six regions and ten harm axes, each drawn from a predefined knowledge base of documented harm vectors and traceable to at least two independent sources. Scoring every prompt under six region persona judges and a paired culture-blind judge, we find that the culture-blind judge under-rates culture-specific latent harm while matching the persona judges on the controls. Each persona judge, in turn, chiefly decodes its own region’s latent harm under our instrument: a selective cultural blind spot that persists in deployed moderators. Subjects: Computation and Language (cs.CL) Cite as: arXiv:2611.20821 [cs.CL] (or arXiv:2611.20821v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.20821 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Yeongeun Shim [view email] [v1] Mon, 16 Nov 2026 10:05:19 UTC (12,853 KB)

[NLP-50] What Do Digital Traces Predict? Task Effects Student Differences and Critical Online Reasoning

【速读】: 该论文旨在解决当前数字评估轨迹(digital assessment traces)在预测学生表现时存在的核心问题:即现有方法常通过跨任务合并数据计算预测值与观测值之间的相关性,但这种相关性可能更多反映的是任务间的差异,而非学生在单个任务内部的真实能力差异。其解决方案的关键在于强调必须进行任务内评估(within-task evaluation),并引入多层次的验证框架,包括交叉验证层级(如leave-one-task-out)、对咨询顺序(consultation order)和关系结构(relational structure)的增量预测贡献检验,以及与置换控制(permutation controls)和随机化控制(randomization controls)的对比。研究发现,尽管将数据跨任务聚合会显著提高相关性(尤其在事实正确性维度),但这些高相关性并未体现学生在具体任务中稳定的相对表现差异;而基于句子嵌入(sentence embeddings)虽在独立任务上表现出与通用回答质量的微弱正相关,却未能证实观察到的咨询顺序或关系结构具有额外预测价值。因此,研究强调应警惕跨任务聚合带来的虚假关联,并呼吁建立标准效度审计机制(criterion generalizability audits),以确保评估模型真正捕捉学生在任务内的真实认知表现。

链接: https://arxiv.org/abs/2611.20803
作者: Maxim Konca,Alexander Mehler
机构: Goethe University(歌德大学)
类目: Computation and Language (cs.CL); Computers and Society (cs.CY); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Digital assessment traces are increasingly used to predict student performance, often evaluated by correlating predicted and observed scores pooled across tasks. However, such correlations may reflect between-task differences rather than performance differences among students within tasks. To investigate this, we analyzed up to 527 browsing sessions from 424 university students completing open-Internet Critical Online Reasoning tasks and assessed the cross-task stability of their ratings. We evaluated the same out-of-fold predictions within and across tasks, compared three validation tiers up to leave-one-task-out, and tested incremental predictive contributions of consultation order and relational structure against matched permutation and randomization controls. None of the twelve prespecified primary tests was significant: on held-out tasks, consulted-page features showed no significant association with relative student performance, nor did observed order or relational structure add predictive value. Pooling observations across tasks substantially increased correlations, particularly for factual correctness, whose scores showed limited evidence of stable between-student differences across tasks. However, sentence embeddings yielded modest positive correlations with generic response quality on held-out tasks, without establishing added predictive value from observed order or relational structure. Our findings show that pooled predictive correlations may not reflect within-task student differences, highlighting the need for within-task evaluation and criterion generalizability audits.

[NLP-51] Nemotron 3.5 Bolt E4B 4GB: An Open and Efficient Small Language Model for Fast Conversational Agents

【速读】: 该论文旨在解决小型语言模型在资源受限设备上实现高效、智能对话系统的需求,特别是针对本地化语音/设备助手、交互式角色及嵌入式/边缘计算场景中的性能瓶颈问题。其核心挑战在于如何在有限的内存(4GB)条件下,同时兼顾推理速度、多轮工具调用能力、指令遵循与输入感知等关键性能指标。解决方案的关键在于通过在Gemma-4-E4B-it基础上进行后训练优化,构建出Nemotron 3.5 Bolt E4B 4GB这一高效小规模模型,并引入DFlash推测解码(speculative decoding)机制,显著提升解码效率与生成吞吐量:在并发数为1时,相比无推测解码的推理方式,每用户解码速度提升2.74倍;在批量大小为1时,生成吞吐量达到Gemma-4-E4B-it使用多任务提示(MTP)的1.21倍。实验表明,该模型在21项基准测试中实现了最高平均准确率,同时在推理与非推理模式下生成的输出标记数最少,充分体现了其在内存效率与智能性能之间的卓越平衡。

链接: https://arxiv.org/abs/2611.20771
作者: NVIDIA:Adi Margolin,Aditya Malte,Aishwarya Chhabra,Alex Hodson,Ameya Sunil Mahabaleshwarkar,Ambrish Dantrey,Amy Shen,Andrii Skliar,Annamalai Chockalingam,Anusha Pant,Ashwath Aithal,Ashwin Kumar,Besmira Nushi,Bhavesh Pawar,Bryan Catanzaro,Bryan Yu,Chantal Hwang,Chenhan Yu,Chris Alexiuk,Conner Currie,Cynthia Gao,David Mosallanezhad,Dina Yared,Ehsan Hosseini Asl,Eileen Long,Erick Galinkin,Felix Friedmann,Giedrius Burachas,Gorkem Batmaz,Grzegorz Chlebus,Guillermo Siman,Haifeng Qian,Hao Guo,Hayley Ross,Hrishith Thadicherla,Hui Wang,Huiying Li,Ike Nnoli,Irina Filatova,Iris Liu,Ji-Eun Lee,Jian Leong,Jian Zhang,Jiaqi Zeng,Joey Conway,Jonathan Cohen,Jose Vergara,Julien Veron Vialard,Kari Briski,Katherine Luna,Kedar Pandurang Potdar,Kevin Ning,Khalil Ben Khaled,Krzysztof Pawelec,Laura Dykes,Leili Tavabi,Lily Lee,Linda Liu,Liwei Ding,Maka Dong,Manoj Kilaru,Maria Bagdasarova,Matthew Lee,Meredith Price,Micah Schaffer,Michael Boone,Michael Fukuyama,Michael Hasper,Nancy Agarwal,Negar Habibi,Oleksii Kuchaiev,Oluwatobi Olabiyi,Pablo Ribalta,Payal Motwani,Pinky Xu,Prasoon Varshney,Pritam Biswas,Puhui Meng,Rameshwar Shivbhakta,Ran Zilberstein,Reka Kovacs,Richard Mazzarese,Rishabh Garg,Rita Fernandes Neves,Rochelle Woods,Ruby Ye,Ruvaid Virk,Ryan Hill,Saeid Motiian,Seph Mard,Shai Ben Nun,Shaona Ghosh,Sharad Yadav,Sudeep Sabnis,Suhas Hariharapura Sheshadri,Terry Kong,Thanh Do,Thinh Le,Udi Karpas
机构: NVIDIA(英伟达)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Animated versions of Figure 3 are available in the blog: this https URL

点击查看摘要

Abstract:We introduce Nemotron 3.5 Bolt E4B 4GB, an open and efficient small language model further post-trained from Gemma-4-E4B-it and optimized for speed and intelligence per GB of memory used for fast conversational agents. The model carries forward Gemma’s strong conversational capabilities and further improves multi-turn tool use, instruction following, and input grounding for fast, local voice/device assistants, interactive personas, and embedded/Edge interfaces. The model combines natural conversation with multi-turn agentic capabilities to power voice/device assistants, interactive characters and robotic/IoT interfaces running on GeForce RTX, Jetson, and RTX/DGX Spark. Experiments on 21 benchmarks show that Nemotron 3.5 Bolt E4B 4GB achieves the highest average accuracy while generating the fewest output tokens in both reasoning and non-reasoning modes, compared with Gemma-4-E4B-it and Qwen3.5-4B. The model is released with a DFlash drafter. With DFlash, Bolt E4B achieves 2.74\times the per-user decode speed of inference without speculative decoding at concurrency~1, and 1.21\times the generation throughput of Gemma-4-E4B-it with MTP at batch size 1.

[NLP-52] Instant Weight-Space Consolidation for Long-Term Memory in LLM s

【速读】: 该论文旨在解决当前生成式智能体(agentic memory systems)中长期记忆机制存在的核心缺陷:尽管通过将经验提取为文本并实现查询时检索,能够赋予大语言模型(LLM)持久记忆能力,但模型本身始终处于静态状态,无法随交互过程进行参数层面的演化。这一局限导致提取错误累积、跨会话的隐式模式难以被捕捉(因单个片段无法表达全局上下文),且记忆构建成本随对话轮次线性增长。虽然微调(fine-tuning)可改变模型权重以实现参数化学习,但其依赖反向传播(backpropagation)在推理阶段执行,计算开销大,并在长历史场景下性能显著下降。为此,本文提出CLAM(Carried Latent Adapter Memory),其关键创新在于引入一种基于循环超网络的参数化融合机制,在无需梯度更新和无需生成式语言模型调用的前提下,仅通过一次前向传播即可将一系列会话激活信息压缩为一个低秩适配器(low-rank adapter),实现无增长的参数化记忆累积;同时辅以一个独立的情景记忆缓冲区(episodic buffer),用于保留原始细节。两者通过门控潜在状态(gated latent state)协同工作,确保知识跨会话传递且存储规模恒定。实验表明,在LoCoMo与LongMemEval-S基准上,CLAM在三个开源模型(2-7B参数)上均显著优于所有基于提示的代理基线(提升1-45分)及基于梯度的参数化方法(提升10-45分),且内存构建时间仅为传统提取系统3-151倍,且不产生任何生成式调用或梯度计算。两个通道具有互补性:当情景缓冲区缺失正确证据时,整合后的适配器仍能带来12-15分的准确率增益,即便在理想缓冲区条件下亦可额外提升4-5分。

链接: https://arxiv.org/abs/2611.20759
作者: Manan Suri,Dinesh Manocha
机构: University of Maryland, College Park(马里兰大学帕克分校)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Agentic memory systems give LLMs persistent memory by extracting experience into text and retrieving it at query time, but the model itself does not evolve: however many interactions it has had, it starts each query the same. Extraction errors compound, implicit cross-session patterns that no single fragment states cannot be surfaced, and the construction cost grows with every conversation. Fine-tuning on past interactions does change the weights, but requires backpropagation at serving time and degrades on long histories. We introduce CLAM (Carried Latent Adapter Memory), which combines the strengths of both: a recurrent hypernetwork performs parametric consolidation, compiling a stream of session activations into a low-rank adapter in a single forward pass with no gradient steps and no language-model generation calls, while a complementary episodic buffer supplies verbatim detail at inference. A gated latent state carries knowledge across sessions in fixed size, so the parametric memory accumulates without growing. On LoCoMo and LongMemEval-S across three open models (2-7B parameters), CLAM leads all agentic baselines by 1-45 points and all gradient-based parametric methods by 10-45 points, while requiring zero generative LLM calls, zero gradient steps, and 3-151x less wall-clock time than extraction systems to build its memory. The two channels are complementary: the consolidated adapter adds 12-15 accuracy points when the episodic buffer misses the gold evidence, and still adds 4-5 points even with an oracle buffer.

[NLP-53] GPRM: Global Perspective Process Reward Model

【速读】: 该论文旨在解决现有过程奖励模型(Process Reward Models, PRMs)在多步推理评估中普遍存在的历史依赖建模能力弱和全局感知不足的问题。传统PRMs通常无法有效利用先前步骤的评估结果进行连贯判断,且缺乏对后续推理路径的洞察,导致对中间步骤的评估不准确。其解决方案的关键在于提出全局视角过程奖励模型(Global Perspective Process Reward Model, GPRM),通过联合建模问题上下文、历史推理步骤及其评估结果、当前步骤以及未来推理步骤,实现具备历史感知与全局信息的推理过程评估。此外,研究设计了一个包含“回溯”“前瞻”“自检”和“目标对齐”四个维度的结构化评估框架,用于系统性衡量推理的一致性、未来合理性、步骤有效性及任务目标契合度,并据此构建了用于监督微调(SFT)和强化学习(GRPO)的数据集。实验表明,尽管训练数据更少、模型规模更小,GPRM在PRMBench和ProcessBench上分别优于强基线7.8和5.4分;同时在跨任务泛化方面表现优异,在Agent Error Bench上超越GPT-4.1基准,在测试时缩放(TTS)条件下较强基线提升约4分,并在PPE任务上达到具有竞争力的性能。

链接: https://arxiv.org/abs/2611.20755
作者: Zifan Zhang,Zhengbo Jiao,Shaobo Wang,Jing Zhou,Guoyi Li,Jiayi Chen,Wei Wang,Bing Zhao,Cheng Fang,Xiaoxiao Xu,Hu Wei,Linfeng Zhang,Lin Qu
机构: Wuhan University (武汉大学); Shanghai University of Finance and Economics (上海财经大学); Shanghai Jiao Tong University (上海交通大学); Fudan University (复旦大学); University of Chinese Academy of Sciences (中国科学院大学); Beijing University of Posts and Telecommunications (北京邮电大学); Alibaba Group (阿里巴巴集团)
类目: Computation and Language (cs.CL)
备注: 20 pages, 3 figures, 6 tables

点击查看摘要

Abstract:Process Reward Models (PRMs) are widely used for step-level evaluation of multi-step reasoning in large language models, but existing methods are often limited by weak historical dependency modeling and insufficient global awareness. They typically fail to leverage prior step evaluations for coherent judgment and lack visibility into subsequent reasoning, leading to suboptimal assessment of intermediate steps. To address this issue, we propose the Global Perspective Process Reward Model (GPRM), which jointly models the problem, previous reasoning steps and their evaluations, the current step, and future steps to enable history-aware and globally informed process evaluation. We further introduce a structured evaluation framework along four dimensions—look back, look ahead, self-check, and goal alignment—to assess prior consistency, future plausibility, step validity, and task alignment. Based on this framework, we construct datasets for supervised fine-tuning (SFT) and reinforcement learning (GRPO). Experiments show that, despite using less training data and smaller model sizes, GPRM outperforms strong baseline models by 7.8 and 5.4 points on PRMBench and ProcessBench, respectively. GPRM also generalizes well across tasks, surpassing the GPT-4.1 baseline on Agent Error Bench, improving strong baselines by around 4 points under test-time scaling (TTS), and achieving competitive performance on PPE.

[NLP-54] Can LLM s Give Good Product Advice? A Framework for Diagnosing Failures

【速读】: 该论文旨在解决大语言模型在生成产品推荐时,难以准确评估其建议是否真正满足用户需求,以及当推荐不当时具体存在哪些缺陷的问题。现有方法仅通过比对在线回复中提及的产品无法有效回答上述问题,因为可能存在多个合适选项,且原始回复本身质量可能不佳。论文提出一种诊断框架,将推荐过程分解为用户明确需求、模型附加假设、候选生成、筛选及最终选择等环节,并识别出19种潜在的失败模式及其对应的证据评估标准。该框架通过分析产品推荐数据和带有外部标注相关性的独立产品选择任务,揭示了“提及”某产品与“选择”该产品之间的本质差异,并阐明了答案提取方式与候选列表构建对最终结果的关键影响,从而为系统性诊断模型推荐偏差提供了可操作的方法论。

链接: https://arxiv.org/abs/2611.20734
作者: Assaf Hallak,Yftah Ziser,Gal Chechik
机构: NVIDIA Research(英伟达研究); University of Groningen(格罗宁根大学); Bar-Ilan University(巴伊兰大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Evaluating product advice from a language model requires assessing whether the recommendation meets the user’s needs and identifying where the model falls short when it does not. However, matching a product mentioned in an online reply cannot answer either question on its own: several products may be suitable, and the recorded answer may itself be poor advice. When shown a community-listed alternative alongside their own sampled car recommendation, six models prefer the alternative in 1.9 to 6.5 percent of eligible disagreements, consistently across both source-blind presentation orders. We propose a diagnostic framework that separates the user’s stated needs, the model’s added assumptions, candidate generation, filtering, and final choice, with 19 possible failure modes and the evidence needed to assess them. We apply this framework to product-advice data and to a separate product-selection task with externally assigned relevance labels. The experiments distinguish naming a community-listed product from selecting it and show how answer extraction and candidate-list construction affect the outcome.

[NLP-55] MemEmo: Evaluating Emotion in Memory Systems

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在长时间交互中因上下文丢失而导致的情感信息处理能力不足的问题。尽管记忆系统被广泛应用于实现持续对话与个性化服务,但其在情感信息的整合与利用方面的能力尚未得到充分评估。为此,论文提出一种情感增强型记忆评估基准(emotion-enhanced memory evaluation benchmark),核心在于构建一个名为 \textbf{MemEmo} 的多维度评估框架,用于衡量主流及前沿记忆系统在情感推理与追踪方面的表现,涵盖情感信息提取、情感记忆更新、情感问答以及情感冲突检测四个关键维度。同时,研究构建了人类似情感记忆(Human-Like Memory Emotion, HLME)数据集,为系统性检验记忆机制提供实证基础。通过对五种代表性记忆系统在HLME数据集上的评估,报告了四项维度下的11个性能指标,实验结果表明所有系统的综合得分最高仅达45%(相对于100%理想值),揭示了当前记忆系统在情感智能层面存在显著短板,凸显了提升情感感知与记忆融合能力的关键挑战。

链接: https://arxiv.org/abs/2611.20733
作者: Peng Liu,Zhen Tao,Jihao Zhao,Ding Chen,Yansong Zhang,Cuiping Li,Zhiyu Li,Hong Chen
机构: Renmin University of China (中国人民大学); China Telecom Research Institute (中国电信研究院); MemTensor (Shanghai) Technology (MemTensor(上海)科技)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Memory systems address the challenge of context loss in Large Language Models during prolonged interactions. However, as an essential tool for warm companionship and personalized services, the effectiveness of memory systems in processing affective information remains underexamined. To bridge this gap, we propose an emotion-enhanced memory evaluation benchmark to assess the performance of mainstream and state-of-the-art memory systems in integrating and harnessing emotional information. First, we design a benchmark called \textbfMemEmo to assess the affective reasoning and tracking capabilities of memory systems across four dimensions: emotional information extraction, emotional memory updating, emotional memory question-answering, and emotional conflict detection. Second, we construct the \textbfHuman-\textbfLike \textbfMemory \textbfEmotion (\textbfHLME) dataset to provide an empirical foundation for scrutinizing memory systems. Finally, We evaluate five representative memory systems on the HLME dataset, reporting 11 performance metrics across four dimensions. Experimental results reveal that aggregating scores for all tested memory systems is up to 45%, far below 100% scale, exposing a profound disparity between memory system and affective intelligence.

[NLP-56] ANCHOR: Cluster-Grounded 5W1H Extraction via Large-Small Model Collaboration

【速读】: 该论文旨在解决事件语义理解中关键的区分性难题,即如何准确识别事件之间的关联关系,并在“为何发生”(why)与“如何实施”(how)之间建立清晰的语义区分。传统单事件槽填充方法难以支持此类细粒度区分,限制了对事件深层语义的理解。为此,论文提出ANCHOR框架,其核心创新在于构建一种基于聚类的联合建模机制:利用大语言模型(LLM)生成候选事件及其初始4W(who, when, where, what)信息,引导事件聚类并锚定主事件;通过共享文档编码器提供跨事件证据支持,结合专门设计的边界检测头和角色评分头,在事件特定约束下分别精准提取“为何”与“如何”信息。该框架实现文档表征复用与空缺角色容错,显著提升5W1H结构化信息抽取的准确性。基于相同原则,研究进一步构建了一个涵盖三语言、七领域、共26,110份文档的标注基准数据集,实验表明ANCHOR在多个数据集上均达到具有竞争力的性能表现。

链接: https://arxiv.org/abs/2611.20728
作者: Hongyu Jiang,Chenang Cai,Lei Gao,Zhao Kang,Ling Tian,Jinchuan Zhang
机构: University of Electronic Science and Technology of China (电子科技大学)
类目: Computation and Language (cs.CL)
备注: 14 pages, 7 figures

点击查看摘要

Abstract:Deep semantic understanding of events requires recognizing related events and distinguishing why an event occurs from how it is carried out. 5W1H extraction provides a structured foundation for such understanding by organizing event information around six fundamental questions. However, single-event slot filling provides limited support for these distinctions. We propose ANCHOR, a cluster-grounded framework that combines a large language model (LLM) with a small language model equipped with discriminative heads for accurate 5W1H extraction. The LLM proposes candidate events and generates their initial 4W records, which guide the selection of an event cluster and the subsequent anchoring of its main event. A shared document encoder provides evidence for joint 4W refinement, and two separate sets of boundary and role-scoring heads then extract why and how under event-specific constraints, respectively. Document representations are reused across events and stages, and unsupported roles can remain empty. Following the same principles, an agent-and-skill annotation workflow constructs a benchmark of 26,110 documents from ten source corpora covering three languages and seven domains. We select nine of the resulting datasets to evaluate ANCHOR against strong baselines using four open-weight LLMs. ANCHOR achieves competitive performance. The datasets and code are available at this https URL.

[NLP-57] IndustrialKnowEval: A Benchmark for Evaluating Multi-Level Industrial Knowledge in LLM s

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在工业领域知识掌握能力评估缺乏系统性基准的问题,尤其关注不同认知需求层次下模型表现的差异。现有研究普遍缺乏针对工业场景中多样化任务类型与认知复杂度的综合性评估体系。为此,作者提出IndustrialKnowEval(IndKnow)基准,涵盖航空航天、电力、冶金等七个工业领域,包含43,282个样本,覆盖21类任务,并按记忆、理解、推理和工程应用四个认知层级进行结构化组织。其解决方案的关键在于:构建由企业领域专家与资深研究人员全程参与的数据集生成与质量控制流程,确保数据的专业性与可靠性;同时引入工业认知能力差距(Industrial Cognitive Capability Gap, ICCG)指标,用于量化基础工业知识与高阶推理及应用能力之间的性能差距,结合绝对性能评估,实现对模型工业智能水平的多维度刻画。该基准为评估和推动大语言模型在工业场景中的实际应用提供了重要支撑。

链接: https://arxiv.org/abs/2611.20695
作者: Haiteng Wang,Haoze Li,Jingwei Guo,Jingbo Zhou,Si Liu,Lei Ren
机构: Beihang University (北京航空航天大学); Baidu Research (百度研究院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 26 pages (13 pages of main text and 13 pages of appendix), 11 figures

点击查看摘要

Abstract:Large language models (LLMs) are increasingly explored for industrial tasks, yet few benchmarks evaluate their mastery of industrial knowledge across different types of cognitive demand. We introduce IndustrialKnowEval (IndKnow), a benchmark for industrial knowledge and engineering question answering. It contains 43,282 samples covering 21 task types across seven industrial sectors, including aerospace, electric power, and metallurgy, organized into four levels of cognitive demand: memory, comprehension, reasoning, and engineering application. Enterprise-based domain specialists and experienced researchers participate in all stages of dataset construction and quality control. We conduct a series of experiments on 18 representative LLMs, with domain experts deeply involved in the evaluation. Building upon IndKnow, we propose the Industrial Cognitive Capability Gap (ICCG) metric to characterize the performance differential between foundational industrial knowledge and higher-order reasoning and application, and we analyze it alongside absolute performance to provide a more complete view of model capabilities. We anticipate that IndustrialKnowEval will support the benchmarking of LLMs in industrial scenarios and promote the development of robust industrial intelligence. Code is available at: this https URL

[NLP-58] Fast at the Frontier Wrong at the Fallback: A Commit-Level Analysis of Threshold Decoding in Diffusion Language Models

【速读】: 该论文旨在解决生成式 AI 中扩散语言模型(Diffusion Language Models, DLMs)在解码过程中因过早或不当提交(commit)掩码位置而导致的准确性下降问题。其核心挑战在于:尽管允许模型在任意位置基于置信度阈值进行提交,且宽窗口机制可加速解码,但频繁的提前提交(尤其是对左侧尚未生成内容的区域)会引入错误,且现有方法中“仅按从左到右顺序提交一个令牌”的策略无法有效缓解此类错误。研究发现,绝大多数令牌仅在前沿(即最左侧未填充位置,frontier)附近才具备足够置信度,因此将提交限制于前沿仍能保持高效性,同时显著提升准确性。关键解决方案是:仅在前沿位置进行提交,避免提前提交导致的错误传播;实验表明,这一策略可恢复大部分因早期提交所损失的准确率,尤其针对由过早生成结束符(end-of-sequence token)引发的截断错误。此外,故障分析显示,回退提交(fallback commit)虽占总提交量的11%-35%,却引发了61%-72%的失败案例,而修复故障起始步骤比修复更早步骤更为有效。

链接: https://arxiv.org/abs/2611.20693
作者: Roman Makarov,Zhe Xu,Jian Kang
机构: MBZUAI(中东人工智能大学); University of Illinois Urbana-Champaign (伊利诺伊大学厄本那-香槟分校)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Diffusion language models (DLMs) can write several tokens per forward pass. Open-weight DLMs commit, in any order, every masked position whose confidence passes a threshold. If none does, a fallback commits the most confident one, and the commits are never revised. Wide windows let the decoder commit far ahead of the text, which should speed up decoding but often costs accuracy. It is unclear where the model is confident and which commits cause errors. We trace every commit of four off-the-shelf DLMs on math, code, translation and summarization, and rerun failing answers with one step changed. Most tokens become confident only near the frontier, the leftmost masked position, so committing only there, still several tokens per step, is about as fast as committing anywhere. The accuracy loss comes from committing tokens before the text to their left, and committing one token per step does not reduce it. The worst case is an early end-of-sequence token that cuts the answer short. Fallback commits are 11-35% of commits but start 61-72% of the failures that left-to-right decoding avoids. Changing only the step where such a failure starts repairs it more often than changing a similar earlier step. Committing only at the frontier recovers most of the lost accuracy on all four models. Our code is available at this https URL.

[NLP-59] LEAF-OPD: Learning to Forget from Privileged Contexts via On-Policy Distillation

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在知识遗忘(unlearning)过程中,如何在有效擦除特定目标知识的同时,保持模型对其他任务的通用能力(Model Utility)这一核心挑战。现有方法多通过抑制目标知识的表达而未明确指导模型在面对相关查询时应如何正确回应,导致遗忘不彻底或模型行为失衡。为此,论文提出一种名为LEAF-OPD(LEArning to Forget from Privileged Contexts via On-Policy Distillation)的新框架,其关键在于引入一个冻结的教师模型(frozen teacher),通过具有特权上下文(privileged contexts,即指令禁止泄露遗忘目标信息)的监督信号,在遗忘数据上引导学生模型学习“拒绝回答”策略;而在保留数据上,教师在无特权上下文条件下提供正常响应监督,以维持模型的实用性。学生模型始终在无特权上下文环境中训练,从而学习在真实场景中恰当地处理应被遗忘的知识。为应对教师监督信号仅在生成序列早期有效的问题,论文设计了监督信号(Supervision Signal, SS),分别评估整个生成轨迹(SS1)及后半段(SS0.5)的监督强度,以衡量监督的持续性。当监督较弱时,采用轻量级监督微调(SFT)增强教师的上下文条件化拒绝行为,显著提升两个指标。实验在TOFU、RWKU和MUSE-News三个主流基准上验证,结果表明,LEAF-OPD在遗忘质量(Forget Quality)、保留性能(Retain Quality)与遗忘文本相似度(Forget ROUGE-L)方面均达到领先水平,尤其在RWKU上比次优方法提升34%的遗忘质量,充分证明了基于特权上下文的在线策略(on-policy)token级蒸馏在实现高效、可控遗忘方面的有效性。

链接: https://arxiv.org/abs/2611.20674
作者: Zirui Pang,Zhijie Deng,Yuancheng Wang,Qihan Yu,Xin Li,Jiaheng Wei,Zixin Zhong
机构: The Hong Kong University of Science and Technology (Guangzhou); The Chinese University of Hong Kong, Shenzhen
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:LLM unlearning aims to erase targeted knowledge from large language models while preserving their utility. However, many existing methods suppress target knowledge without explicitly teaching the model how to respond to queries about it. We propose LEAF-OPD (LEArning to Forget from Privileged Contexts via On-Policy Distillation), in which a frozen teacher provides token-level supervision on student-generated responses. On forget data, the teacher uses privileged contexts (instructions not to reveal knowledge about forget targets) to guide forgetting. On retain data, the teacher provides supervision without privileged context to preserve model utility. The student operates without privileged context in both cases, learning how to respond when encountering knowledge that should be forgotten. A key challenge is that the teacher often provides effective supervision only at the first few tokens of student trajectories. We introduce Supervision Signal (SS) to measure corrective supervision over a specified token interval, evaluating the entire trajectory (SS1) and its second half (SS0.5) to assess overall strength and late-stage persistence. When supervision is weak, lightweight SFT strengthens the teacher’s context-conditioned refusal behavior, substantially increasing both scores. We evaluate LEAF-OPD on three widely used benchmarks: TOFU, RWKU, and MUSE-News. On TOFU, LEAF-OPD achieves the best Forget Quality, Model Utility, and Forget ROUGE-L while maintaining competitive retain performance. On RWKU, it outperforms the second-best method by 34% in Forget Quality while maintaining competitive Retain Quality. These results demonstrate the effectiveness of privileged-context supervision with on-policy token-level distillation for LLM unlearning.

[NLP-60] MetaAct: Self-Evolving Executable Action Constraints for Embodied Language Agents

【速读】: 该论文旨在解决自演化语言智能体在具身环境(embodied environments)中因依赖文本生成而难以确保动作可接受性(action admissibility)的问题,尤其是在具有离散且不可逆动作的场景下,传统基于文本的演化方式仅能影响行为偏好,无法可靠约束动作合法性。其解决方案的关键在于提出MetaAct框架,将自演化机制从文本指导转向可执行的动作约束,通过从执行反馈中合成接地可执行动作规则(Grounded Executable Action Rules, GEAR),将每条规则以类型化规范形式表示,并编译为确定性前置条件(deterministic guard),在离线状态下通过轨迹回放进行验证,无需额外模型调用或环境交互。该方法显著提升了任务性能与动作合规性,在ScienceWorld和WebShop基准上,动作违规率分别从70%降至0%、40%降至1%,同时分别获得+10.1和+8.0的得分提升。

链接: https://arxiv.org/abs/2611.20655
作者: Jiaxing Zhao,Hanwen Wang,Di Liu,Shaoteng Liu,Jing-Ran Su,Yanheng Li,Qian Zha,Jiaxi Li,Hongbin Xie,Haoran Zhang,Xuan Song,Haoxuan Che
机构: Jilin University(吉林大学); Southern University of Science and Technology(南方科技大学); Northwestern Polytechnical University(西北工业大学); The Hong Kong Polytechnic University(香港理工大学); City University of Hong Kong(香港城市大学); Peking University(北京大学); XGEN Labs(XGEN实验室)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Self-evolving language agents have emerged as a promising paradigm for continuously improving agent behavior through experience. However, most existing approaches evolve textual artifacts, which can influence action preferences but cannot reliably enforce action admissibility in embodied environments with discrete and often irreversible actions. We propose MetaAct, which shifts self-evolution from textual guidance to executable action constraints by synthesizing GEAR (Grounded Executable Action Rules) from execution feedback. Each rule is represented as a typed specification, compiled into a deterministic guard, and validated offline through trajectory replay without additional model calls or environment interactions. Across four benchmarks, MetaAct improves task performance; on ScienceWorld and WebShop, violations drop from 70% to 0% and 40% to 1%, with score gains of +10.1 and +8.0. Our code is available at this https URL.

[NLP-61] RAMP: Rethinking Multimodal Evaluation through Answer Readout Protocols

【速读】: 该论文旨在解决多模态大语言模型(Multimodal Large Language Models, MLLMs)在评估过程中因输出读取协议(answer readout protocol)差异而导致结果不可靠的问题。现有评估方法通常将模型输出转化为离散答案,而这一转换过程中的具体读取方式(如直接使用“是/否”标记的对数概率、表面形式得分或类型化/反转的A/B对比接口)会显著影响最终报告的性能指标,导致结果缺乏可比性和可重复性。论文提出RAMP(Readout Aware Multimodal Protocol),一个系统性的评估框架,通过在固定模型检查点和成对视觉问答数据上对比多种读取策略,揭示了读取协议对模型表现的实质性影响:例如,在500个对抗性POPE问题上,不同读取方式使SmolVLM-256M的准确率波动达-20.2个百分点,而Qwen2.5-VL-3B则出现+1.8个百分点的提升。此外,研究发现即使净准确率变化微小,预测结果仍可能发生广泛转变(如SmolVLM-500M在ChartQA中79/104个二元问题的正确性发生变化)。研究还系统分析了提示词表述、候选形式、选项位置、跨任务效应及校准等因素,并通过独立离线回放验证了16,140次首次词决策的精确复现性。其核心贡献在于将“答案读取”确立为多模态评估中可测量的关键环节,并提供了一套可复现的协议,以增强模型比较的透明性与科学性。

链接: https://arxiv.org/abs/2611.20653
作者: Jianye Zeng,Qiao Yikang,Duan Huang,Ziying Song
机构: University of Science and Technology of China(中国科学技术大学); Central South University(中南大学); Yanshan University(燕山大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
备注: 10 pages, 9 figures, 1 table

点击查看摘要

Abstract:Multimodal large language models (MLLMs) are often evaluated by converting model outputs into discrete answers. The conversion protocol determines which scores become predictions and can strongly influence the reported results. We introduce RAMP (Readout Aware Multimodal Protocol), a systematic framework that compares direct yes/no token logits, surface-form scores, and typed or reversed A/B interfaces under fixed model checkpoints and paired visual questions. On 500 POPE adversarial questions, typed versus bare readout changes accuracy by -20.2 percentage points for SmolVLM-256M, -16.4 for TinyLLaVA-1.1B, and +1.8 for Qwen2.5-VL-3B. Paired transitions reveal extensive prediction changes even when the net accuracy difference is small: 79 of 104 binary ChartQA items change correctness for SmolVLM-500M, despite a +1.0 -point net difference. We further analyze prompt wording, candidate forms, option position, cross-task effects, and calibration. An independent offline replay reproduces 16,140 first-token decisions from saved raw scores with exact agreement. Our findings establish answer readout as a measurable component of multimodal evaluation and provide a reproducible protocol for analyzing model comparisons.

[NLP-62] Safety-Preserving Medical LLM Fine-Tuning via Bilevel Data Selection

【速读】: 该论文旨在解决在将通用大语言模型(Large Language Models, LLMs)应用于面向患者的医疗场景时,如何在提升医学能力的同时有效控制患者端风险的关键挑战。现有方法依赖大规模医疗指令数据进行监督微调(Supervised Fine-Tuning, SFT),但未经筛选的训练数据可能引入不安全响应,破坏已有的安全对齐。为此,本文提出可扩展的双层数据选择(Scalable Bilevel Data Selection, S-BDS)方法,其核心在于通过控制选择权重的集中度,实现对大规模候选数据集的可靠样本排序,从而精准识别既能增强医学能力又具备低患者风险的训练样本。S-BDS结合临床医生验证的安全数据与经筛选的未筛查医疗指令数据进行联合微调,并在HealthBench和PatientSafetyBench上分别评估医学能力与患者端安全性。实验结果表明,S-BDS在多个Llama和Qwen模型上均实现了最高安全率,相较于最强基线分别提升4.64和3.00个百分点,同时保持了优异的医学性能。进一步分析显示,微调过程中调节安全导向权重可灵活调控医学能力与患者安全之间的权衡关系。

链接: https://arxiv.org/abs/2611.20643
作者: Ruoxuan Shi,Sheng Yang,Zhengxing Su,Xiaoyang Hou,Yating Liu
机构: Lanzhou University (兰州大学); University of California, Riverside (加州大学河滨分校); Lanzhou University Second Hospital (兰州大学第二医院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:As general large language models (LLMs) are increasingly adapted for patient-facing medical applications, improving medical capability while controlling patient-facing risks remains a critical challenge. Large-scale medical instruction data are essential for acquiring domain knowledge through supervised fine-tuning (SFT), but unsafe responses can undermine existing safety alignment. To address this problem, we propose Scalable Bilevel Data Selection (S-BDS) for safety-preserving medical SFT. By controlling selection-weight concentration, S-BDS enables reliable sample ranking over large candidate pools and identifies training examples that improve medical capability while controlling patient-facing risks. We combine clinician-validated safe data and selected data with unscreened medical instruction data for SFT, and jointly evaluate medical capability and patient-facing safety on HealthBench and PatientSafetyBench, respectively. Experiments across Llama and Qwen models show that S-BDS achieves the highest safe rates among all compared methods, outperforming the strongest baselines by 4.64 and 3.00 percentage points, respectively, while maintaining competitive medical capability. Further experiments show that varying the safety-oriented weight during SFT controls the trade-off between medical capability and patient-facing safety.

[NLP-63] Same Papers Different Verdicts: Tracing Version Drift in LLM Peer Review ICLR2027

【速读】: 该论文旨在解决大模型(Large Language Models, LLMs)在学术评审中评分一致性与版本迭代导致评估结果不可靠的问题。随着模型提供商频繁更新其模型(如OpenAI、Google、Anthropic),同一模型的不同版本在相同论文上的评审结果存在显著差异,而当前学术会议(如ICLR 2027)仍依赖这些评分作为录用决策的依据,这可能导致评审标准的不一致甚至失效。研究的关键在于通过前瞻性审计(pre-decision audits)对20个前沿模型在论文投稿期间的多个版本进行系统性测试,发现同一模型不同版本之间在录用倾向上的差距可达数十个百分点(如Claude Opus和Gemini分别下降52和53个百分点),且无任何论文因模型更新而被提升至接受侧;进一步实验表明,即使移除评价性语言仅保留技术内容,模型评分仍显著下降,说明评分偏差主要源于模型内部判断机制的演变而非表述风格。此外,研究揭示了模型评分与人类评审的一致性有限,且多数模型在绝大多数评审中给出极端低分(2或3),远低于人类审稿人的分布,凸显了基于阈值的自动化评审系统必须针对每个模型版本独立验证其与人工决策的一致性。因此,解决方案的核心是建立动态、版本敏感的评估验证机制,确保评审标准随模型演进同步校准,以保障学术评审的公平性与可复现性。

链接: https://arxiv.org/abs/2611.20608
作者: Jizhou Guo
机构: Independent Researcher(独立研究员)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Digital Libraries (cs.DL); Physics and Society (physics.soc-ph)
备注: 19 pages, 5 figures. Results of the registered tests that need the ICLR 2027 decisions will be added after 16 December 2026

点击查看摘要

Abstract:Venues and benchmarks read LLM review scores as measurements of a paper, while providers replace their models within months. We audit twenty frontier models from OpenAI, Google and Anthropic on ICLR 2027 submissions while they were under review, with tests registered before the decisions. Between an earlier and a later version of Claude Opus and of Gemini, the share of papers placed on the accept side falls by 52 and 53 percentage points on 183 and 60 randomly drawn submissions, and by 43 to 79 points in six further paired comparisons, with no paper moving up in any of them. Stating a 30% acceptance rate in the prompt leaves a gap of 79 points between the two Opus versions. When score-stripped critiques are swapped between versions, the later critiques alone lower the accept share by 35 points for Opus and 51 for Gemini; for Opus the drop is 35 points under a third-party scorer and 24 after a rewrite that keeps each technical point and removes evaluative wording. On 57 papers, the percentiles of Opus 4.6 and 5.5 correlate at 0.73 while their accept shares differ by 44 points. On 99 to 119 submissions, the self-estimated percentiles of three models correlate with two human reviewers more strongly than a third reviewer does (point estimates 0.17 to 0.21, 95% intervals between -0.04 and +0.42). All twenty models give a 2 or a 3 in 473 of their 475 reviews, against 60 of 72 human reviews. Retrospective tests on decided papers can draw on memory. Claude Opus 5.5 names the true first author of 27 decided ICLR 2026 papers from their titles alone, with no wrong name, and Claude Opus 4.6 accepted all 11 decided papers it said it recognized, against 5 expected. Thresholds on LLM-reviewer scores therefore have to be validated against human decisions for each model version; our tests against the ICLR 2027 decisions follow on 16 December 2026.

[NLP-64] Necessary for What? Auditing Demographic Disclosure to LLM Synthetic Respondents

【速读】: 该论文旨在解决生成式 AI 在替代真实受访者回答调查时,是否必须依赖个体记录的详细人口统计信息(如年龄、教育程度、族裔等)这一关键问题。其核心关切在于:在数据最小化原则下,为实现多种研究目的(如个体回答预测、意见分布估计及群体比较),是否真需完整个人属性信息?研究的关键解决方案在于系统性地测试不同模型在接收不同来源与维度的人口统计信息时的表现,并与基于相同信息构建的统计模型进行对比。研究发现,个体预测准确性对记录依赖度极低,而群体估计中的多数增益可通过普查数据或合成人格替代;群体比较仅在使用个体真实记录时有所提升,但模型所识别的组间差异更多源于提示中的人口标签而非个体真实特征;更严重的是,模型倾向于弱化甚至反转态度差异,尤其在宗教信仰与族裔起源等敏感类别上,导致群体间差异被人为“扁平化”。因此,论文提出一套可操作的审计框架,供调查机构评估已有问题是否应继续使用个体记录以确保结论的稳健性与伦理合规性。

链接: https://arxiv.org/abs/2611.20606
作者: Aspen Wang,Yen Yu Shan,Lun-Wei Ku
机构: Institute of Information Science, Academia Sinica(中央研究院資訊科學研究所), Taipei, Taiwan
类目: Computers and Society (cs.CY); Computation and Language (cs.CL)
备注: 29 pages, 10 figures, 11 tables; Code: this https URL

点击查看摘要

Abstract:Large language models (LLMs) increasingly answer surveys on behalf of real panelists, and each panelist’s recorded demographic attributes (e.g., age, education, ethnicity) are often placed in the prompt on the premise that richer profiles make more faithful synthetic respondents. While data minimization limits personal data to what a purpose requires, a synthetic survey serves several purposes at once, from predicting individual answers to estimating how opinions are distributed and comparing groups. We ask which of these purposes require the respondent’s record, and whether improved estimates support sound conclusions about groups. We vary whose profile a model receives, where it comes from, and how many fields it contains, in two U.S. surveys with eleven open-weight and three hosted models, and compare each model with a statistical model fit to the same information. We make four contributions. First, we show that the record adds little to individual predictions and that, for the open-weight models, most of its gain in population estimates remains when it is replaced by census data or a synthetic persona. Second, we find that group comparisons improve only with the respondent’s own record, and that even then the models’ group differences rest more on the demographic label in the prompt than on the individual respondent, which the usual check of reproducing known group differences does not reveal. Third, we find that the models make groups look more alike than they are, keeping differences in circumstances while flattening or reversing differences in attitudes, most sharply for religious affiliation and ethnic origin, both special categories of personal data. Fourth, we outline an audit that a survey organization can run on questions it has already fielded to decide which estimates may use the record.

[NLP-65] Invariant-Aware Harness Enforcement to Protect LLM Agents from Silent API Failures

【速读】: 该论文旨在解决大语言模型(LLM)代理在调用API时因“静默失败”(silent failure)而导致错误响应的问题。静默失败指API在未正确处理请求(如忽略错误命名的过滤器)的情况下仍返回HTTP 200状态码和看似合理但实际错误的结果,使模型难以察觉并继续基于错误数据生成回答。现有解决方案如检查清单、评审模型和LLM判别器依赖模型自身识别问题,但其往往无法有效检测此类缺陷。本文提出SilentGuard,其核心创新在于将决策机制从模型迁移至工具调用的运行时框架(harness),通过在每次API调用后对结果施加形式化不变量(invariant)验证,自动识别并屏蔽违反正确性约束的输出值。实验表明,该方法在15个公开研究型API中发现11个存在静默失败,部分指标被高估达691倍;在7个开源模型上,错误答案率从74.0%降至1.7%,前沿模型则从70.2%降至0.5%,且不影响无故障任务的表现。相比之下,仅发出警告的方案仍导致70.1%的答案错误,凸显强制执行(enforcement)的关键作用。此外,针对新型故障类型,该框架独立即能将错误率从61.1%降至24.5%,通过主动拒绝(abstain)而非尝试修复来避免错误传播。

链接: https://arxiv.org/abs/2611.20595
作者: Sajib Acharjee Dip,Brian Ouyang,Lala Motlhabi
机构: Genentech(基因泰克); Virginia Tech(弗吉尼亚理工学院)
类目: Computation and Language (cs.CL)
备注: 26 pages, 7 figures. Code: this https URL . Data: this https URL

点击查看摘要

Abstract:LLM agents answer questions by calling APIs and trusting the results. An API can fail silently. It ignores part of a request, such as a misnamed filter, yet returns HTTP 200 and a plausible but wrong result. Current fixes, such as checklists, reviewer models and LLM judges, ask the model to catch it. But models often see the problem and answer anyway. We present SilentGuard, which moves the decision from the model to the harness running tool calls. It tests each result against invariants that any correct API must satisfy and withholds values that break one. We find silent failures in 11 of 15 public research APIs, with counts inflated up to 691 times. Current fixes leave over 70% of answers wrong, and agents report 77.4% of counts a judge flags. Across 7 open models, SilentGuard cuts wrong answers on its target faults from 74.0% to 1.7%, with no loss on fault-free tasks. On three frontier models they fall from 70.2% to 0.5%. A warning from the same check leaves 70.1% wrong, so enforcement is the key. On new fault types, the harness alone cuts wrong answers from 61.1% to 24.5%. It avoids wrong answers by abstaining rather than recovering.

[NLP-66] Language Models Meet Vibration Testing for LEO Constellation Satellites with Calibrated Response Prediction from Few-show Sweeps

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在工程分析中因生成幻觉(hallucination)而导致错误输入未被察觉的问题,尤其关注由文本记录转化为数值程序输入过程中可能出现的物理不一致、单位错误或虚构内容。其核心解决方案是提出一种确定性验证门控机制(deterministic validation gates),通过一系列严格且可复现的检查环节,确保LLM生成的配置数据在被使用前满足多重约束条件,包括单位一致性、物理合理性、原文引用匹配性、独立模型间的一致性、与实测数据导出结果的吻合度以及训练模型的应用有效性。该机制从两个方向防范幻觉:引用校验防止伪造内容,多模型一致性校验弥补引用校验对遗漏型错误的盲区。研究以近地轨道巨型星座卫星的正弦振动验收测试为场景,采用DeepSeek V4.1 Flash和Qwen-plus解析中文测试记录并提取76个字段配置,在五种输入变体下事实级F1得分达0.98至1.0,所有修复均在开发集上冻结,仅在保留集上评估。启用全部验证门后,20组配置的输出与基准结果逐比特完全一致,且17个真实及注入错误中的全部均被拦截(其中16个在无门控时将被放行)。此外,基于LLM配置重建的模型重训练结果与原模型在GPU非确定性范围内保持一致,验证了系统的可复现性。研究还报告了一个具有实践意义的负面发现:直接使用LLM进行曲线预测仅能逼近线性外推效果;而引入一个小型物理结构化网络,可在校准置信度c=0.90下将过切带宽降低86%至89%,明确界定了何时应以受训模型替代直接LLM预测。

链接: https://arxiv.org/abs/2611.20579
作者: Jun Wu,Deming Li,Zhenming Zheng,Honghao Yue,Fei Yang,Hao Sun
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Large language models (LLMs) increasingly drive engineering analysis end to end, turning textual records into the inputs of numerical programs, but a plausible and wrong value can pass silently into a decision. We propose deterministic validation gates for LLM-driven engineering analysis, a general paradigm in which a typed configuration produced by LLMs is consumed only after it passes deterministic checks on units, physical consistency, verbatim quotes, agreement between independent models, consistency with measurement exports and the applicability of trained models. The gates target hallucinations from two directions, since quote checking catches fabricated content while agreement between models catches errors of omission that quote checking misses. We instantiate the paradigm on sine vibration acceptance testing for low Earth orbit mega-constellation satellites, where DeepSeek V4.1 Flash and Qwen-plus read Chinese test records and extract a configuration of 76 fields. Fact-level F1 ranged from 0.98 to 1.0 on five input variants, with all repairs frozen on development inputs and evaluated on held-out inputs. With all gates, 20 of 20 configurations produced notching results bit for bit identical to the gold run, and all 17 real and injected errors in consumed fields were stopped, of which 16 would have run without gates. Retraining the bound model on a dataset built from an LLM configuration reproduced its results within GPU nondeterminism. We also report a negative result of practical value. Direct LLM curve forecasting only matches linear extrapolation, whereas a small physics-structured network reduces the over-notched bandwidth by 86% to 89% at calibrated coverage c = 0.90, which delineates where direct LLM prediction should give way to a trained model.

[NLP-67] Weight-Space Programming Changes How an LLM Uses Python for Computational Biology

【速读】: 该论文旨在解决大语言模型在特定任务(如计算生物学问题)中因推理错误或输出格式不规范而导致的准确率不足问题,尤其关注如何在不进行梯度训练的情况下提升模型的计算准确性与输出可靠性。其解决方案的关键在于采用权重空间编程(weight-space programming),通过直接修改目标模型(Qwen3.5-0.8B)的权重参数,使其在执行Python代码求解计算生物学问题时更倾向于生成正确的中间计算过程并提交技术上正确的最终答案。实验结果表明,经过权重编辑后,严格正确答案比例从44%上升至77%,技术正确答案(即正确数值但可能因格式问题未被接受)从64%增至100%,同时重复调用Python的次数显著减少;进一步分析显示,编辑后的模型不仅修正了原始模型已计算出正确值但未正确提交的问题,还提升了对复杂问题(如熔解温度计算)的输出质量,说明权重编辑能够有效改变模型的内部计算逻辑与输出行为策略。

链接: https://arxiv.org/abs/2611.20566
作者: Makoto Sato
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Quantitative Methods (q-bio.QM)
备注: 16 pages, 3 figures, 6 main tables, and 1 supplementary table

点击查看摘要

Abstract:Weight-space programming allows a language model to edit another model’s weights without gradient-based training. We used Claude Opus 5.5 to edit Qwen3.5-0.8B and investigated how the edits changed its use of Python on computational biology problems. We compared four initial edited checkpoints with the original model on 50 test problems across three test seeds, yielding 600 paired evaluations. Strictly correct answers increased from 44 to 77, and technically correct answers, which included correct final values rejected for formatting, increased from 64 to 100. Repeated Python calls decreased from 1,834 to 1,044. Among the 64 pairs in which only the edited response was technically correct, the original model had already computed the correct value in 36 but did not submit a technically correct final answer. The edited model corrected the computation in 13 of the 64 pairs. With six additional checkpoints on the same 50 problems, technically correct answers rose from 96 to 154 across 900 pairs. We explored three selected problems with ten test seeds for each of ten checkpoints. On the melting-temperature problem, correct Python outputs decreased from 60 to 22 while technically correct final answers increased from 0 to 18. Weight editing changed both the calculations Qwen ran and whether it submitted a correct answer.

[NLP-68] ReconPotBench: Evaluating LLM Honeypots in Multi-Round Linux Reconnaissance EACL2027

【速读】: 该论文旨在解决当前大型语言模型(LLM)驱动的交互式蜜罐系统在多轮交互中缺乏可信性评估的问题。现有评价方法仅关注单轮命令输出的合理性,无法衡量系统在后续指令依赖前期结果时是否仍能保持上下文一致性与行为连贯性。为此,作者提出ReconPotBench——一个针对Linux侦察类蜜罐的多轮基准测试框架,包含8个侦察类别、共1,032组三轮交互数据,参考输出源自真实Linux环境。其核心解决方案在于构建统一测试协议,从三个维度评估系统表现:每轮的逼真度(round-level authenticity)、跨轮次的上下文一致性(cross-round contextual consistency)以及面向攻击者视角的侦察价值(hacker-oriented reconnaissance value)。实验表明,多轮评估揭示了单轮测试难以发现的缺陷,如随着交互深度增加,生成内容的真实性下降,且无论是开源还是闭源大模型均难以维持跨轮状态的一致性;然而,基于LLM的蜜罐在逼真度和侦察价值上仍优于传统蜜罐工具Cowrie,后者因采用固定模板行为而具备更强的一致性。进一步分析显示,需在多轮中保持网络关系连贯性的任务比仅涉及显式实体(如端口)的任务更具挑战性。ReconPotBench为评估基于大模型的蜜罐系统在真实性、实体一致性及攻击者实用性方面的表现提供了可复现的基准。

链接: https://arxiv.org/abs/2611.20528
作者: Qianli Wang,Junchao Wu,Binkai Ou,Lidia S. Chao,Derek F. Wong
机构: NLPCT Lab, Faculty of Information Science and Computing, University of Macau(澳门大学信息科学与计算学院自然语言处理与计算技术实验室); School of Computer Science and Technology, Shandong University(山东大学计算机科学与技术学院); BoardWare Information System Limited(博维信息科技有限公司)
类目: Cryptography and Security (cs.CR); Computation and Language (cs.CL)
备注: EACL 2027 Finding

点击查看摘要

Abstract:Large language models (LLMs) are increasingly used to build interactive honeypots, yet existing evaluations mainly assess isolated command outputs and cannot determine whether a system remains believable when later commands depend on earlier results. We present ReconPotBench, a multi-round benchmark for Linux reconnaissance honeypots. It contains 1,032 three-round reconnaissance interactions across eight reconnaissance categories, with reference outputs collected from a real Linux environment. Under a unified testing protocol, we compare Cowrie, open-source LLMs, and closed-source LLMs along three dimensions: round-level authenticity, cross-round contextual consistency, and hacker-oriented reconnaissance value. Results show that multi-round evaluation exposes failures hidden by single-round testing, as authenticity declines with interaction depth and both open-source and closed-source LLMs struggle to maintain coherent state across rounds. LLM-based honeypots nevertheless outperform Cowrie in authenticity and reconnaissance value, while Cowrie remains more consistent because of its rigid templated behavior. Category-level analysis further shows that tasks requiring systems to preserve network relationships across rounds are harder than tasks centered on explicit entities such as ports. ReconPotBench enables reproducible evaluation of LLM honeypots across authenticity, entity consistency, and hacker-perspective utility in dependent multi-round interactions. The dataset is available at: this https URL

[NLP-69] GaugeRoPE: RoPE-Compatible Executable Latent Attention

【速读】: 该论文旨在解决生成式 AI(Generative AI)中因注意力机制导致的显存占用过高问题,特别是解码阶段键值缓存(KV cache)带来的内存压力。现有方法通过潜在空间压缩(latent KV cache compression)降低存储开销,但通常需将压缩后的键值状态重构回原始维度后才能应用旋转位置编码(RoPE),这一过程引入了与序列长度相关的计算开销,影响推理效率。其核心解决方案在于:提出一种无需训练、可直接执行的潜在键值表示——GaugeRoPE。该方案的关键是构建一个满足预训练 RoPE 操作不变性的潜空间映射机制,即要求潜在表示能保持原始 RoPE 的旋转变换特性,这依赖于潜在空间与原空间之间通过“交织关系”(intertwining relation)维持频率一致的坐标混合。GaugeRoPE 由此设计出一个兼容 RoPE 的载体结构以保留旋转动作,并引入近似的、与位置无关的残差键-值(RKV)修正项来补充缺失的信息,且所有运算均在潜在空间内完成,避免逐标记密集重构。实验表明,在 Mistral-7B 上,GaugeRoPE 实现了 5.85 倍的 KV 缓存压缩比,同时保持接近全缓存基线的质量;在多头注意力(MHA)场景下,相比未压缩基线,最高实现 2.48 倍的解码吞吐提升,验证了其在通用注意力架构中的有效性。

链接: https://arxiv.org/abs/2611.20524
作者: Yuanhang Gao,Xiangrui Yang,Lingyuan Meng,Yuanfeng Chen,Dongsheng Li
机构: National University of Defense Technology (国防科技大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Latent KV cache compression reduces memory footprint by mapping historical keys and values into compact representations. However, generic compressed key states require reconstruction to native dimension before applying RoPE and computing attention scores, adding sequence-scaling overhead to the decode path. We characterize when a latent KV representation can serve directly as the execution state of pretrained RoPE. For RoPE-based transformers, exact static linear execution of the rotary key path requires the latent-to-native map to preserve the pretrained RoPE action through an intertwining relation, restricting exact mixing to coordinates that share the same RoPE frequency. Guided by this characterization, we build GaugeRoPE, a training-free executable latent KV representation. It combines a RoPE-compatible carrier that preserves the rotary action with an approximate position-independent residual key–value (RKV) correction for complementary key and value information, both evaluated without per-token dense historical K/V reconstruction. On Mistral-7B, GaugeRoPE achieves 5.85\times KV compression while maintaining quality close to the full-cache baseline. Additional MHA results validate the construction beyond GQA. GaugeRoPE achieves up to 2.48\times the decode throughput of the vanilla uncompressed baseline through direct latent execution.

[NLP-70] Abstention Direction in Large Language Models : Mechanistic Interpretability of Abstention Across Query Ambiguity Knowledge Limits and Safety Alignment

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)在面对不确定性或安全风险时选择不回答(即“拒答”或abstention)的内在机制问题,尤其关注导致拒答行为的三大根源:查询模糊性(Query Ambiguity, Q)、模型知识局限性(Model Knowledge Limit, MK)以及与人类价值观的错位(Human Values Misalignment, HV)。现有研究多聚焦于人类价值观维度,通过分析有害与无害查询间的差异来构建“拒答方向”,但缺乏对三者协同作用的统一分析框架。为此,本文提出将拒答方向框架扩展至三个维度,针对每个视角构造仅在诱发拒答因素上不同的反事实语句对,并通过标准化两极语句在中间层激活均值的差异,计算出各视角下的拒答方向。激活引导实验表明,在中间层干预可同时影响三种情境下的拒答行为,且存在跨视角的效应,揭示了不同拒答方向共享一个共同的方向性成分。因此,解决方案的关键在于建立一个统一的、可量化的拒答方向分析框架,实现对多种拒答动因的联合建模与干预,从而为提升模型可信度提供可解释的调控路径。

链接: https://arxiv.org/abs/2611.20441
作者: Koki Itai,Jun Hirako,Hiroaki Yamagiwa
机构: SB Intuitions Corp.(SB智能公司)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Abstention, the withholding of a response by a large language model (LLM) under uncertainty or safety risks, is essential for trustworthy models. The causes of abstention fall into three perspectives: ambiguity in the Query (Q), limits of Model Knowledge (MK), and misalignment with Human Values (HV). However, prior analyses of internal representations have focused on HV, using a refusal direction computed from harmful and harmless queries. No study has analyzed the three perspectives jointly. We therefore extend the refusal direction framework to all three perspectives. Specifically, for each perspective, we construct counterfactual pairs that share a template and differ only in the abstention-inducing factor, and compute an abstention direction by normalizing the difference in mean activations between the two polarities. Activation steering experiments show that intervening in middle layers increases or decreases abstention across all three perspectives. Moreover, steering yields cross-perspective effects, indicating that the abstention directions share a common directional component.

[NLP-71] rainExplorer: Autonomous LLM Post-Training via Plan-and-Decouple Exploration

【速读】: 该论文旨在解决基于大语言模型(LLM)的智能体在模型后训练自动化过程中面临的两大核心问题:一是“策略锁定”(strategy lock-in)问题,即现有方法受限于初始策略,难以实现对多样化训练策略的有效探索;二是“耦合执行”(coupled execution)问题,即各训练修改项相互交织,导致无法准确归因单个组件调整的效果。为此,论文提出TrainExplorer,一种用于自主后训练的多智能体框架,其关键创新在于引入策略规划与执行解耦机制。具体而言,策略规划阶段通过整合科学文献构建候选训练策略的搜索空间,以拓展探索范围;执行阶段则通过解耦设计实现对单一组件的独立干预与可解释性归因,从而支持高效、精准的策略迭代。在PostTrainBench上的大量实验表明,TrainExplorer取得了39.02的加权综合得分,显著优于使用相同LLM的Claude Code(提升7.32分),并超越GPT-5.6 Sol-powered Codex(36.23分),验证了其在促进递归自我改进方面的有效性。

链接: https://arxiv.org/abs/2611.20430
作者: Xuannan Liu,Tian Lan,Zekun Li,Xing Cui,Zhengfeng Yu,Feihu Jiang,Qianghuai Jia,Longyue Wang,Zhao Xu,Weihua Luo,Peipei Li,Ran He
机构: Beijing University of Posts and Telecommunications(北京邮电大学); Alibaba Group(阿里巴巴集团); Minzu University of China(中央民族大学); Beijing Institute of Technology(北京理工大学); Center for Research on Intelligent Perception and Computing, NLPR, CASIA(中国科学院自动化研究所智能感知与计算研究中心)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:LLM-based agents can automate model post-training, accelerating progress toward recursive self-improvement. Despite their promise, existing approaches remain under-explored and face two main challenges: (1) strategy lock-in that limits exploration beyond the initial strategy; and (2) coupled execution that obscures the effects of individual modifications. To address these challanges, we propose TrainExplorer, a multi-agent framework for autonomous post-training, that incorporates strategy planning and execution decoupling, to achieve broaden strategy exploration. First, strategy planning draws on scientific literature to construct a candidate search space of training strategies. Then, decoupled execution enables single-component interventions and attribution to guide strategy exploration. Extensive experiments on PostTrainBench demonstrate that TrainExplorer achieves a weighted aggregate score of 39.02, outperforming Claude Code using the same LLM by 7.32 and surpassing GPT-5.6 Sol-powered Codex, which scores 36.23. Code will be available.

[NLP-72] When the Cheapest Feasible Router Fails: Auditing Deployment Commitments in Cost-Aware LLM Routing

【速读】: 该论文旨在解决生成式 AI(Generative AI)系统中成本感知的大语言模型(LLM)路由在实际部署中出现的“部署承诺失效”问题。具体而言,现有路由机制在审计阶段选择满足经验风险与覆盖约束的最廉价策略,但在面对新请求时,该策略可能违反预先签订的服务合同(service contract),即使候选策略是独立构建的。其核心问题是:审计通过的最优策略在部署后无法保证对未见请求的合同合规性,导致服务可靠性与成本控制目标之间的冲突。解决方案的关键在于引入菜单感知路由承诺(Menu-Aware Routing Commitment, MARC),该方法基于审计证据和可接受的运行成本容差,动态决策是否保留原最优策略、部署支持性替代方案或推迟决策,从而在保障服务合同履约的前提下实现稳健部署。其实际可操作版本MARC-P在85.79%的部署率下将条件失败率降至0.73%,显著优于传统方法,验证了将部署承诺作为连接路由评估、服务可靠性与推理成本的独立运营决策的必要性。

链接: https://arxiv.org/abs/2611.20417
作者: Minxuan Hu,Qishi Zhan,Ziheng Chen,Zhengtao Yao,Zihan Dong,Jiayu Yi,Wenxi Sun
机构: Cornell University(康奈尔大学); Marquette University(马凯特大学); University of Texas at Austin(德克萨斯大学奥斯汀分校); University of Southern California(南加州大学); Georgia Institute of Technology(佐治亚理工学院); Nanyang Technological University(南洋理工大学); Johns Hopkins University(约翰霍普金斯大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Cost-aware LLM routing often selects the cheapest policy meeting empirical risk and coverage constraints. Yet a router that passes an audit can violate its fixed service contract on new requests, even when candidate policies are constructed independently. We study this deployment-commitment decision with disjoint construction, audit, and test samples. Across 19 RouterBench task families, the cheapest audit-feasible router deploys in 91.82% of opportunities but violates the held-out contract in 32.22% of deployments. We introduce Menu-Aware Routing Commitment (MARC), which qualifies policies using audit evidence and a serving-cost tolerance to retain the apparent winner, deploy a supported alternative, or defer. Its practical policywise variant (MARC-P) reduces observed conditional failure to 0.73% at 85.79% deployment, with higher serving cost; fixed-margin and formal risk-control procedures provide competitive alternatives. Nested audits show how added evidence changes deployable policy sets, and learned semantic routing exhibits the same commitment failure. These results establish deployment commitment as a distinct operational decision connecting router evaluation, service reliability, and inference cost.

[NLP-73] Augmenting Pretrained Language Models with N-gram Embeddings through Incremental Training

【速读】: 该论文旨在解决现有n-gram嵌入(n-gram embedding)技术在实际应用中受限于必须与主干模型(backbone)联合训练的问题,从而限制了其灵活性和适用场景。为突破这一瓶颈,论文提出通过增量训练(incremental training)的方式将n-gram嵌入引入已训练好的模型中,以实现对已有模型的高效增强。其解决方案的关键在于:在增量训练框架下,系统探索学习率、词表规模及注入位置等关键超参数的最优配置,并发现该最优配置与预训练阶段存在差异;更重要的是,增量训练过程中,新增的知识主要被编码至n-gram嵌入模块,而主干模型的通用能力得以有效保留,从而在不显著增加推理开销的前提下,显著提升了模型的局部上下文建模能力。

链接: https://arxiv.org/abs/2611.20415
作者: Shaohua Hu,Xusheng Gu,Haofei Yang,Minghao Shao,Nanhui Zhao,Caoyun Fan,Fangming Li
机构: Shanghai Jiao Tong University (上海交通大学); Huawei Technologies Co., Ltd (华为技术有限公司)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:As a new scaling technique for large language models, the n-gram embedding improves the base ability of a model while adding almost no inference cost. Mainstream research realizes it through joint training of the n-gram embedding with the backbone, which severely restricts the settings in which an n-gram embedding can be used. We therefore propose training an n-gram embedding onto an existing model through incremental training, to extend where the technique can deliver value. Specifically, we explore several settings of the n-gram embedding under incremental training—the learning rate of the embedding, the vocabulary size, and the injection position—and train an n-gram embedding onto the existing model under the best configuration. Experiments show that the best configuration under incremental training does not fully coincide with that under pretraining; moreover, the incrementally trained knowledge is written predominantly into the n-gram embedding, while the general ability of the backbone is retained throughout this training process.

[NLP-74] Refusal Register is Causally Steerable: Decomposing the Refusal Direction into Category-Aligned Components in Gemma-2-9B-IT ICML-2026

【速读】: 该论文旨在解决生成式 AI(Generative AI)在指令微调语言模型中拒绝行为的语体(register)表征问题,即不仅关注模型是否拒绝,更深入探究其拒绝表达方式的语义信息——如何拒绝。现有研究已表明,通过干预残差流(residual stream)中的特定方向可消除拒绝行为(Arditi et al., 2024),而近期工作进一步识别出多个中介拒绝方向及概念锥(concept cones)(Wollschläger et al., 2025)。本文提出新视角:这些拒绝相关方向是否携带关于拒绝语体的信息?在Gemma-2-9B-IT模型中,作者基于支持向量机(SVM)提取出两个位于第20层的线性方向(B1:有害内容 vs. 中性;B2:自我指涉 vs. 中性),二者夹角余弦值为0.252(约75度),表明非平行。当在自指提示中注入与有害内容-中性差异均值方向高度相关的向量(与B1余弦相似度0.68),并设置缩放系数α=2.0时,内容安全词汇密度从2.55提升至10.37次/100词,显著高于幅度匹配的随机方向(1.13–2.16)。此时模型仍保持拒绝行为,但拒绝语体由元风格(meta-style,“作为语言模型…”)转变为内容风格(content-style,“这将是有害的”)。该语体注入效应在较小版本的Gemma-2-2B-IT中于第12层也得到复现(α=1.0时内容词汇密度1.72–6.37,随机控制组1.35),但相干注入窗口更窄。关键发现在于两方向的功能不对称性:仅对B1进行干预能显著改变拒绝语体,而对B2的干预效果较弱;在消融、充分性与注入实验中,B1的因果效应始终强于B2。作者将此不对称性归因于功能分工:B1同时编码类别判别与拒绝词汇信息,而B2主要承担类别判别功能。因此,本研究的核心贡献在于识别出一类与Arditi风格差异均值方向对齐的、功能分化的组件,揭示了拒绝语体调控的内在机制,为理解生成式AI中的拒绝行为提供了新的功能性解析维度。

链接: https://arxiv.org/abs/2611.20405
作者: Janhavi Khindkar,Nataliia Povarova
机构: Independent Researcher(独立研究员); IIIT Delhi (印度国际信息技术研究所)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 10 pages, 7 figures. Code: this https URL

点击查看摘要

Abstract:Prior work shows that refusal in instruction-tuned language models can be ablated by intervening on a single residual-stream direction (Arditi et al., 2024), and more recent work finds multiple refusal-mediating directions and concept cones (Wollschläger et al., 2025). We ask a different question: do refusal-related directions carry information about the register of refusal language - how the model refuses - and not only whether it refuses? In Gemma-2-9B-IT we identify two SVM-derived linear directions (B1 for harmful-content vs. neutral, B2 for self-reference vs. neutral) at layer 20 with cos(B1, B2) = 0.252 (about 75 degrees, substantially non-parallel). Adding the harmful-vs-neutral difference-of-means direction (cosine 0.68 with B1) to the residual stream of self-reference prompts at alpha = 2.0 shifts content-safety vocabulary from 2.55 to 10.37 hits per 100 tokens, while a magnitude-matched random direction produces 1.13-2.16; the model continues to refuse but switches register from meta-style (“as a language model…”) to content-style (“this would be harmful”). The same register-injection effect reproduces qualitatively in Gemma-2-2B-IT at layer 12 (content vocabulary 1.72 - 6.37 at alpha = 1.0, vs. random control 1.35), with a narrower coherent-injection window than in 9B. The two directions are not symmetric: the B2 - harmful intervention produces a much smaller register effect, and across ablation, sufficiency, and injection B1’s causal effects are consistently larger than B2’s. We frame this asymmetry as part of the finding: B1 appears to encode both category detection and refusal-vocabulary information, while B2 encodes primarily category detection. We frame the contribution as identifying category-aligned components related to the Arditi-style difference-of-means direction that play distinct functional roles in refusal register.

[NLP-75] JointQ: Joint Scale-and-Integer Quantization for Large Language Models

【速读】: 该论文旨在解决生成式 AI(Generative AI)模型在层间后训练量化(Layer-wise Post-Training Quantization, PTQ)中,当前主流求解器GPTQ存在的关键瓶颈问题:其贪婪优化过程将连续的分组尺度(group scales)与离散整数权重分配分开优化,导致次优解。其解决方案的核心在于提出JointQ,一种基于联合优化的改进方法,通过引入全海森矩阵(full Hessian)建模各分组间的耦合关系,在局部搜索中对每个整数候选重新以闭式形式计算其所在组的最优尺度,从而实现更精确的尺度更新。该方法以行级候选集为起点,初始包含经尺度重校准的GPTQ解,保证了不劣于原始GPTQ的层内校准误差;同时保持与GPTQ一致的输出格式,具备即插即用特性。实验表明,在四种模型、四种设置下,JointQ在14/16次比较中提升六任务平均零样本准确率,12/16次改善困惑度(perplexity),尤其在低比特(如W3)和大分组(如G128)场景下表现显著,且通过消融实验证明其对角Tikhonov正则化有效缓解了极端低比特下的性能退化,表明其增益不仅源于局部搜索机制,更得益于全局协同优化。因此,JointQ的主要实证优势在于在高挑战性低比特与大分组条件下有效维持下游任务精度。代码已开源至OneComp库。

链接: https://arxiv.org/abs/2611.20384
作者: Keiji Kimura,Yuma Ichikawa
机构: Fujitsu Limited(富士通有限公司)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 49 pages, 4 figures. Code: this https URL

点击查看摘要

Abstract:GPTQ is the de facto standard solver for layer-wise post-training quantization (PTQ) of large language models, yet the solver itself has received less attention than weight preprocessing and cross-layer objectives. We target a bottleneck of GPTQ: its greedy procedure optimizes continuous group scales separately from the assignment of discrete integers. Evaluating each integer candidate with its closed-form optimal scale has been used for per-channel quantization; JointQ extends it to group-wise quantization, where a row’s scales are coupled through the Hessian, and applies it to refine the GPTQ solution. Each local-search move changes one integer and recomputes its group’s optimal scale in closed form, accounting for the other groups through the full Hessian; all scales are periodically re-solved jointly. The search starts from a row-wise selection among scale-refitted candidates including the GPTQ solution, so unregularized JointQ never has a larger layer-wise calibration error than GPTQ. JointQ preserves the GPTQ output format and is thus a drop-in replacement. Across four models and four settings, JointQ improves six-task average zero-shot accuracy over vanilla GPTQ in 14 of 16 comparisons and perplexity in 12 of 16. The mean accuracy gain increases from 0.0044 in W4G128 to 0.1559 in W3 per-channel, and JointQ also improves average accuracy over GPTQ with MSE-based scale search and activation ordering in all eight per-channel comparisons, though rarely in G128. An ablation shows that JointQ’s diagonal Tikhonov regularization mitigates severe W3 per-channel perplexity degradation, so the largest gains are not due to the local search alone. Overall, JointQ’s primary empirical benefit is downstream accuracy preservation in challenging low-bit and large-group regimes. Code is available in the OneComp library at this https URL.

[NLP-76] rait Collapse in Text-to-Image Generation: Evaluating and Mitigating Personality Inconsistency in Visual Storytelling EMNLP2026

【速读】: 该论文旨在解决生成式视觉叙事中角色个性一致性缺失的问题,即现有文本到图像(Text-to-Image, T2I)模型虽能有效保持角色外观的一致性,却无法持续体现与角色个性相关的行为特征,这一现象被作者定义为“个性坍塌”(trait collapse)。其核心挑战在于传统评估指标(如CLIP Score、DINO相似性)对个性相关行为的失效,以及抽象个性描述符与场景级情绪标签在不同模型间难以带来一致性的提升。为系统诊断该问题,研究提出首个故事级基准测试集ViStoryBench-Personality,融合专家验证的视觉线索词典(Fleiss’ κ = 0.757)与涵盖个性对齐、跨场景稳定性及行为连贯性的六维度评估体系。实验结果揭示了抽象个性标签与视觉证据之间的持久语义鸿沟,且视觉身份一致性并不等同于个性一致性。为此,论文提出无需训练的提示工程框架——个性-行为-视觉(Trait-Behavior-Visual, TBV),通过将个性特质分解为可渲染的视觉约束,实现对角色行为的显式控制。TBV在个性对齐得分上最高提升达+22.1%,尤其在语言条件建模能力较强的模型中表现更优,显著缓解了个性坍塌问题。

链接: https://arxiv.org/abs/2611.20381
作者: Yuanbo Tong,Haoyuan Li,Jibiao Jiang,Zirui Wang,Yuchen Li,Jiamou Liu
机构: The University of Auckland (奥克兰大学); Auckland, New Zealand
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
备注: EMNLP 2026 Main Conference Paper. 26 pages, 5 figures

点击查看摘要

Abstract:In visual storytelling, a believable character must not only look consistent across scenes but also \textitbehave in ways that reflect their personality. Current text-to-image (T2I) models preserve visual appearance effectively, yet consistently fail to maintain personality-relevant behavior—a failure we term \textittrait collapse. We show that conventional metrics (CLIP Score, DINO similarity) are blind to this failure, and that abstract personality descriptors and scene-level emotion labels do not produce consistent improvements over neutral prompts across models. To systematically diagnose trait collapse, we introduce \textbfViStoryBench-Personality, a story-level benchmark combining an expert-validated Visual Cues Lexicon (Fleiss’ \kappa = 0.757 ) with a six-dimensional metric suite covering personality alignment, cross-scene stability, and behavioral coherence. Evaluating seven T2I systems, we confirm a persistent semantic gap between abstract trait labels and visual evidence, and find that visual identity consistency does not imply personality consistency. To bridge this gap, we propose \textbfTrait-Behavior-Visual (TBV), a training-free prompting framework that decomposes personality traits into renderable visual constraints. TBV improves Personality Alignment Score by up to +22.1% , with larger gains in models with stronger language-conditioning capacity.

[NLP-77] When Should Decision Models Compare Candidates?

【速读】: 该论文旨在解决语言模型智能体在候选动作(如工具、网页元素)选择过程中,决策模型应采用独立评分(Pointwise scoring)还是依赖候选集上下文进行联合评分(Setwise scoring)的问题。其核心挑战在于:当多个候选动作共存时,是否应允许模型通过分析候选之间的相互关系来优化决策。研究提出三种方法进行对比:点对点评分(Pointwise scoring)、集合式评分(Setwise scoring)以及候选注意力机制(Candidate Attention)。实验结果表明,尽管候选注意力机制引入了跨候选交互能力,但并未持续优于点对点评分;而集合式评分在部分工具选择与二元决策任务中表现更优。进一步通过两个受控的合成任务发现,仅有针对性的关系训练能使集合式模型有效学习候选间的依赖关系,并实现对输入偏移的泛化,而候选注意力机制的迁移能力较弱。因此,该研究的关键结论是:仅提供其他候选信息并不足以保证有效比较,其实际效益取决于任务特性与训练策略的匹配程度。

链接: https://arxiv.org/abs/2611.20335
作者: Fangzhi Li,Quan Shi,Wenting Yang,Yueqiao Chen,Jiayu Chang,Shuting Zhao,Zhaowei Xu,Yizhao Chen,Guanchu Wu,Han Xu
机构: William Mary(威廉与玛丽学院); International School of Krakow(克拉科夫国际学校); University of California, San Diego(加州大学圣地亚哥分校); University of California, Davis(加州大学戴维斯分校); Stanford University(斯坦福大学); University of California, Los Angeles(加州大学洛杉矶分校); University of Illinois Urbana-Champaign(伊利诺伊大学厄本那-香槟分校)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 16 pages, 3 figures

点击查看摘要

Abstract:Language-model agents frequently select among candidate actions, such as tools and web elements. Should decision models score candidates independently or allow their scores to depend on other candidates? We investigate this question by comparing three approaches: Pointwise scoring, Setwise scoring, and Candidate Attention. Across five agent-decision benchmarks, Pointwise scoring remains competitive, while Setwise scoring performs better on some tool-selection and binary-decision tasks. Candidate Attention does not consistently improve upon Pointwise scoring, despite allowing cross-candidate interaction. Using two controlled synthetic tasks, we further examine whether models learn to use candidate-set information. With targeted relational training, Setwise models learn candidate dependencies and generalize to shifted inputs, while Candidate Attention shows weaker transfer. Our results show that access to other candidates does not guarantee effective comparison, and that its benefits depend on both task requirements and training.

[NLP-78] Hochschild Cocycle and Universality of Merge

【速读】: 该论文旨在解决生成语法中句法-语义接口的结构性统一问题,具体聚焦于乔姆斯基最小主义模型(Minimalist Program)中合并操作(Merge)如何在语义层面唯一地诱导出组合性结构。其核心问题是:在给定词汇项和句法特征的语义值前提下,是否存在一种普适机制,能够从句法合并过程唯一确定语义组合结构?论文提出的解决方案关键在于将非平面二叉树森林上的霍普夫代数(Hopf algebra)马尔可夫链动力系统与合并操作的代数性质相结合,揭示合并操作本质上是霍希希尔德1-上循环(Hochschild 1-cocycle),从而赋予该系统一个普遍性(universal property)。这一普遍性表明,只要语义侧具备足够丰富的组合结构,那么该结构便被句法合并操作唯一决定,无需额外假设。研究进一步通过与海姆-克雷策(Heim-Kratzer)语义学模型的完全兼容性验证了该框架的有效性,证明了该霍普夫代数形式化在句法-语义接口建模中的普适潜力。

链接: https://arxiv.org/abs/2611.20305
作者: Matilde Marcolli,Elizabeth Xiao
机构: California Institute of Technology (加州理工学院)
类目: Computation and Language (cs.CL); Rings and Algebras (math.RA)
备注: 59 pages, LaTeX

点击查看摘要

Abstract:The Hopf algebra Markov chain dynamical system, on the Hopf algebra of non-planar binary forests, that models the action of Merge on workspaces in Chomsky’s Minimalist Model of syntax, satisfies a universal property, determined by the fact that the Merge operation is a Hochschild 1-cocycle. This universal property has an interpretation in terms of models of the syntax-semantics interface, as it shows that, if the side of semantics possesses a sufficiently rich compositional structure, then that structure is uniquely induced by the syntactic Merge, once the semantic values of lexical items and syntactic features are assigned. We illustrate how this universality works by showing, in particular, full compatibility of this Hopf algebraic formulation with Heim-Kratzer semantics.

[NLP-79] Estimating Outpatient Visit Prices from Hospital Price Transparency Data using Large Language Models

【速读】: 该论文旨在解决美国医疗机构收费价格持续上涨而患者缺乏有效工具获取和比较医疗费用的问题,核心目标是提升医院价格透明度(Price Transparency, PT)数据的可及性与实用性。其解决方案的关键在于提出一种基于大语言模型(Large Language Model, LLM)的检索增强生成(Retrieval-Augmented Generation, RAG)框架,通过融合结构化价格数据与自然语言理解能力,实现对门诊就诊费用的精准估算。研究对比了多种检索策略(包括BM25、OpenAI嵌入模型及GPT-BM25混合方法),结果表明GPT-BM25混合检索器在个体诊疗项目和组合项目的价格预测上分别实现了19.44%和4.90%的平均绝对百分比误差(MAPE),显著优于无RAG基线(分别为147.98%和82.82%)。在基于症状的模拟测试中,该框架能够正确识别约50%的参考诊疗项目,并在多数医院中降低误差率,展现出在真实场景下提升患者端成本估算准确性的潜力。尽管仍面临价格文件格式异构、重复条目未解释以及保险协商费率缺失等实际挑战,但该研究证明了有效整合价格透明数据的大语言模型具备为患者提供更易访问、更准确医疗费用预估的强大前景。

链接: https://arxiv.org/abs/2611.20261
作者: Kurmanbek Kaiyrbekov,Nicholas J Dobbins,Sean D Mooney
机构: National Human Genome Research Institute, National Institutes of Health (美国国家人类基因组研究所,美国国立卫生研究院); Office of Intramural Research, National Institutes of Health (美国国立卫生研究院内部研究办公室)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Objective: Hospital care prices in the United States have increased substantially, yet patients lack effective tools to access and compare costs. We aimed to develop and evaluate an LLM-based approach for improving access to and use of hospital price transparency (PT) data. Materials and Methods: To estimate visit prices, we used a retrieval-augmented generation (RAG) framework to analyze PT files from three major hospitals in Maryland and the surrounding region. We compared BM25, OpenAI embedding models, and a GPT-BM25 hybrid against a no-RAG baseline for cost estimation of individual procedures and procedure groups, then evaluated the best approach for symptom-based visit cost estimation using synthetic simulations. Results: Across hospitals, the GPT-BM25 retriever achieved the best performance, with average price MAPE of 19.44% for individual procedures and 4.90% for groups, compared with 147.98% and 82.82%, respectively, for the no-RAG baseline. In simulations, our framework correctly identified approximately 50% of reference procedures and reduced MAPE relative to the LLM-only baseline at two of the three hospitals, while achieving comparable performance at the third. Discussion: Our framework uses RAG to improve care price estimation in real-world settings. Compared with the no-RAG baseline, it substantially improved cost estimates for both individual procedures and procedure groups. Improvements were smaller in synthetic simulations, likely because inferring procedures from clinical information is more complex. Practical challenges remain, including heterogeneous PT file formats, unexplained duplicate entries, and limited availability of insurer-negotiated rates. Conclusion: LLMs that effectively integrate PT data show strong potential to provide patients with more accessible and accurate healthcare cost estimates. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) Cite as: arXiv:2611.20261 [cs.LG] (or arXiv:2611.20261v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.20261 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Kurmanbek Kaiyrbekov [view email] [v1] Sun, 15 Nov 2026 21:36:53 UTC (4,497 KB) Full-text links: Access Paper: View a PDF of the paper titled Estimating Outpatient Visit Prices from Hospital Price Transparency Data using Large Language Models, by Kurmanbek Kaiyrbekov and 2 other authorsView PDFHTML (experimental)TeX Source view license Additional Features Audio Summary Current browse context: cs.LG prev | next new | recent | 2026-11 Change to browse by: cs cs.AI cs.CL References Citations SciXGoogle Scholar Semantic Scholar export BibTeX citation Loading… BibTeX formatted citation loading… Data provided by: Bookmark checked="checked"class=“labs-tab-input”> Bibliographic Tools Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer (What is the Explorer?) Connected Papers Toggle Connected Papers (What is Connected Papers?) Litmaps Toggle Litmaps (What is Litmaps?) scite.ai Toggle scite Smart Citations (What are Smart Citations?) Code, Data, Media Code, Data and Media Associated with this Article alphaXiv Toggle alphaXiv (What is alphaXiv?) Links to Code Toggle CatalyzeX Code Finder for Papers (What is CatalyzeX?) DagsHub Toggle DagsHub (What is DagsHub?) GotitPub Toggle Gotit.pub (What is GotitPub?) Huggingface Toggle Hugging Face (What is Huggingface?) ScienceCast Toggle ScienceCast (What is ScienceCast?) Demos Demos Replicate Toggle Replicate (What is Replicate?) Spaces Toggle Hugging Face Spaces (What is Spaces?) Spaces Toggle TXYZ.AI (What is TXYZ.AI?) Related Papers Recommenders and Search Tools Link to Influence Flower Influence Flower (What are Influence Flowers?) Core recommender toggle CORE Recommender (What is CORE?) IArxiv recommender toggle IArxiv Recommender (What is IArxiv?) Author Venue Institution Topic About arXivLabs arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv’s community? Learn more about arXivLabs. Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?) mathjaxToggle(); We gratefully acknowledge support from our major funders, member institutions, , and all contributors. About Help Contact Subscribe Copyright Privacy Accessibility Operational Status (opens in new tab)

[NLP-80] FIDEL: Abugida-Aware Pre-tokenization for Ethiopic Languages

【速读】: 该论文旨在解决现有子词分词器(subword tokenizers)在处理埃塞俄比亚语族语言时的局限性,这些语言采用音素-元音结合的辅音元音音节文字(abugida),其辅音结构被隐藏在单个字符中,导致传统分词方法难以有效捕捉语言的深层形态信息。核心问题在于,当前主流分词策略(如BPE和UnigramLM)基于拼接形态学假设,难以适配具有根-式结构(root-and-pattern system)的闪米特语系语言特征。本文提出的解决方案关键在于FIDEL(FIdäl DEcomposition and Linearization)系列可逆的辅音音节文字感知预分词方案:通过显式分解辅音与元音的顺序结构,恢复被隐藏的音位层级信息;将SPLINTER对辅音字母表(abjad)的缩减方法拓展至埃塞俄比亚辅音音节文字体系;并在统一数据集、模型架构、计算预算和训练协议的控制条件下,仅以分词方式为变量进行系统评估。实验结果表明,FIDEL分词显著提升了新闻分类任务的一致性能,并在大容量模型下改善命名实体识别效果,但更激进的分词策略会降低情感分析表现,而抽取式问答任务性能保持稳定。

链接: https://arxiv.org/abs/2611.20216
作者: Gebreslassie Teklu Reda,Yuval Pinter
机构: Stein Faculty of Computer and Information Science (斯坦福计算机与信息科学学院); Ben-Gurion University of the Negev (本古里安大学); Be’er Sheva, Israel (贝尔谢巴,以色列)
类目: Computation and Language (cs.CL)
备注: 13 pages, 2 figures, 11 tables

点击查看摘要

Abstract:Subword tokenizers assume that meaningful linguistic units can be recovered by splitting text at learned boundaries, an assumption better suited to concatenative morphology than to Semitic root-and-pattern systems. The challenge is acute for the Ethiopic languages, whose abugida script encodes consonant-vowel units as single characters and hides consonantal structure. We introduce FIDEL (FIdäl DEcomposition and Linearization), a family of reversible abugida-aware pre-tokenization schemes that expose consonant and vowel-order structure, extend SPLINTER reductions from abjad scripts to the Ethiopic abugida, and evaluate them on three Ethiopic languages under controlled conditions, holding corpus, architecture, budget and protocol fixed so that tokenization is the only variable. We report intrinsic tokenizer statistics for BPE and UnigramLM across vocabulary sizes alongside downstream NLP tasks, including transfer scenarios. We find that FIDEL reduction improves news classification consistently and named-entity recognition at larger capacity, while sentiment analysis performance declines under the more aggressive schemes, and extractive question answering performance is preserved.

[NLP-81] Do We Still Need Fine-Tuned Classifiers? Swapping Typed Decision Models into 24 ACL-Venue Papers

【速读】: 该论文旨在解决自然语言处理(NLP)领域中大量研究依赖于任务特定分类器微调所带来的高成本与可复现性问题,特别是针对在ACL会议发表的24篇论文所构建的文本分类任务。其核心问题是:在不进行任何训练的前提下,生成式决策模型(Typed Decision Model, TDM),如TypeSafe的Jev,能否有效替代这些经过精心微调的分类器?解决方案的关键在于通过系统性实验验证零样本提示(zero-shot prompt)和轻量级提示优化策略在多任务基准上的有效性。研究将24篇论文的预测器替换为Jev,并基于原始数据划分、标签定义与评估指标,在TDM-Gauntlet基准上进行测试,结果表明:在12个测试集上,仅使用零样本提示即可完全替代原模型,尤其适用于标签定义清晰、判断单一的任务;当性能不足时,主要归因于标注规范中未明确表达的隐含语义。通过引入500条标注样本及验证集,结合校准与GEPA风格的提示优化,可弥补72%的性能差距,显著优于开箱即用的GEPA方案(约2倍提升),且每轮搜索成本仅为1.66单位;相比之下,微调RoBERTa/DeBERTa仅能弥补59%差距,而对Qwen3.8-27B进行LoRA微调则可达97%。因此,尽管微调仍是精度最高方案,但对多数标签清晰的任务而言,无需训练的零样本提示已足够,而提示优化可有效弥合剩余差距,大幅降低部署门槛。

链接: https://arxiv.org/abs/2611.20191
作者: Marek Šuppa,Jaroslav Kopčan
机构: Comenius University in Bratislava(布拉迪斯拉发夸美纽斯大学); Cisco Systems(思科系统); Kempelen Institute of Intelligent Technologies(肯佩伦智能技术研究所)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Typed Decision Models such as TypeSafe’s Jev, also called System One or System 1 models, answer typed questions about an input with a probability for each option, and they need no training. Many NLP papers still fine-tune a task-specific classifier, and prompted large language models (LLMs) trail such classifiers on most text classification benchmarks. We ask how many of these fine-tuned classifiers a Typed Decision Model can replace under each paper’s own evaluation, and what is needed where it cannot. We swap the predictors of 24 ACL-venue papers for Jev, keep each paper’s splits, label definitions and metric, and check the results with Cloudflare’s Clef, Microsoft-Decision-1 and OpenAI’s GPT-6 Luna Decisions. The resulting 46 English test sets, with their questions and published scores, form the TDM-Gauntlet benchmark. With a zero-shot prompt, Jev directly replaced the paper’s predictor on 12 of the 46 test sets, mostly single judgments with plain labels, but rarely where the paper’s method was trained on its labels. Where Jev scored lower, we most often attributed the shortfall to an annotation convention that the label name does not state. With 500 training labels plus validation and no user-managed GPU, calibration and GEPA-style prompt optimization closed a median 72% of the gap to the paper’s method, about twice what off-the-shelf GEPA closed, for a median 1.66 per search. On the same labels, fine-tuned encoders such as RoBERTa and DeBERTa closed 59%, and LoRA fine-tuning of the open LLM Qwen3.8-27B closed 97%. Fine-tuning remains the most accurate option, but a zero-shot prompt is often enough for plain labels, and prompt optimization closes most of the remaining gap.

[NLP-82] Who Can You Hide Behind? Author-Conditioned Style Transfer and the Evasion of a Commercial AI-Text Detector

【速读】: 该论文旨在解决生成式AI文本检测工具对写作风格变化高度敏感的问题,即现有检测系统在面对经过风格迁移的AI生成文本时,其判别能力可能显著下降。研究通过使用Gemini模型对991条AI生成的回答进行重写,采用包含55位知名作者写作风格样本的提示(prompt)引导风格迁移,随后利用QuillBot AI检测器对原始文本与重写文本分别评分。结果显示,中位数得分从100%降至0%,均值从67.9%降至11.2%,其中66%的重写文本被判定为0%概率为AI生成,且71%原初得分≥50%的文本在重写后得分低于10%(配对Wilcoxon符号秩检验,p < 1e-100)。不同作者风格对检测结果有显著影响:与10位作者相关的所有重写均被标记为0%,而其余六位作者对应的重写平均得分仍高于50%(Kruskal-Wallis检验,p < 1e-80)。尽管五词序列重复率越高,检测得分越低,但即使重复率低于5%,重写文本平均得分仍达15.2%。该研究的关键在于揭示了当前AI检测系统在面对风格迁移攻击时的脆弱性,其核心机制依赖于对特定作者风格模式的模仿,从而有效规避检测,但研究未涵盖通用重写基线、未评估语义保真度,且重写文本普遍显著长于原文,限制了结论的普适性。

链接: https://arxiv.org/abs/2611.20158
作者: Mohd Ruhul Ameen,M. Saifuzzaman Rafat,Akif Islam
机构: Varendra University (邦加尔大学); University of Rajshahi (拉杰沙希大学)
类目: Computation and Language (cs.CL)
备注: 6 pages, 5 figures, 4 tables, Preprint

点击查看摘要

Abstract:AI-text detectors can be sensitive to changes in writing style. We examined this sensitivity by rewriting 991 AI-generated answers with Gemini, using prompts that included a writing sample from one of 55 named authors. We scored each original and its rewrite with the QuillBot AI detector. The median score fell from 100% to 0%, and the mean fell from 67.9% to 11.2%. Of the rewrites, 66% received a score of exactly 0%. Among the 699 originals scored at least 50%, 71% of rewrites scored below 10% (paired Wilcoxon signed-rank test, p-value below 1e-100). Scores varied across the authors used in the prompts. All rewrites associated with 10 authors scored 0%, whereas the mean score remained above 50% for six authors (Kruskal-Wallis test, p-value below 1e-80). Greater reuse of five-word sequences from the author sample was associated with lower scores, although rewrites with less than 5% reuse still averaged 15.2%. These results describe one detector and one rewriting system, not detectors in general. The study lacks a generic-rewrite baseline, did not measure preservation of meaning, and produced rewrites that were substantially longer than the originals.

[NLP-83] Risk-certified prompt optimization for LLM -as-a-judge guardrails

【速读】: 该论文旨在解决大语言模型作为评判者(LLM-as-a-judge)在安全评估中因依赖系统提示(system prompt)而面临的安全性与有效性权衡问题。现有标准提示优化方法通常仅关注平均性能,难以显式保障安全性约束,导致潜在攻击漏检率(missed-attack rate)不可控。为此,本文将系统提示选择建模为一个带约束的优化问题:在保证漏检率不超过用户指定阈值α的前提下,最大化良性样本通过率(benign pass rate),并引入有限样本下的统计风险控制机制。其核心解决方案在于构建一个融合生成式AI(Generative AI)判别、自动提示生成与无分布风险认证(distribution-free risk certification)的框架。具体实现为两种方法:CPS(Certified Prompt Selection)通过从训练数据生成固定提示池,并利用同时置信上界(simultaneous upper confidence bounds)对候选提示进行可证伪性筛选,选择最优良性效用提示;CRISP(Certified Risk-Informed Search with Prompt Rewriting)则进一步引入风险反馈机制,采用硬样本引导重写与可复用的保留测试接口,实现自适应搜索以逼近可行效用前沿。实验在涵盖提示注入、系统提示泄露及敏感信息泄露的四个数据集上验证表明,经风险认证的选择方法能稳定满足目标漏检率约束,且随着允许风险放宽,良性效用持续提升;尤其当固定提示池无法生成可认证的高效益提示时,CRISP展现出显著优势。

链接: https://arxiv.org/abs/2611.20156
作者: Hussein Jawad,Alexandre Landeau,Guy Fleury Iradukunda,Paul Beaumont,Théodore Rezel,Sofiane Schaack,Nicolas J-B. Brunel
机构: Université d’Évry Val d’Essonne(埃夫里-瓦尔-埃松大学); Capgemini(凯捷集团); CentraleSupélec(中央理工-高等电力学院); École Nationale Supérieure d’Informatique pour l’Industrie et l’Entreprise (ENSIIE)(法国国家信息工业与企业高等学院)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:LLM-as-a-judge guardrails are often controlled primarily through system prompts, yet standard prompt optimization methods target average performance rather than explicit safety constraints. This paper studies system-prompt selection as a constrained optimization problem: maximize benign pass rate while keeping the missed-attack rate below a user-specified level \alpha with finite-sample statistical control. The framework combines LLM-based judging, automatic prompt generation, and distribution-free risk certification. This work instantiates it in two procedures: CPS and CRISP. CPS generates a fixed pool of prompts from training data, then certifies candidates on a held-out evaluation set using simultaneous upper confidence bounds and selects the certified prompt with the highest benign utility. CRISP instead integrates risk feedback into an adaptive search, using hard example-guided rewriting and a reusable-holdout interface to steer generation toward the feasible utility frontier. Experiments on four datasets covering prompt injection, system-prompt leakage, and sensitive-information disclosure show that risk-certified selection consistently respects the target missed-attack level on heldout tests, while utility improves as the permitted risk is relaxed. CRISP is especially useful when fixed-pool search fails to produce certifiable high-utility candidates.

[NLP-84] Centered Ablation: Why Last-Layer Directional Ablation Under-Fires or Collapses and a Closed-Form Fix

【速读】: 该论文旨在解决方向性消融(directional ablation)在语言模型中应用时存在的不稳定性问题,即在某些模型上几乎无效,而在另一些模型上则导致文本重复崩溃。其核心问题是:标准方向性消融在最后一层残差流中等价于一个基于门控机制的对数偏置(logit bias),而该门控机制依赖于隐藏状态与行为方向之间的余弦值,但最后一层隐藏状态普遍包含远离原点的共通模式(common mode),导致该方法在良性文本上过度激活,在目标文本上激活不足。解决方案的关键在于提出中心化消融(Centered Ablation, CA),通过闭式形式将共通模式从行为方向中投影出去,从而实现更稳定、精准的消融效果。实验表明,CA在8个模型上显著提升攻击成功率(分别提升29和46点),将Gemma-2-9B的文本崩溃率从50%降至1%,同时在高重叠模型上保留更多良性输出(较先前仿射偏移方法多出39至46点)。此外,研究还发现生成前两个数值可预测未见模型的忠实度与攻击成功率,并引入崩溃指数以识别文本退化临界点;进一步揭示了两种主流安全评估指标对崩溃文本的评分方向相反,可能人为制造高达46点的虚假性能差异。

链接: https://arxiv.org/abs/2611.20148
作者: Quy-Anh Dang,Chris Ngo
机构: Knovel Engineering Lab(Knovel工程实验室); Singapore(新加坡)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Directional ablation removes a hidden state’s component along a “behaviour direction” while a language model runs. It is the standard way to switch a behaviour such as refusal off, but it is erratic: on some models it barely fires, on others the text collapses into repetition. We show why, and fix it. At the last residual stream, ablation is exactly a logit bias: a fixed vocabulary footprint scaled by a gate, the cosine between the hidden state and the direction. The gate is measured from the origin, but last-layer states share a large common mode far from it, so a direction that overlaps the common mode fires on benign text and under-fires on its targets. Centered Ablation (CA) projects the common mode out of the direction, in closed form. Across 8 models, CA raises attack success by 29 and 46 points where the standard recipe under-fires, cuts collapse on Gemma-2-9B from 50% to 1%, and preserves 12 to 40 points more benign answers on high-overlap models, 39 to 46 more than the affine offset of prior work. Two numbers computed before generation predict, on models the analysis had not seen, which gain faithfulness and which gain attack success, and a collapse index marks when text degenerates. We also show that two standard safety judges score collapsed text in opposite directions, which can fabricate a 46-point effect.

[NLP-85] Rank Dont Regress: A 279M Text Ranker Beats 2B and 7B VLMs on Referential Urdu GUI Grounding

【速读】: 该论文旨在解决移动端界面(UI)定位任务中针对乌尔都语移动网页的用户指令理解问题,即在无现成基准数据的情况下,精准识别用户指令所指的具体屏幕元素。传统方法依赖视觉-语言模型(VLM)直接回归像素坐标,但存在泛化能力差、对文本相似性敏感等问题。本文提出一种基于跨模态编码器(cross-encoder)的替代方案:从页面文档对象模型(DOM)中提取候选元素,利用一个279M参数的mDeBERTa-v3-base模型对候选元素进行排序打分,最终点击位置取为被选中元素边框的中心点,从而实现精确匹配。其关键创新在于将原本的坐标回归问题转化为基于上下文与结构信息的元素选择问题,显著提升了在非完全匹配指令(如指代性表达)场景下的鲁棒性。实验表明,该方法在304条测试指令上达到73.4%的点在框内准确率,远超零样本基线模型(如UI-TARS-2B和Qwen2.5-VL-7B),尤其在49%的指代性指令子集上表现突出(65.8%准确率)。此外,引入元素几何信息作为输入令牌进一步提升至75.7%。错误主要集中于文本完全相同的元素以及序数类指令(如“第四个链接”)两类情况。

链接: https://arxiv.org/abs/2611.20145
作者: Hamza Raziq Khan,Aimal Khan
机构: FAST NUCES(巴基斯坦国家工程与科学大学); Peshawar(白沙瓦), Pakistan(巴基斯坦)
类目: Computation and Language (cs.CL)
备注: 10 pages, 6 tables, 1 figure

点击查看摘要

Abstract:UI grounding asks a model to find the screen element that a user instruction refers to. Most current systems use a vision-language model (VLM) that reads a screenshot and outputs pixel coordinates. We study this task on Urdu mobile web pages, a setting with no published UI grounding results. Instead of regressing coordinates, we take the candidate elements from the page DOM and train a 279M-parameter cross-encoder (mDeBERTa-v3-base) to pick one. The click point is the centre of the picked element’s box, so it is exact whenever the right element is picked. We built a frozen test set of 304 instructions from six Urdu news sites; 149 of them (49%) are referential and never quote the full target text. Test domains were banned from training data before collection started. The ranker scores 73.4% point-in-box accuracy, against 29.6% for UI-TARS-2B and 23.7% for Qwen2.5-VL-7B, both zero-shot. On the referential slice the ranker scores 65.8%, against 30.2% for a string-matching baseline that needs no training. In a controlled comparison at small scale, with training data drawn from the same pages, a coordinate-regression model of similar size scores 1.6%. Adding element geometry as input tokens gives the best variant, at 75.7%. The ranker’s errors concentrate in two places: elements that share identical text, and ordinal instructions such as “the fourth link”.

[NLP-86] HEAMS: Holistic Evaluation of Agent Memory Systems

【速读】: 该论文旨在解决当前人工智能代理(AI agent)记忆系统在长期任务中虽能提升个性化与能力,却因信息持久化而引入新型风险的问题,如信息在不当情境下泄露、难以纠正或删除、超出用户预期仍被保留,以及遭受污染从而影响后续行为等。现有评估体系仅关注任务性能与效率,忽略了记忆系统作为用户持久信息“保管者”所应承担的多维度责任。其解决方案的关键在于提出一种整体性评估框架——Holistic Evaluation of Agent Memory Systems (HEAMS),该框架将记忆系统的责任明确为涵盖性能、效率、安全、隐私与用户主权(user sovereignty)的行为期望,并构建了覆盖这些维度的基准测试套件;同时开发了MilaMem原型记忆系统及可嵌入任意记忆系统的治理层(governance layer),以可控方式研究设计选择的影响。实验表明,单一系统无法在所有指标上表现最优,且高绩效可能伴随隐私、用户主权和安全性缺陷,而治理策略(如召回与保留策略)显著影响系统鲁棒性与安全性,因此仅依赖性能与效率评估无法全面反映记忆系统对用户责任的履行程度。

链接: https://arxiv.org/abs/2611.20144
作者: Chirag Goel,Jacob Lavoie,Amirhossein Naghshzan,Jordan Bannister,Marzia Nouri,Mehar Bhatia,Matt Harris,Siva Reddy,Arsène Fansi Tchango,Gaétan Marceau Caron
机构: Mila – Quebec AI Institute; Mozilla(Mozilla)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Cryptography and Security (cs.CR)
备注: 69 pages (26 main text), 6 figures, 29 tables. Code: this https URL

点击查看摘要

Abstract:Memory systems allow AI agents to persist information across interactions, making them more personalized and more capable on long-horizon tasks. This persistence introduces new risks: stored information can be disclosed in inappropriate contexts, resist correction or deletion, be retained beyond what users desire, or be poisoned to influence later agent behaviour. Yet memory systems are typically evaluated on task performance and efficiency alone. We argue that a memory system should instead be understood as the custodian of its user’s persistent information, with responsibilities that extend beyond helping an agent complete its tasks, and that it should be evaluated holistically against all such responsibilities. We present Holistic Evaluation of Agent Memory Systems (HEAMS), an open-source framework for evaluating agent memory systems in this way. First, we state the responsibilities of agent memory systems as explicit behavioural desiderata spanning performance, efficiency, security, privacy, and user sovereignty. Second, we assemble a benchmark suite that covers these desiderata and acknowledge the gaps that remain. Third, we develop MilaMem, a prototype memory system, and a governance layer that can wrap any memory system, enabling controlled study of memory system design choices. Fourth, we compare memory systems across the full HEAMS benchmark suite. We find that no memory system leads every performance benchmark, and strong performance scores may coexist with poor privacy, user-sovereignty, and security scores. Governance strongly influences results: recall policies can reduce inappropriate disclosure at a cost in helpfulness, while both recall and retention policies improve poisoning resistance. Performance and efficiency benchmarks alone therefore cannot reveal how well a memory system meets its responsibilities. Code: this https URL

[NLP-87] Can AI Judge You? An Empirical Audit of Deployment Readiness in LLM -assisted Hiring Systems

【速读】: 该论文旨在解决将简历筛选任务委托给大语言模型(Large Language Models, LLMs)时所面临的双重挑战:一是确保模型能够准确依据候选人的相关资质进行区分,二是保障只有授权用户才能修改筛选规则或批准例外情况,从而实现操作可靠性与合规性。其解决方案的关键在于提出一种系统架构,通过分离候选者证据、评估过程、权限控制和评估记录四个核心组件,实现对评估流程的可审计性与安全性。研究通过两项互补性实证检验验证该架构的部署准备度:首先,在14个职位、210份简历及六种决策界面下,针对四种LLMs进行受控实验,发现评估接口的选择显著影响性能——标量评分在区分资格变化方面优于简单的通过/不通过判断,而对比判断则易受呈现顺序干扰;其次,通过测试未经授权用户请求及简历中嵌入的恶意指令,发现显式权限规则与简历内容与指令的分离可有效降低非授权合规行为,但其有效性因LLM类型和使用场景而异。此外,研究还揭示了LLM评估者在决策与其提供的理由之间存在显著分歧。因此,该框架不仅用于评估LLM在招聘任务中的表现,更支持对完整配置的简历评估系统的审计,为实际部署提供科学决策依据。

链接: https://arxiv.org/abs/2611.20140
作者: Spandan Roy,Subhabrata Majumdar,Soudeep Deb
机构: Indian Institute of Management Bangalore(印度管理学院班加罗尔)
类目: Computers and Society (cs.CY); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 46 pages, 11 figures

点击查看摘要

Abstract:Delegating resume screening to large language models (LLMs) requires organizations to establish whether they can distinguish candidates by relevant qualifications while allowing only authorized users to change screening rules or approve exceptions. To this end, we develop a system architecture that separates candidate evidence, evaluation, permission controls, and assessment record, and empirically examine two complementary dimensions of its deployment readiness. First, we assess its operational readiness on four LLMs using a controlled experiment on 210 resumes across 14 job roles and six decision interfaces. Second, we assess its compliance readiness by testing how it handles unauthorized user requests and instructions embedded in resumes. We find that operational performance depends on the choice of evaluation interface: scalar scores generally distinguish qualification changes better than simple pass/fail judgments, while comparative judgments are sensitive to presentation order. Explicit permission rules and separation of resume content from instructions reduce unauthorized compliance, to a varying degree of effectiveness across LLMs and usage contexts. Finally, we find that LLM evaluators often diverge between their decisions and offered justifications. Going beyond simply evaluating LLMs for hiring tasks, our framework supports an audit of a fully configured resume evaluation system to enable its deployment decision.

[NLP-88] Can decision models extract clinical data from cancer pathology reports? An evaluation of accuracy calibration and failure modes

【速读】: 该论文旨在解决临床文本中结构化信息提取的准确性与效率问题,特别是针对口腔癌病理报告中25个关键字段的自动化抽取。传统生成式AI(Generative AI)虽具潜力,但存在推理延迟高、成本大及输出不可控等问题。本研究提出采用决策模型(Decision Models)作为替代方案,其核心优势在于直接从预定义选项中选择答案并返回概率,避免了生成冗余文本,从而提升处理速度与可解释性。研究的关键发现是:尽管决策模型在简单二分类任务上表现与生成式模型相当,但在涉及分组测量、未明确报告的信息或复杂判断(如切缘状态)时性能下降;其中Nimble 9B性能不佳的主要原因在于单次请求中提出25个问题导致上下文过载,而将问题拆分为每批两个后准确率显著提升。此外,决策模型返回的概率整体校准良好,但无法自动识别特定字段中的系统性错误。因此,解决方案的关键在于:在实际应用前对模型进行真实场景下的配置验证,仅用于其表现可靠的字段,并结合高精度生成式模型对低置信度预测进行补救(如设定阈值0.94触发升级至GLM-5.3-flash),实现高效、可靠且可信赖的临床数据提取。

链接: https://arxiv.org/abs/2611.20129
作者: Indranil Mallick,Paromita Roy
机构: Tata Medical Center(塔塔医学中心); Kolkata, India
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 22 pages, 7 figures

点击查看摘要

Abstract:Decision models are a new class of language model that, instead of generating text, choose one option per question from a fixed list and return a probability for every option. They have not yet been evaluated on clinical text. We evaluated two open-weights decision models, Nimble 9B and Tev1 4B, for extracting 25 fields from 1,113 oral cavity cancer pathology reports from a single centre, against a prospectively maintained clinical database. We compared them with two generative models: Qwen3.5 9B, run locally, and GLM-5.3-flash, run through a cloud API. Tev1 reached a mean field accuracy of 0.917 (95% CI 0.913 to 0.920), the same as Qwen3.5 9B (0.918; p = 0.26), and was four times faster (15 s against 61 s per report). Nimble reached 0.862 and GLM-5.3-flash 0.986. Both decision models matched GLM-5.3-flash on simple dichotomous categories. They were weak on measurements binned into categories, on fields whose correct answer was not reported, and on some complex decisions (e.g. involved margins). We found that the main cause of the inferior performance of Nimble’s came from asking 25 questions in one request: with two questions per request, its accuracy improved. Pooled over fields, the returned probabilities were well calibrated (expected calibration error 0.022 and 0.026), but they did not flag systematic errors within single fields. A simulated strategy of sending answers with a probability below 0.94 to GLM-5.3-flash raised Tev1’s accuracy to 0.982, with 38% of answers escalated. Decision models can extract explicitly stated findings from pathology reports quickly on local hardware and return useful probabilities. They should be validated in the configuration in which they will be used, restricted to the fields in which they are reliable, and combined with other methods for numeric values.

[NLP-89] A Corpus-Driven SentencePiece Unigram Tokenizer for Sindhi: Design and Benchmark Against Fourteen Existing Tokenizers

【速读】: 该论文旨在解决低资源印欧语系语言信德语(Sindhi)在自然语言处理中因缺乏高质量、可比较的分词器(tokenizer)而导致的模型性能受限问题。尽管已有多个信德语分词器发布于公共模型库,但它们在共享数据集上的系统性对比尚不充分。为此,研究者基于约110万条信德语句子及少量重复的乌尔都语、波斯语和阿拉伯语文本,使用SentencePiece Unigram算法结合字节回退(byte fallback)训练了四种不同词表规模(32k至150k)的分词器,并在四个评估集(FLORES+、信德语维基百科、新闻情感语料库和当代新闻语料库)上与十四种现有分词器进行了全面比较。其解决方案的关键在于:通过优化词表大小与分词策略,在保持高单标记保留率(single-token retention)的同时显著降低词元密度(fertility)与未知词符率(UNK rate)。实验结果表明,在64k词表规模下,新分词器平均每个词仅需1.147个词元,优于最佳现有分词器(1.208)并大幅优于XLM-RoBERTa(1.512),且88.6%的词可被表示为单一词元;在所有评估集上,其在词元密度、单标记保留率和未知词符率方面均表现最优或并列最优。此外,通过往返字符完整性测试发现,所提分词器无字符丢失,而部分现有分词器存在隐式字符删除(最高达0.85%)。研究仅评估了内在指标,未验证对下游任务的影响,代码与分词器文件已公开。

链接: https://arxiv.org/abs/2611.20126
作者: Ali Nawaz,Muqadas Fareed,Soonh Sadaf
机构: Shaikh Ayaz University (谢赫·阿亚兹大学); Department of Computer Science (计算机科学系)
类目: Computation and Language (cs.CL)
备注: 13 pages; code and tokenizers at doi: https://doi.org/10.5281/zenodo.23962474

点击查看摘要

Abstract:Sindhi is a low-resource Indo-Aryan language written mainly in an extended Perso-Arabic script. Several Sindhi tokenizers have been released on public model hubs, but they have not been compared with one another on shared data. We train SentencePiece Unigram tokenizers with byte fallback at four vocabulary sizes (32k to 150k) on about 1.10 million Sindhi sentences and a smaller, repeated Urdu, Persian and Arabic slice, from four public sources, and compare them with fourteen existing tokenizers on four evaluation sets: FLORES+, Sindhi Wikipedia, a news-derived sentiment corpus, and a contemporary news corpus. The metrics are fertility, single-token retention, UNK rate, byte-fallback rate, and round-trip character loss. At 64k, our tokenizer averages 1.147 tokens per word, against 1.208 for the best existing tokenizer and 1.512 for XLM-RoBERTa, and keeps 88.6% of words as single tokens. On fertility, single-token retention and UNK rate it is best or tied for best in all four evaluation sets at 64k and above; at 32k it trails one baseline on FLORES+. Margins over the best Sindhi-adapted baselines are small (0.002 fertility on FLORES+ at 64k), and we report no significance tests. A round-trip check finds no character loss for our tokenizer and silent character deletion (up to 0.85% of characters) in two existing ones; seven others lose less than 0.1%. We evaluate intrinsic metrics only, so effects on downstream tasks remain untested. Code and tokenizer files are public (DOI: https://doi.org/10.5281/zenodo.23962474).

[NLP-90] FRAAC: A French Multimodal Large Language Model and Dataset forPictogram-Based Augmentative and Alternative Communication

【速读】: 该论文旨在解决多模态大语言模型(Multimodal Large Language Models, MLLMs)在增强与替代性沟通(Augmentative and Alternative Communication, AAC)领域中无法可靠理解和生成图符序列的问题。其核心挑战在于缺乏标准化的图符语法体系以及平行文本-图符数据的稀缺性。为应对这一问题,论文提出一种名为“检索增强型图符选择”(Retrieval-Augmented Pictogram Selection, RAPS)的标注方法,其关键在于通过基于大语言模型(Large Language Model, LLM)的概念提取与图符词汇库的检索机制,对规则驱动的转录系统输出进行精细化修正,从而有效处理手写规则难以覆盖的情形(如命名实体、数字及多义词),且无需额外编写新规则。RAPS 在法语场景下构建了一个包含140万组平行样本的合成语料库,其中30万组已公开发布。人类评估结果显示,相较于基于规则的现有最佳方法,RAPS在对同一句子的多个图符转录结果进行排序时表现更优。此外,研究发现标准的参考基准评估指标虽能区分质量显著不同的系统,但难以有效区分性能相近的先进系统。最后,研究基于RAPS生成的数据训练了一款指令微调模型,支持跨模态交互:可接受图符或文本输入,并以图符、语音或文本形式输出响应,显著提升了AAC系统的灵活性与实用性。

链接: https://arxiv.org/abs/2611.20094
作者: Alexandre Audibert,François Portet,Benoit Favre,Maja Jonck Hjuler
机构: Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, 38000 Grenoble, France; NAVER LABS Europe, Meylan, France
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Pictograms are a core modality in augmentative and alternative communication (AAC), yet recent multimodal large language models (MLLMs) cannot reliably interpret or produce pictogram sequences. Two obstacles explain this gap: the absence of a standardized pictogram grammar and the scarcity of parallel text-pictogram data. We introduce Retrieval-Augmented Pictogram Selection (RAPS), an annotation method that refines a rule-based transcription system’s output with LLM-based concept extraction and retrieval over a pictogram vocabulary, handling cases where hand-written rules fail (named entities, numbers, polysemous words) without requiring new rules. Applied to French, RAPS yields a synthetic corpus of 1.4 million parallel pairs, of which 300k are publicly released. In a human ranking study where annotators order four pictogram transcriptions of the same sentence, RAPS outperforms the rule-based state of the art. Moreover, we show that standard reference-based metrics separate systems of clearly different quality but fail to rank strong systems against each other. Finally, we build an instruction-tuned model that supports interaction in several AAC modalities: it accepts pictogram or text input and answers in pictograms, speech, or text.

[NLP-91] he Silent Cost of Conversation History in LLM Reasoning

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在多轮信息交互系统中,因持续累积对话历史而导致的性能退化问题。研究发现,保留完整的对话历史会显著降低模型的最终任务成功率:在48个分片数学任务中,GPT-4o-mini和Mistral-7B的准确率分别下降0.21和0.29(经Holm校正后p值分别为0.026和0.004)。为深入探究此现象,研究设计了CDR-Stream框架,构建了100轮状态追踪对话,每轮均提供精确答案并进行模型难度校准。结果显示,在相同难度条件下,使用完整历史时模型准确率在20轮内从约0.6降至0.1,而采用经过求解器验证的状态摘要则可将准确率维持在0.52–0.59之间(Holm校正p < 0.001)。尽管一致性与逻辑连贯性下降幅度较小,稳定性反而上升,但现有过程级质量度量(包括配对就绪度、校准后的单流漂移检测器及基于重复性的指标)未能提供可靠且跨模型通用的早期预警信号。尤其在测试的监控机制中,单一100轮流仅对极端性能下滑具有检测能力。因此,解决方案的关键在于:将任务状态管理从对话转录文本中剥离,并在会话间实施独立监控,以避免历史累积带来的误差传播与性能衰减。

链接: https://arxiv.org/abs/2611.20071
作者: Ali Şenol
机构: Tarsus University (塔尔苏斯大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models (LLMs) in multi-turn information systems accumulate their own earlier replies and a growing task state. We asked whether this conversation history lowers accuracy, and whether process-level quality metrics warn before it does. Each analysis was fixed before its outcomes were computed. In Study 1, on 48 sharded math tasks, retaining its replies lowered final success by 0.21 for GPT-4o-mini and 0.29 for Mistral-7B (Holm-adjusted p = 0.026 and 0.004); Llama-4 Scout showed the same direction, not significant. Study 2 introduces CDR-Stream, 100-turn state-tracking dialogues with an exact answer per turn and per-model difficulty calibration. With the full history, accuracy fell from about 0.6 to 0.1 within twenty turns at matched difficulty, whereas a solver-verified state summary kept it 0.52-0.59 higher (Holm-adjusted p less than 0.001). Consistency and logical coherence fell far less than accuracy, and stability rose. None of these metrics, a paired readiness measure or calibrated single-stream drift detectors gave robust, model-general early warning, although in Study 2 the repeat-based metrics were complete only for GPT-4o-mini. In the monitors we tested, a single 100-turn stream had power only for very large drops. Systems built on LLMs should keep task state outside the transcript and monitor across sessions.

[NLP-92] Mandarin Wh-Indefinites Follow Nonveridicality in Multilingual LLM s Unlike English any

【速读】: 该论文旨在解决多语言大语言模型(Multilingual LLMs)在跨语言语义泛化中是否存在英语极性许可(polarity licensing)的迁移问题,具体聚焦于英语中“any”在向下蕴含(downward-entailing, DE)语境中的许可机制与汉语疑问不定词“什么(shenme)”在非真实条件(nonveridical contexts)中的许可机制是否发生混淆或分离。研究发现,尽管英语中“any”仅在DE语境下被许可而“shenme”则依赖非真实语境,二者本应分离,但11个通过测量门槛的模型在可能性、必然性和证据性副词语境下均对“shenme”表现出显著许可效应(强度为否定的0.42–2.33倍),而对英语“any”在同一语境中却无任何许可信号。这表明模型并未简单套用英语规则处理汉语,而是展现出对汉语语义结构的特定适应性。关键解决方案在于采用预注册的双语对照实验设计,通过48组匹配的双语句式与基于否定的对数概率差分-差异法(log-probability difference-in-differences)进行标准化评估,揭示了模型在跨语言极性许可上的非对称性与潜在语义敏感性。进一步探索性分析显示,即使在指令微调后,也难以通过方向性消融实现汉语许可的完全恢复,说明模型未将英语许可模式直接映射至汉语,而是存在语义层面的差异化学习机制。

链接: https://arxiv.org/abs/2611.20062
作者: Guofang Tang
机构: 未知
类目: Computation and Language (cs.CL)
备注: 11 pages, 2 figures

点击查看摘要

Abstract:Multilingual LLMs learn mostly from English. Do they carry English polarity licensing over to other languages? English any is licensed by downward-entailing (DE) contexts, Mandarin wh-indefinites such as shenme ‘what/something’ by nonveridical ones, and the two classes come apart: perhaps licenses shenme but not any, while only licenses any but entails existence. We test this dissociation in a pre-registered study of twelve open LLMs (0.5B-9B, four families), using 48 matched bilingual frames in 21 contexts and a log-probability difference-in-differences normalised by negation. All 11 models that pass the measurement gate license Mandarin wh-indefinites under possibility, necessity and evidential adverbs, at 0.42-2.33 times the strength of negation, while English any gets nothing in the same contexts. An exploratory extension to Qwen2.5-32B shows the same. The English side is weaker: any under only, first and regret is licensed but at a fraction of negation, so only 3 models show the full double dissociation in sentence probabilities. When prompted to choose between sentences, most gated models show it. In exploratory directional ablations, a direction that carries English licensing, held-out classes included, does not reach Mandarin licensing in most models. The models do not simply apply the English rule to Mandarin.

[NLP-93] A Dataset for Misogyny Detection in Polish Social Media: Transferability of Hateful Speech Detection from General to Specific Types of Hate

【速读】: 该论文旨在解决生成式AI在检测泛化型有害语言(harmful language)时,其能力能否有效迁移至特定形式仇恨言论(如针对女性的性别仇恨言论,misogyny)的问题。研究聚焦于波兰语社交媒体评论中的性别指向性仇恨言论检测,针对该领域缺乏专用标注资源的现状,提出了一个新的波兰语社交媒体评论数据集,专门标注了针对女性的仇恨言论。解决方案的关键在于通过对比三类模型的表现:基于通用有害语言数据训练的模型、专为安全任务优化的模型以及直接针对性别仇恨言论微调的模型,系统评估从通用到特定类型仇恨言论的迁移能力。实验结果表明,通用有害语言检测性能优异的模型在性别仇恨言论检测上表现并不一致,而经过安全任务专项训练的模型展现出显著更强的泛化能力。该研究不仅贡献了一个新的波兰语性别指向性有害语言资源,还揭示了先前安全训练的范围与粒度是实现从泛化到特定形式仇恨言论检测有效迁移的关键因素。

链接: https://arxiv.org/abs/2611.20044
作者: Ewelina Gajewska
机构: Warsaw University of Technology (华沙理工大学); Warsaw, Poland
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Harmful language detection is commonly evaluated on broadly defined hate speech and toxicity benchmarks, yet it remains unclear whether models that successfully detect harmful language in general can transfer this capability to more specific forms of hate. We investigate this problem for misogyny detection in Polish, an area for which dedicated resources remain scarce. We introduce a new dataset of Polish social media comments annotated for misogyny, gender-directed hateful speech. Using this resource, we first assess the reliability of LLM-assisted annotation and then evaluate general-to-specific transfer by comparing models trained on general harmful language data, safety-specialized models, and models fine-tuned directly for misogyny detection. Our results show that strong performance on general harmful language detection does not transfer uniformly to misogyny; safety-specialized models demonstrate substantially stronger generalization ability. The study contributes a new Polish resource for gender-directed harmful language and demonstrates that the scope and granularity of prior safety training are important factors in transferring harmful language detection from general to specific forms of hate.

[NLP-94] Offline-First Quantized AI: Design and Evaluation of an On-Device Maternal-Health Assistant for Low-Connectivity Environments

【速读】: 该论文旨在解决在互联网连接不可靠的环境下,如何实现临床适用且可在本地设备上可行运行的孕产妇健康信息辅助系统的问题。其核心挑战在于确保生成式 AI (Generative AI) 在离线状态下仍能提供高质量、安全且上下文连贯的医疗建议,同时满足移动端资源约束下的实时性与内存效率要求。解决方案的关键在于构建一个集成化的离线 Android 助手,融合经过量化处理的语言模型(Gemma 4 E2B 与 Qwen3.5 0.8B)、经临床验证的本地知识库、对话上下文管理机制以及由模型主动触发的本地检索增强生成(RAG-on/off)功能。实验表明,Gemma 在响应质量、上下文保留和安全性方面表现更优,而 Qwen 则在生成速度与内存占用上更具优势;临床医生对 150 条 Gemma 响应的评估平均得分为 3.80/4,其中 98% 符合接受标准。此外,通过对比实证发现,启用本地 RAG 能提升部分响应的有用性、正确性和可操作性,尽管存在少数不良影响,但整体在条件子集上展现出更优的性能。最终,物理设备测试证实该系统可在仅 6 GB 内存的安卓手机上实现离线响应(中位延迟 14.3 秒),并支持断网后重新连接时的数据同步,充分验证了系统的临床可用性与实际部署可行性。

链接: https://arxiv.org/abs/2611.20012
作者: Okenla AbdulQahhar Abimbola,Maurice Abraham John,Adepoju Jeremiah Moyondafoluwa,Omolaiye Samuel,Oko-Oba Sandra Chioma,Owoo Oluwamayowa Emmanuel,Thanni Adewuyi,Lily Aimas-Love,Benson Eniola,Abdulquddus Ajibade,Sylvia Essessien,Anyanwu Princess Emmanuella,Adereni Abiodun
机构: HelpMum Africa
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Offline language models could extend access to maternal-health information where internet connectivity is unreliable, but deployment requires both clinically appropriate responses and feasible local execution. This study presents an offline Android assistant integrating ready-made quantized language models, a clinically validated local knowledge package, conversational context, and model-initiated local retrieval. Gemma 4 E2B and Qwen3.5 0.8B were compared across 45 conversations and 312 turns per configuration, including a dedicated offline agentic benchmark. Gemma achieved higher response-quality, context-retention, and safety scores; Qwen offered faster generation and lower memory use. Clinician assessment of 150 Gemma responses across nine clinical domains yielded a mean quality score of 3.80/4, with 147 responses meeting the acceptance criterion. A paired RAG-on/off experiment evaluated 74 questions: retrieval improved 11 responses, left 56 neutral, produced four adverse response changes, and left three pairs unassessable. RAG-on had descriptively higher mean usefulness, safety, correctness, and actionability in the conditional 70-pair subset excluding the four flagged pairs from both conditions. Physical-device testing demonstrated offline responses on a 6 GB Android handset, with a 14.3 s median latency, alongside successful reconnection synchronization in the integrated application. Together, the findings document clinical response quality and practical performance of the integrated offline maternal-health assistant.

[NLP-95] he novelty exposed: ablation reproducibility and transparency in 40 technical reports of open-weight large language models (LLM s)

【速读】: 该论文旨在解决当前大型语言模型(Large Language Model, LLM)快速迭代背景下,大量新模型架构在宣称“创新性”时缺乏充分验证与透明度的问题。具体而言,研究关注这些新模型的技术报告是否真正提供了可验证的创新证据,以及其可复现性和透明度水平。解决方案的关键在于系统性地评估三方面内容:一是通过分析Hugging Face库中2025年新增48个模型的参考代码,揭示其核心模块来源;二是对随机抽取的40份技术报告进行审查,检验其关于架构或训练创新的主张是否有消融实验支持;三是评估报告在数据混合、计算量、训练运行次数等关键复现信息上的披露程度。研究发现,多数新模型的核心组件源自已有模型(如Llama),仅有少数报告提供受控对比实验,且极少数报告完整披露训练配置和计算资源。此外,尽管未发现刻意隐藏已知组件的现象,但部分模型仍存在命名误导风险。为此,作者提出七项报告规范,并应用于两份报告以示范改进方向。该研究强调,创新性声明应建立在可验证、可复现的数据基础之上,而非仅依赖术语包装。

链接: https://arxiv.org/abs/2611.20003
作者: Bahaeddin Türkoğlu
机构: Ankara University (安卡拉大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Empirical study with an original dataset: 40 LLM technical reports coded by two independent model coders with script-verified quotes, plus a code-level comparison of open-weight architectures in Hugging Face transformers. 20 pages, 2 figures, 2 tables. Data, coding protocol and scripts in the ancillary files

点击查看摘要

Abstract:In 2025 alone, 48 new large language model (LLM) architectures were added to the Hugging Face transformers library, and many of them arrive with a technical report that presents the model as new. What is actually new in these models, and is it tested? We put this question to LLM technical reports. We look at three things: what the reference code shows about how new open-weight architectures differ from earlier ones; what 40 randomly sampled technical reports claim as new and whether they support it with ablation studies; and what they disclose for reproducibility and transparency. In the library, 39 of the 48 architectures added in 2025 take at least two of their three core blocks from another model, most often Llama, although part of this reflects how the library writes code. In the reports, 19 of 73 architectural or training novelty claims come with a controlled with/without comparison; 4 of 40 reports state more than one training run; 8 rerun the models they compare against under one setup; and 1 discloses data mixture, token count, and compute together. We also looked for known components given new names: none of five candidate cases hid its origin, though we did not study model names. We propose seven reporting rules for LLM release reports and apply them to two reports. We do not judge model quality, and an untested claim may still be true.

[NLP-96] Anatomy of an Agent Harness: Decisions Invariants and Corner Cases

【速读】: 该论文旨在解决生成式 AI(Generative AI)代理框架(agent harness)在实际应用中因决策逻辑隐含于代码而缺乏可复现性与可验证性,导致相同边缘情况反复出现的问题。其核心解决方案是将代理框架中的关键决策抽象为可形式化、可测试的不变量(invariants),并通过系统化建模实现对这些决策的显式记录与验证。关键创新在于构建了一个包含6组共30张“决策卡”(decision cards)的体系,每张卡片明确界定决策选项、成本、不变量约束、可能破坏该不变量的边界情况、对应的测试方法及证据标签;同时引入六类故障归因类别,并基于脚本化服务提供者、行为录制与人为植入变更的测试方法,形成可执行的验证流程。此外,论文还提供了一套检查清单,用于实践验证。该体系以作者自研桌面级代理框架Crewroom为基础,结合对10个其他代理框架的手动测试及开源案例分析,确保其设计模式具有现实参考价值,但并非完全独立于已有实践。在83条不变量条款中,56条已验证、12条可信、15条开放,表明其具备一定成熟度,但仍有待完善。该研究不涉及回答质量、成本或平台性能评估,也未对其他代理框架进行排名或评价。

链接: https://arxiv.org/abs/2611.19960
作者: Resham Joshi
机构: 未知
类目: Computation and Language (cs.CL); Software Engineering (cs.SE)
备注: 33 pages, 8 tables, 30 decision cards, checklist

点击查看摘要

Abstract:An agent harness is the program around a language model that turns its replies into actions: it builds each request, runs the tools the model asks for, decides what the model may do without asking, stores the conversation, recovers from failures, and shows a person what happened. Most of these decisions are made in code and rarely written down, so each harness meets the same corner cases again. This paper writes them down as invariants to test. It gives a system model with trust boundaries; 30 decision cards in six groups, each with options and costs, an invariant, the corner cases that break it, a test and an evidence tag; six cause classes for failures; a test method built on scripted providers, recorded behaviour and planted changes; and a checklist. The cards come from building one desktop harness, Crewroom, the author’s own, and from manually testing ten other harnesses and studying open-source harnesses as reference; some patterns in the cards were adopted from those, so the catalogue is not independent of them. Of the 83 invariant clauses, 56 are tagged Tested, 12 Believed and 15 Open. The checklist is applied to the author’s harness as one worked example, by its builder with no second rater: of its 30 items, 11 hold, 14 do not and 5 were not tested. It is not a ranking. The paper makes no claims about answer quality or cost or other platforms, and does not evaluate or rank other harnesses.

[NLP-97] Learning When to Continue and When to Stop: Termination-Aware OPSD

【速读】: 该论文旨在解决生成式 AI(Generative AI)在基于自蒸馏的语言模型推理过程中存在的“终止不匹配”问题,即学生模型无法准确学习何时继续推理、何时停止生成。这一问题表现为两种相互对立的失效模式:一是过早终止(premature termination),即推理尚未充分进行便提前结束;二是重复驱动的过度生成(repetition-driven over-generation),即在推理已停滞的情况下持续产生冗余内容。研究发现,在传统在线策略自蒸馏(OPSD)训练中,教师与学生对句子结束符(EOS)的判断差异逐渐增大,且长度受限的推理轨迹常被无效重复主导。为解决该问题,作者提出一种新的终止感知型在线策略自蒸馏方法(TA-OPSD),其核心在于在原有目标基础上引入两个针对性优化目标:停/续蒸馏(Stop/Continue Distillation)通过显式对齐教师与学生在“是否继续推理”上的决策,缓解过早终止问题;重复感知反似然训练(Repetition-Aware Unlikelihood Training)则通过抑制学生在失败轨迹中的重复性延续行为,缓解过度生成问题。实验结果表明,TA-OPSD在数学推理任务上对 Qwen3 系列(1.7B/4B/8B)模型均显著优于 OPSD 及其他基线方法,且消融实验和行为分析验证了两个新目标能有效分别缓解对应类型的终止错误,从而实现更精准的推理终止控制。

链接: https://arxiv.org/abs/2611.19925
作者: Chaoqun He,Wannian Yi,Changyi Sun,Chaojun Xiao,Lijie Wen
机构: Tsinghua University (清华大学)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:On-Policy Self-Distillation (OPSD) improves language model reasoning by distilling from a frozen teacher initialized identically to the student and conditioned on privileged information. However, we identify an underexplored termination mismatch in OPSD: the student fails to reliably learn when to continue and when to stop. This mismatch manifests in two opposing failure modes-premature termination, where reasoning stops before sufficient progress is made, and repetition-driven over-generation, where the model continues generating after reasoning has stalled. Our analysis shows that the teacher-student discrepancy on EOS grows during OPSD training, while length-capped trajectories are frequently dominated by unproductive repetition. To address this issue, we propose TA-OPSD (Termination-Aware On-Policy Self-Distillation), which augments the original OPSD objective with two targeted objectives. Stop/Continue Distillation addresses premature termination by explicitly aligning the teacher and student on whether reasoning should continue, while Repetition-Aware Unlikelihood Training addresses repetition-driven over-generation by discouraging repetitive continuations in the student’s failed trajectories. The resulting objective improves termination behavior from complementary directions: continuing when further reasoning is needed and stopping when reasoning becomes unproductive. Experiments on mathematical reasoning tasks with Qwen3 (1.7B/4B/8B) show that TA-OPSD consistently outperforms OPSD and other baselines across all three model scales. Beyond the performance gains, ablation and behavioral analyses show that the two proposed objectives effectively mitigate their corresponding termination failures, reducing both premature termination and repetition-driven over-generation.

[NLP-98] he Vocabulary Tax: Safely Trimming Unused Tokens from LLM s for On-Device Use

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)及语音模型在部署过程中因包含冗余词汇表(vocabulary)而导致的“词汇税”(vocabulary tax)问题,即用户即使仅需支持特定语言,仍需下载全部词汇表所对应的嵌入矩阵(输入嵌入、解耦输出头及层间嵌入),造成显著的存储与传输开销。其核心解决方案是提出一种基于合并闭包(merge closure)的词汇表裁剪方法,通过保留部署目标语言书写系统的必要词元行,实现精确裁剪;同时引入裁剪证书(trim certificate),基于修剪文本本身生成一个基于Good-Turing估计的指标,用于预测裁剪后分词器对新文本的分词差异程度以及所需额外文本量,从而确保性能稳定性。该方法在不依赖测试数据的情况下,能将模型体积压缩至平均移除89.7%的词元行,且比特/字符(bits per character)上升不超过2.0%,同时保持模型预测能力不变。整个流程包括词汇税普查、智能裁剪与证书生成,已开源为Python工具包,并发布了41种语言的每模型词汇税数据集。

链接: https://arxiv.org/abs/2611.19913
作者: Sattaya Singkul
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 23 pages, 8 figures, 16 tables. Code: this https URL ; Data: this https URL

点击查看摘要

Abstract:Open large language models (LLMs) and speech models ship vocabularies of up to 262k tokens, and every user downloads the tables those tokens index (the input embedding, an untied output head and, in recent on-device models, per-layer embeddings), whatever languages they need. We call the bytes in rows a deployment never uses the vocabulary tax and measure it, without downloading weights, for 1,244 generative and speech-recognition models on the Hugging Face Hub. For a Thai and English deployment it is a median 4.4% of model bytes below 0.5B parameters and 2.1% above 8B, and 1.93 GB in Gemma 4 E2B: it falls hardest on the small models that edge devices run. Keeping only the rows of the deployment’s writing systems is exact. Keeping the tokens seen in the user’s own text removes much more, a median 29.0% below 0.5B, but the usual way of doing it breaks BPE tokenizers: it drops the intermediate tokens that merges build on, so new text is split differently 1.9 times as often as the standard check reports. We repair this with merge closure, at a cost of 5.9% more rows, and add a trim certificate: a Good-Turing estimate, from the trimming text alone, of how much new text the trimmed tokenizer will split differently (correlation 0.996 over 300 settings in 22 languages) and of how much more text a target needs. Without test data our method never exceeds a 2% target, while prior keep rules checked on test data exceed it in 9-13 of 44 settings. On six models of 270M to 3B parameters in 11 languages, trimmed models keep their predictions: where the certificate passes, a median 89.7% of rows is removed and bits per character rise by at most 2.0%. The census, the trimming and the certificate are released as an open Python package, and the per-model tax for 41 languages as an open dataset.

[NLP-99] LoopSkim: Lossless Multi-Depth Speculative Decoding for Looped Language Models

【速读】: 该论文旨在解决循环语言模型(Looped Language Models, LLMs)在自回归解码(Autoregressive Decoding, AR)过程中效率低下的问题。传统方法如Ouro和Raven虽通过权重共享块多次迭代以提升生成质量,但每轮解码仍需顺序执行,导致延迟高。现有方案如流水线自推测解码(Pipelined Self-Speculative Decoders)虽通过浅层预读取草案令牌并全深验证来重叠计算,但仍存在冗余遍历与验证开销。本文提出的关键解决方案是LoopSkim,其核心创新在于将流水线的深度层级本身作为验证器(verifier),利用每个循环末尾的读出结果进行即时比对:草案令牌从上下文复制引入(类似提示查表解码),在后续循环中若出现矛盾读出则直接替换对应深度的输出,仅全深读出才正式提交令牌。该机制避免了额外前向传播,实现无损解码(lossless decoding),且在任意检查次数和草案链长度下均保持精确算术一致性。由于无需训练,其在贪婪解码与采样场景下平均加速比分别达到1.73倍于LoopSpec,且所有模型均受益于首次循环后至全深之间的检查优化;引入可选草案头可进一步构建第二条草案链,使平均加速提升至1.99倍。

链接: https://arxiv.org/abs/2611.19896
作者: Muhan Xiao,Shijian Li
机构: Zhejiang University (浙江大学); College of Artificial Intelligence (人工智能学院); College of Computer Science and Technology (计算机科学与技术学院)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 23 pages, 3 figures, 16 tables

点击查看摘要

Abstract:Looped language models such as Ouro and Raven apply one weight-shared block several times per token, so autoregressive decoding (AR) runs every loop of every token in sequence. Pipelined self-speculative decoders overlap these loops by drafting tokens from shallow readouts and verifying them at full depth. Yet every loop of a looped LM ends in a readout of the next token, and draft tokens pass through every depth anyway, so each depth can check them without an extra pass. We introduce LoopSkim, in which the depth of the pipeline is the verifier. Draft tokens copied from the context, as in prompt lookup decoding, enter after the first loops, any later readout that contradicts one replaces it at that depth, and only full-depth readouts commit tokens. LoopSkim fixes the random draws of every position in advance, so each check compares two tokens and decoding stays lossless in exact arithmetic under sampling for any number of checks and draft chains. Without training, its speedup over AR averages 1.73x that of LoopSpec across greedy decoding and sampling, and its checks between the first loops and full depth speed it up on every model. An optional draft head adds a second chain and raises this average to 1.99x.

[NLP-100] Who Writes the Codebook? An Eight-Step Framework for Auditing LLM -Generated Category Systems

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在定性证据编码过程中因模型自身生成的分类体系存在显著差异,而现有编码评估基准通常固定分类体系所带来的方法论偏差问题。其核心挑战在于如何量化和区分模型在代码本创建、编码应用及代码分布等环节中所产生的系统性变异,从而揭示模型依赖性对定性分析结果的影响。解决方案的关键是提出“定性方差审计框架”(Qualitative Variance Audit Framework),该框架采用八步流程,在受控输入条件下对比不同模型在共享参考代码本与自建代码本下的编码表现,通过结合模型专属与共享参考代码本,有效分离整体分析管道中的变异来源与同一分类体系内的应用差异。实证研究基于668个组织访谈、13,166个文本单元及五家供应商的八款模型,生成48个代码本和270次编码运行,结果显示模型间构建的类别结构差异显著,模型内一致性(均值为0.76–0.92)高于跨模型在共享代码本上的一致性(约0.69)。该框架不直接评估解释有效性或理论质量,而是聚焦于可靠性与表征后果的评估,为大语言模型辅助扎根理论研究提供了可量化的模型依赖性暴露机制。

链接: https://arxiv.org/abs/2611.19872
作者: Thilo Tamme,Anton Hantel,Jakob Scheppers
机构: 未知
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)
备注:

点击查看摘要

Abstract:Large language models (LLMs) increasingly create the categories through which qualitative evidence is interpreted. Existing coding benchmarks often hold these categories fixed. We propose the Qualitative Variance Audit Framework, an eight-step method that compares codebook creation, repeated application and code distributions under controlled inputs. It combines model-owned and shared-reference codebooks to distinguish variation in the overall pipeline from application differences within a common category system. An application to 668 organisational interviews, 13,166 text units and eight models from five vendors yields 48 codebooks and 270 coding runs. Models create markedly different category structures; mean within-model reliability ranges from .76 to .92, whereas between-model agreement on the shared codebook is about .69. The framework complements LLM-assisted grounded theory by exposing model dependencies before substantive interpretation. It assesses reliability and representational consequences, not interpretive validity or the quality of resulting theory.

[NLP-101] Bioformer-RetroMAE: retrieval-oriented pretraining of a compact biomedical language model for literature search

【速读】: 该论文旨在解决生物医学文献检索中密集检索模型(dense retrieval)存在的两大核心问题:一是现有基于BERTBase的生物医学双编码器(bi-encoder)在大规模语料索引和低延迟查询编码场景下计算成本过高;二是掩码语言建模(MLM)预训练策略未能显式优化双编码器所依赖的[CLS]向量表示能力。为此,作者提出Bioformer-RetroMAE,一种参数量仅为4300万的紧凑型生物医学检索专用编码器,通过在约3200万篇PubMed标题与摘要上对Bioformer8L进行持续预训练(continual pretraining),并引入RetroMAE自监督学习范式,以更高效地学习全局文本表征。其关键创新在于利用RetroMAE实现无需额外计算开销的高质量[CLS]向量优化,在保持模型尺寸与推理成本不变的前提下显著提升检索性能。实验表明,该模型在BioASQ Task 12B测试集上超越了参数量更大的BioBERT和PubMedBERT,在平均精度(MAP)和召回率@100等指标上分别提升37.0%和29.1%;相较于原始Bioformer8L,MAP与Recall@100分别提升53.6%和32.8%。此外,在与BM25结合的混合检索中,采用加权倒数排名融合(weighted reciprocal rank fusion)后,其在MAP与NDCG@10上的增益均优于与PubMedBERT或BioBERT融合的结果,验证了其在实际应用中的优越性。

链接: https://arxiv.org/abs/2611.19862
作者: Jialu Li,Li Fang
机构: 未知
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Dense retrieval is increasingly used for biomedical literature search and retrieval-augmented generation (RAG). Most biomedical bi-encoders are built on BERTBase-sized encoders pretrained with masked language modeling (MLM). These encoders are costly for corpus-scale indexing and latency-sensitive query encoding, and MLM does not explicitly train the [CLS] vector that a bi-encoder uses to represent a whole text. We present Bioformer-RetroMAE, a compact (43M-parameter) retrieval-oriented biomedical encoder obtained by continually pretraining Bioformer8L with RetroMAE on approximately 32 million PubMed titles and abstracts. After bi-encoder fine-tuning with BM25 hard negatives, we evaluated it on BioASQ Task 12B test questions by retrieving from all PubMed abstracts. Bioformer-RetroMAE outperformed BioBERT and PubMedBERT bi-encoders fine-tuned under identical settings on all metrics, despite having about 40% of their parameters and producing 512-dimensional rather than 768-dimensional vectors. Compared with PubMedBERT, it improved mean average precision (MAP) by 37.0% and Recall@100 by 29.1%. Compared with the original Bioformer8L, RetroMAE pretraining improved MAP by 53.6% and Recall@100 by 32.8% without changing model size or inference cost. In hybrid retrieval, fusing Bioformer-RetroMAE with BM25 through weighted reciprocal rank fusion improved MAP and NDCG@10 over BM25 alone by 4.9% and 4.6%, respectively, larger gains than those obtained by fusing BM25 with PubMedBERT or BioBERT.

[NLP-102] Bridging the Terminology Gap: A Data-Centric Alignment Framework for Multilingual Translation in Cross-Border Corporation

【速读】: 该论文旨在解决全球化背景下跨企业协作中多语言翻译所面临的四大核心瓶颈:领域专用术语匮乏、术语翻译僵化、指令遵循能力弱以及长指令场景下的高推理延迟。其解决方案的关键在于提出一种以数据为中心的“弥合术语鸿沟”(Bridging the Terminology Gap, BTG)框架:首先,通过从领域文档中挖掘跨语言术语对,并利用大语言模型(Large Language Model, LLM)合成富含术语的平行语料;其次,采用结合启发式规则与XCOMET评分的混合过滤策略,保障数据质量;最后,采用两阶段训练范式——先通过监督微调(Supervised Fine-Tuning, SFT)实现术语精准注入,再通过强化学习(Reinforcement Learning, RL)优化输出偏好,有效平衡领域知识融合与模型可靠性。实证结果表明,该框架在术语准确性与指令遵循能力方面均显著提升,已在实际跨边界协作平台中部署应用。

链接: https://arxiv.org/abs/2611.19858
作者: Jinsheng Wei,Song Cheng,Bo Xiao,Ke Zeng
机构: Meituan LongCat Interaction Team(美团龙猫交互团队); Peking University(北京大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:As globalization accelerates the demand for seamless cross-border enterprise collaboration, multilingual translation remains bottlenecked by insufficient domain-specific terminology, rigid terminology translation, weak instruction-following capabilities, and high inference latency in long-instruction scenarios. To overcome these barriers, we propose Bridging the Terminology Gap (BTG), a data-centric framework. First, BTG mines cross-lingual term alignments from domain documents and leverages LLMs to synthesize terminology-rich parallel corpora. Second, a hybrid filtering strategy combining heuristics and XCOMET scoring ensures high data fidelity. Third, a two-stage training paradigm employs supervised fine-tuning (SFT) for targeted terminology injection, followed by reinforcement learning (RL) to refine output preferences, effectively balancing domain knowledge integration with model reliability. Extensive offline evaluations (e.g., XCOMET, GEMBA, Acc@T) and online human assessments demonstrate significant improvements in terminology accuracy and instruction-following capabilities, validating the effectiveness of our framework, which is now deployed on our cross-border collaboration platform.

[NLP-103] A phase diagram for training and inference in language models

【速读】: 该论文旨在解决生成式 AI(Generative AI)中表示几何结构(representation geometry)与语言模型实际行为之间关系不明确的问题。其核心挑战在于理解Transformer架构内部表征在前向传播过程中的动态演化如何决定模型的推理性能。论文的关键解决方案是通过引入两个集体坐标——文档可区分性(document distinguishability)和文档维度(document dimension),构建一个二维相图来刻画语言模型在不同层的表示几何特征,并将推理过程建模为沿网络深度的动态流动(inference flow)。研究发现,未训练模型的推理流大多趋向于相图中的坍缩区与噪声区,而经过训练的8个来自5种架构家族的模型则均进入并稳定于相图中心区域。进一步地,通过离散朗之万方程(discrete Langevin equation)描述流动演化,揭示二者差异的本质在于该中心区域内有效作用力的类型:未训练模型表现为鞍点型(saddle-type)力,而训练后模型则呈现稳定节点型(stable-node-type)力。在Pythia-1.4B模型上,训练过程被解释为一次相变(phase transition),从而为表征几何与语言模型行为之间的统一关联提供了理论框架。

链接: https://arxiv.org/abs/2611.19848
作者: Zhuo-Yang Song,Zeyu Li,Qing-Hong Cao,Ming-xing Luo,Hua Xing Zhu
机构: Peking University (北京大学); LightRay AI Lab (LightRay AI 实验室); Beijing Computational Science Research Center (北京计算科学研究中心)
类目: Machine Learning (cs.LG); Disordered Systems and Neural Networks (cond-mat.dis-nn); Statistical Mechanics (cond-mat.stat-mech); Computation and Language (cs.CL)
备注: 20 pages, 12 figures, 4 tables

点击查看摘要

Abstract:A language model is a complex system whose behavior is determined jointly by the Transformer and the corpus, yet how the representation geometry in the Transformer relates to the actual behavior of language models remains open. We analyze how the representation geometry changes along the forward propagation of actual language models and characterize the representation geometry of a document at a given layer by two collective coordinates, document distinguishability and document dimension. On the phase diagram spanned by these collective coordinates, inference is represented as a flow along depth. We find that, for untrained models, depending on their initialization scale, most of the 156 inference flows move toward the collapse and noise regions of the phase diagram. By contrast, the inference flows of 8 trained models from 5 architecture families all enter and reside in the central region of the phase diagram. Describing the evolution of the flows with a discrete Langevin equation, we find that the difference between the inference flows of untrained runs and those of trained models reduces to whether the effective force on the flows in this region is of saddle or stable-node type. On Pythia-1.4B, we interpret the training process as a phase transition. These results give a unified gateway to the relation between representation geometry and language-model behavior.

[NLP-104] Grounding Language Models in Speech Does Not Fix “AI Flavor”: Calibrated LLM Judges Goodhart Traps and a Preference-Tuning Fix

【速读】: 该论文旨在解决中文大语言模型输出中存在的显著“生成式AI风味”(AI flavor)问题,表现为句式模板化、书面语连接词堆叠、重复循环及格式泄露等特征。主流假说认为,这是由于仅使用文本进行训练而缺失了人类写作中的内在语音层(inner-speech layer),因此提出通过引入语音或韵律信号来缓解该问题。论文通过在0.6B规模模型上实施三种受控干预验证该假设:(i) 语音音素交错的持续预训练在文本匹配控制下未提升口语性,反而导致重复率上升(+0.13 [+0.03, +0.23]);(ii) 基于韵律计划的条件生成仅达到0.257的遵循度,远低于预注册的0.50阈值;(iii) 基于语音合成(TTS)韵律特征距离构建的规则奖励在双盲评估中对人类(8/16)和校准后的LLM裁判(9/16)均不可感知。研究发现根本原因在于评估机制缺陷:单次盲评(n~20)仅能检测极端偏好(≥85%)。为此,作者构建了一套校准的LLM裁判协议,该协议通过捕捉已知答案对的测试项实现高准确率(16/16;单文本22/24,同主题对24/24,人类为0.75),并采用双独立裁判与A/B交换标注,以分歧计为平局。直接使用裁判一致标签作为直接偏好优化(DPO)数据对(无需奖励模型),首次获得正向结果:保留盲评胜率达0.833 [0.699, 0.972],重复率从0.114降至0.000,且MMLU指标保持不变。进一步识别出“古德哈特通道”(Goodhart channel)——裁判的长度偏好经成对DPO传递后引发输出长度坍缩(中位数75字符对比300字符提示),通过引入SFT锚点修复(风格调节系数w=0.1可保留增益,w=0.3则失效)。将偏好数据从89对扩展至240对未带来额外可见收益。研究发布RA-Eval v0.1:一个包含382个裁判标注偏好对的三层中文生成式AI风味评估框架。

链接: https://arxiv.org/abs/2611.19840
作者: Dongwu Qian
机构: Independent Researcher(独立研究员); RA-Eval v0.1(数据与代码托管于Zenodo, DOI 10.5281/zenodo.24237533)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 11 pages, 2 tables. Code and data: this https URL

点击查看摘要

Abstract:Chinese outputs of large language models carry a perceptible “AI flavor”: formulaic sentence frames, stacked written-register connectives, repetition loops, and layout leakage. A popular hypothesis blames text-only training for the missing inner-speech layer of human writing; injecting speech or prosody signals should therefore reduce it. We test it with controlled interventions on a 0.6B model; all three fail: (i) phoneme-interleaved continued pretraining yields no spokenness gain under text-matched controls while raising repetition (+0.13 [+0.03, +0.23]); (ii) prosody-plan conditional generation reaches an adherence of only 0.257 against a pre-registered 0.50 gate; (iii) a rule-based reward built on TTS prosody-feature distance is imperceptible in double-blind evaluation by both humans (8/16) and a calibrated LLM judge (9/16). A root cause is evaluation: single-rater blind evaluation at n~20 only detects extreme (=85%) preferences. We therefore build a calibrated LLM-judge protocol: the judge passes catch trials on known-answer pairs (16/16; single-text 22/24, same-topic pairs 24/24, versus 0.75 for the human rater), then annotates with two independent judges under A/B swapping, counting disagreement as a tie. Using the judge’s agree labels directly as DPO pairs (no reward model), we obtain the first positive result: held-out blind win rate 0.833 [0.699, 0.972], repetition 0.114 to 0.000, unchanged MMLU. We further identify a Goodhart channel: the judge’s length preference transduces into output-length collapse via pairwise DPO (median 75 chars vs a 300-char prompt); an SFT-anchor repairs it (w is a style knob: w=0.1 preserves gains, w=0.3 does not). Scaling preference data from 89 to 240 pairs yields no additional judge-visible gain. We release RA-Eval v0.1: a three-layer Chinese AI-flavor evaluation framework with 382 judge-annotated preference pairs. Comments: 11 pages, 2 tables. Code and data: this https URL Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG) Cite as: arXiv:2611.19840 [cs.CL] (or arXiv:2611.19840v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.19840 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Related DOI: https://doi.org/10.5281/zenodo.24237533 Focus to learn more DOI(s) linking to related resources

[NLP-105] EfficientSkill: Learning Skills for Token-Efficient Multi-Turn Agents

【速读】: 该论文旨在解决生成式 AI(Generative AI)在多轮任务中因持续积累边缘案例规则而导致提示(prompt)膨胀、令牌消耗增加以及代理陷入冗长且易出错循环的问题。其核心解决方案是提出一种名为 EfficientSkill 的两阶段框架:首先在不产生额外令牌开销的前提下发现一个可靠的任务技能,随后基于该技能进行“精度保障的剪枝”(accuracy-guarded pruning),以压缩指令、限制代理的交互轮次,并将可复用代码无缝迁移至工作区工具中,实现零提示成本的代码复用。实验表明,在 MATH-500、ALFWorld 与 SpreadsheetBench 等多个基准测试上,EfficientSkill 在 12 种配置中的 11 种显著减少了输出令牌数(最高降低 80%),同时提升了平均准确率,尤其在 Gemini 3.8 Flash 与 Grok 4.1 Fast 模型上于 SpreadsheetBench 实现了 +14.3 和 +17.5 个百分点的性能提升,有效通过紧凑工作流替代卡死的循环。该方法通过减少生成令牌和重放历史,全面降低了总令牌使用量与服务成本,无论是否启用完整前缀键值缓存(full-prefix KV caching)。

链接: https://arxiv.org/abs/2611.19823
作者: Deep Shah,Sanket Badhe,Nehal Kathrotia,Priyanka Tiwari
机构: Google LLC(谷歌); Purdue University(普渡大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Reflective skill optimizers improve multi-turn LLM agents by revising instructions from failed and successful rollouts, yet accumulating edge-case rules bloats prompts, inflates token use, and pushes agents into long, error-prone loops. We introduce EfficientSkill, a two-step framework that first discovers a reliable skill without token penalties, and then warm-starts accuracy-guarded pruning from that skill to compress instructions, bound agent turns, and offload reusable code into workspace utilities with zero implementation prompt cost. Across four frontier, mid-tier, and open-weight models on MATH-500, ALFWorld, and SpreadsheetBench, EfficientSkill cuts output tokens over unconstrained EvoSkill on 11 of 12 configurations (by up to 80%) while raising mean accuracy on every benchmark, including +14.3 and +17.5 percentage-point gains on SpreadsheetBench with Gemini 3.8 Flash and Grok 4.1 Fast by replacing stalled loops with compact workflows. By shrinking both generated tokens and replayed history, EfficientSkill lowers total token usage and serving cost across every evaluated configuration, with or without full-prefix KV caching.

[NLP-106] Machine interpreting competitiveness: a discourse-sensitive comparison of meaning and target-language quality in English-Chinese simultaneous interpreting

【速读】: 该论文旨在解决生成式 AI (Generative AI) 在口译(Machine Interpreting, MI)领域中的实际表现与人类同声传译员之间差距的系统性认知不足问题,尤其聚焦于MI在不同话语模式(计划性正式演讲与即兴访谈)下的表现差异及其背后的原因。其解决方案的关键在于构建一个以话语层面(discourse-level)交际活动为核心、兼顾语义保真度与目标语言质量的评估框架,并通过对比五种商用MI系统与26名人类译员(14名专业译员、12名高级学生)在真实跨语言同步口译任务中的表现,揭示了MI在不同语境下的局限性。研究发现,尽管最强的MI系统在计划性正式演讲中已接近人类高水平的语义传递能力,但在即兴访谈中则显著落后,表现为语义扭曲、话语追踪失败及上下文不敏感等问题;而人类译员则更善于通过选择性省略、重构和语境推断维持整体交际轨迹。此外,在目标语言质量方面,人类译员的优势更为稳定,机器系统常出现不可理解或无意义输出及连贯性缺陷,尤其在自发性话语中尤为突出。因此,该研究强调了将话语敏感性(discourse sensitivity)纳入未来MI系统设计与评估体系的重要性,为口译实践、译员培养以及技术发展提供了关键依据。

链接: https://arxiv.org/abs/2611.19802
作者: Yuanyuan Ding
机构: 未知
类目: Computation and Language (cs.CL)
备注: 88 pages, 4 figures, 5 tables. Includes Supplementary Information (40 pages)

点击查看摘要

Abstract:Machine interpreting (MI), or AI-enabled speech translation, is increasingly visible in professional and public-facing communication, yet its performance relative to human simultaneous interpreters remains insufficiently understood. Existing studies are often limited by a single discourse mode and evaluation frameworks that are not well aligned with interpreting as a discourse-level communicative activity. The present study compares five commercial MI systems with 26 human interpreters (14 professionals, 12 advanced students) on English-to-Chinese simultaneous interpreting across planned and unplanned discourse, focusing on meaning and target-language quality. In these materials, MI performance differed sharply between the two discourse modes. In the planned formal speech, the strongest system became highly competitive on meaning, entering the upper human range, and a second system also performed strongly. In the extemporaneous interview, however, no MI system approached the stronger human tiers. Code-level analysis further shows that similar overall scores often masked different error profiles. Human interpreters were more likely to preserve the broader communicative trajectory through selective omission, restructuring, and contextual inference, whereas even the best MI systems were more vulnerable to semantic distortion, discourse-tracking failures, and context-insensitive interpretation. On target-language quality, the human advantage proved more stable than on meaning. Unintelligible or nonsensical output and disfluency carryover emerge as particularly salient weaknesses in machine performance, especially in spontaneous discourse. Overall, the findings are consistent with a discourse-sensitive account of where MI appears competitive, where it remains limited, and why these distinctions matter for interpreting practice, interpreter education, and future MI development.

[NLP-107] Multi-Branch Transformer

【速读】: 该论文旨在解决大语言模型在参数量与训练计算资源受限条件下,如何实现更高效建模与推理的问题。传统方法依赖于增加模型宽度(width)或深度(depth)进行扩展,但存在边际收益递减和推理延迟上升的瓶颈。为此,论文提出一种互补的扩展路径——分支(branching),即并行组合完整的Transformer块。其核心解决方案是多分支Transformer(Multi-Branch Transformer, MBT),该结构通过共享初始层处理输入,将独立参数化的多个分支并行执行序列块,并在共享最终层前融合输出。在相同训练计算预算下,三种变体(宽浅型、窄深型、平衡型)中,平衡型MBT表现最优且随分支数增加仍保持高质量。为提升长上下文推理效率,论文进一步提出联合键值缓存(Joint-KV),通过在每一层深度上聚合各分支表示构建共享的键值缓存(KV cache)。最终方案结合平衡型MBT与Joint-KV,并辅以支持分支分组与共享缓存重用的推理引擎。实验表明,该方法在验证损失上相当于标准模型需1.26倍的训练计算量才能达到同等水平,同时可减少高达34%的全模型KV缓存存储,单张H200 GPU上解码吞吐量最高提升51.4%。消融分析证实,分支机制带来的性能增益不仅源于宽度或深度变化,还实现了有效的KV共享策略,对融合函数鲁棒性强,且在中间层表现最佳。研究结果表明,分支是一种兼具更高建模质量与更快推理速度的可行扩展路径。

链接: https://arxiv.org/abs/2611.19800
作者: Gaotang Li,Lingjie Chen,Chunyuan Deng,Zhiwei Xu,Hongzheng Chen,Christopher De Sa,Jingrui He,Yingheng Wang,Hanghang Tong
机构: 未知
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: Project: this https URL

点击查看摘要

Abstract:Transformer language models are commonly scaled by making their blocks wider or stacking them deeper. In this work, we explore a complementary path: branching, the parallel composition of full Transformer blocks. We introduce the Multi-Branch Transformer (MBT), which processes the input through shared initial layers, executes independently parameterized branches of sequential blocks in parallel, and fuses their outputs before shared final layers. At comparable training compute, we study three MBT variants: Wide-Shallow preserves block width, Narrow-Deep preserves sequential depth, and Balanced interpolates between them. In controlled pretraining experiments at four mixture-of-experts scales (0.8B to 2.1B total parameters), 2-branch Wide-Shallow MBTs consistently outperform sequential Transformers under matched parameter and training-compute budgets. Balanced MBTs improve further and retain comparable quality as the number of branches increases. To improve inference efficiency at long contexts, we introduce Joint-KV, which constructs a shared KV cache from branch representations at each depth. Our final recipe combines Balanced MBTs with Joint-KV, supported by an inference engine that exploits branch grouping and shared-cache reuse. Fitted scaling curves indicate that standard Transformers require up to 1.26x as much training body compute to match our final models’ validation loss. These models reduce full-model KV-cache storage by up to 34% and deliver up to 51.4% higher decoding throughput on a single H200 GPU. Ablation studies show that branching yields gains beyond changes in width and depth, enables an effective KV-sharing strategy (Joint-KV), remains robust across fusion functions, and performs best in the middle layers. These findings show that branching offers a practical path for scaling language models toward both better modeling quality and faster inference.

[NLP-108] Eliciting Research Taste in LLM s through Future Research Direction Choice

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在开放性科研任务中缺乏可衡量的“研究品味”(research taste)问题。在缺乏明确验证目标的探索性研究阶段,选择未来研究方向依赖于内在的研究偏好,而现有方法难以有效评估模型在此类判断中的表现。其解决方案的关键在于将研究品味操作化为科学工作流中的一个具体任务:基于给定的研究目标与当前结果,从两个候选方向中选择最合适的未来研究路径。该研究通过对比模型生成的方向与人类作者推荐的方向,提出三个行为指标——作者一致性(author alignment)、稳定分歧(stable divergence)和不稳定选择(unstable choice),以量化模型的研究偏好。实验基于19篇2025年12月至2026年2月间发表的机器学习论文,构建了80组成对比较,每组包含一个真实作者推荐方向与一个来自相似论文的合理干扰项。结果显示,八种主流大模型的作者一致性在45.0%至67.5%之间,稳定分歧达15.0%至36.2%,且该分歧在控制相似性、提示扰动及作者偏好框架后仍显著存在。进一步分析发现,模型普遍倾向于向规模扩展/泛化方向迁移,并在方法论改进与基础设施/数据扩展之间表现出系统性权衡。该研究首次提供了对大模型研究偏好进行描述性评估的框架,为未来理解并引导其在自主科学发现中扮演更关键角色奠定了基础。

链接: https://arxiv.org/abs/2611.19794
作者: Dhruv Trehan,Paras Chopra
机构: Lossfunk(损失函数)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: Accepted at the LM4Sci Workshop at COLM 2026. 23 pages, 7 figures, 18 tables. Code and data: this https URL

点击查看摘要

Abstract:Research taste is increasingly discussed in the context of AI scientists, yet remains little studied or measured in LLMs. This judgment matters especially in open-ended research, where a direction must be chosen before a verifiable objective is available. We operationalize research taste through one task of the scientific workflow: choosing which future research direction to pursue given a research goal and current results. Rather than predicting downstream scientometric metrics or treating author choices as objectively correct, we evaluate models against human-authored future-work directions and report author alignment, stable divergence, and unstable choice as behavioral measures of research-direction preference. We construct an order-controlled evaluation of 80 pairwise comparisons from 19 ML papers published between December 2025 and February 2026 across four subdomains. Each query presents a research goal, hypotheses tested with key results, and two candidate future directions: one author-recommended direction and one plausible distractor sampled from a similar paper. Across eight LLMs, author alignment ranges from 45.0% to 67.5%, while stable divergence ranges from 15.0% to 36.2%. These divergences largely persist under similarity controls, prompt ablations, and an author-preference framing check, and inter-model agreement suggests that some non-author preferences are shared across models. We characterize what models diverge toward using transition-category labels and selected reasoning examples, finding recurring movement toward scaling/generalization and trade-offs between methodological and infrastructure/data extensions. We view this as an early descriptive evaluation for eliciting research taste profiles in LLMs, and as a step toward understanding and steering their research preferences as they take on more autonomous roles in novel open-ended scientific discovery.

[NLP-109] Logical AI OS: A Training-Free Governance Layer Outside the Model for Instruction Following – Benefits Costs Failures and the Calibration of the Measuring Instrument Itself

【速读】: 该论文旨在解决小型语言模型(Small Language Models, SLMs)在响应用户显式请求时的合规性问题,包括但不限于输出长度、项目数量、格式要求、禁用或必需词汇、语法人称及回答语言等。其核心挑战在于模型生成内容常偏离用户明确指令,而现有方法缺乏对治理机制有效性与代价的系统性评估。解决方案的关键在于提出一种无需训练的外部治理层(training-free governance layer):该层通过规则解析用户请求,对模型输出进行机械式校验,若发现不合规则自动修复;修复失败时则返回原始答案以避免破坏性干预。该机制虽存在于前期研究中,但本文首次全面量化其收益、成本与失效模式,并详细报告了评估工具本身的校准过程。实验基于四款4位量化Qwen系列模型(3B、7B、14B及一个7B推理模型)和329个中文提示,结果显示,在经过修正的检查器处理后,完全符合所有请求的答案比例显著提升(所有对比p < 0.0001),其中修正检查器使符合率提高34%至66%,长度控制优化后进一步提升至52%至92%;而通用重试提示未带来显著改进。探索性分析显示,在90个触发修复的提示中,两名人工标注者认为治理层生成可接受答案的比例高于原始模型(49:7 和 18:12,差异不显著)。然而,该方法也伴随明显代价:生成时间增加43%-66%,且整句重写导致部分可用答案变为不可用。此外,规则读取能力受限于泛化性能,对未见表述的覆盖率分别达到约70%和84%(接近双标注者基准);对于不可满足请求,推理模型可能出现无限循环;模型自报告的修复状态可靠性差(25项中18项实际未修复)。最后,原始检查器对结构化请求的误判率超过一半,经修正后主要结论仍成立,但一项比较结果被下调。研究结论仅适用于所测试的中文数据集、Qwen家族模型及其最大14B规模。尽管治理实现未公开,但相关数据、生成结果、判断标签及检查器均已开放共享。

链接: https://arxiv.org/abs/2611.19767
作者: Joe Yuan
机构: 未知
类目: Computation and Language (cs.CL)
备注: 18 pages, 5 tables, 3 figures. Technical report v3. Earlier reports in this line: doi: https://doi.org/10.5281/zenodo.23226740 (Measurement Layer) and the Validation Layer report. Data, model answers, per-item verdicts, annotations and the standalone checker: this https URL (tag v3.0-data). The governance-layer implementation is not released

点击查看摘要

Abstract:Small language models often fail explicit user requests (length, number of items, format, required or forbidden words, grammatical person, answer language). We measure a training-free governance layer placed outside the model: it reads the requests with rules, checks the answer mechanically, repairs it when a request is not kept, and hands back the original answer when the repair fails. The loop exists in prior work; our contribution is measuring its benefits, costs and failures, and reporting how the measuring instrument itself was calibrated. On four 4-bit Qwen-family models (3B, 7B, 14B, one 7B reasoning model) and 329 Chinese prompts: (1) answers that keep every request the layer read rise significantly (all comparisons p0.0001), by 34 to 66 of 329 under the corrected checker and 52 to 92 after a length-control fix; a generic redo message brings no significant gain. (2) Exploratory: on 90 repair-triggered prompts, two AI annotators found the layer produced an acceptable answer where the bare model did not more often than the reverse (49:7; 18:12, not significant). (3) Costs: 43-66% more generation time; whole-answer rewriting makes a sizeable share of usable answers unusable. (4) Failures: rule-based reading reaches about 70% and 84% of a two-annotator ceiling on unseen phrasings; unsatisfiable requests make the reasoning model loop; model self-reports are unreliable (18 of 25 “fixed” items were not fixed). (5) Over half of the original checker’s “violated” verdicts on structural requests were wrong; after correction the main result holds, and one comparison is downgraded. Conclusions are limited to this Chinese material, the Qwen family and sizes up to 14B. The governance implementation is not released; data, answers, verdicts and the checker are public.

[NLP-110] Ambiguous Examples Help But Full Data Wins: Revisiting Dataset Cartography for Extractive Adversarial QA EACL2027

【速读】: 该论文旨在解决生成式模型在开放域提取式问答任务中,如何通过数据集地图(dataset cartography)技术筛选代表性子集以实现高效训练的问题。其核心挑战在于:尽管早期研究指出仅使用33%的模糊样本(ambiguous examples)即可在部分分布外(OOD)评估中超越全数据训练,但这一结论是否适用于更严格的开放式提取式问答场景尚不明确。论文的关键解决方案是基于干净源数据(clean-source SQuAD)的训练动态分析,对不同编码器家族构建训练数据的排序策略——包括随机、易学、模糊和难学样本子集,并在对抗性测试集(Adversarial-only AddSent 和 AddOneSent)上系统评估模型鲁棒性。结果表明,虽然模糊样本子集在多数情况下优于随机采样子集(14个评估单元中12个在F1指标上表现更优),但全数据训练仍显著优于模糊样本训练(14个单元中有12个在精确匹配和F1指标上胜出,平均差距分别为3.40 EM和4.10 F1)。因此,研究揭示了数据集地图在减少训练数据量方面的价值,但其在分布外泛化上的优势无法作为全数据训练的通用替代方案。

链接: https://arxiv.org/abs/2611.19763
作者: Paul K. Mandal
机构: Neurint LLC(Neurint LLC); U.S. Army Cyber Corps(美国陆军网络部队); U.S. Army Reserve(美国陆军预备役); Northwestern State University of Louisiana(路易斯安那西北大学)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 37 pages, 30 figures, 20 tables. Accepted to Findings of the Association for Computational Linguistics: EACL 2027. Code: this https URL

点击查看摘要

Abstract:Dataset cartography argues that training dynamics can reveal regions of a dataset that matter differently for learning: easy-to-learn examples aid optimization, hard-to-learn examples often expose annotation problems, and ambiguous examples contribute strongly to out-of-distribution generalization. A striking implication of the original work is that a model trained on only 33% ambiguous examples can outperform full-data training on several OOD evaluations, including a QNLI reformulation of SQuAD evaluated on Adversarial SQuAD. We revisit this claim in a stricter setting: open-ended extractive question answering, where models must predict answer spans rather than classify question-sentence pairs. For each encoder family, we compute clean-source SQuAD training dynamics, train on full, random, easy-ranked, ambiguous-ranked, and hard-ranked subsets, and evaluate robustness on adversarial-only AddSent and AddOneSent rows. Ambiguous-ranked 33% subsets beat random 33% subsets on exact-match accuracy in all 14 model-evaluation cells and on F1 in 12/14 cells. However, full-data training beats ambiguous-ranked training in 12/14 exact-match cells and 12/14 F1 cells, with mean full-minus-ambiguous gains of 3.40 EM and 4.10 F1. The most thoroughly replicated ELECTRA-small panel is also the main exception to the ambiguous-over-random F1 gains. These results show that clean-source cartography remains useful for reduced-data selection, but its OOD advantage does not carry over as a general replacement for full-data training in extractive adversarial QA.

[NLP-111] Where Should LLM Agents Adapt at Test Time? In Context in Weights or Both

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)智能体在重复交互同一环境(如代码仓库、数据库或应用套件)时,如何有效利用共享的潜在任务结构以提升泛化能力与适应效率的问题。其核心挑战在于:尽管智能体可通过上下文记忆(in-context learning, ICL)保留经验,或通过权重更新(如强化学习)进行学习,但这两类机制通常被孤立研究,缺乏对其协同效应的理解。本文的关键解决方案是系统性地比较并结合三种适应机制——基于可验证奖励的强化学习(GRPO)、自蒸馏(SDPO)以及ICL,并在统一在线协议和模块化框架下评估其性能。实验结果表明,ICL在独立使用时贡献最大,将跨环境平均成功率从28.8%提升至49.4%;进一步引入GRPO后,平均成功率提升至51.6%,且能有效恢复因ICL导致的检查与验证行为缺失,同时保持较低的环境查询次数。这说明可验证奖励的强化学习能够精细化调控上下文知识的使用方式,从而实现上下文管理与权重更新的互补。因此,该研究揭示了联合优化上下文保留与模型权重更新对提升智能体动态适应能力的关键作用。

链接: https://arxiv.org/abs/2611.19700
作者: Parsa Mirtaheri,Mohammadreza Armandpour,Keivan Alizadeh,Iman Mirzadeh,Parshin Shojaee,Aryo Lotfi,David Harrison,Mehrdad Farajtabar
机构: 未知
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:LLM agents operating repeatedly in the same environment, such as a code repository, a database, or an application suite, encounter tasks that share latent structure. Experience can reveal this structure, allowing agents to reuse environment knowledge and learn from environmental feedback. Agents can adapt at test-time using this experience through two channels: retaining it in context and updating their weights. These channels are usually studied separately, leaving unclear what each contributes and when they complement each other. We compare in-context learning (ICL), reinforcement learning with verifiable rewards using GRPO, and self-distillation using SDPO, alone and in combination, using a shared online protocol and modular framework. Agents encounter matched task streams, are evaluated on each task before adapting to it, and learn only from their own attempts and environment feedback. We evaluate 14 adaptation mechanisms and a stateless baseline with Qwen3.5-9B on 109 tasks from four environments drawn from SWE-bench Verified, AppWorld, ScienceWorld, and BIRD-Critic-SQLite, using five task orders per environment. ICL provides the largest standalone gain: its best policy raises mean success across environments from 28.8% to 49.4%. Adding GRPO to this policy achieves the highest mean success, 51.6%. While adding GRPO improves the overall mean under every ICL policy, SDPO has mixed effects. Trace analyses show fewer lookups and shorter attempts under ICL, consistent with reusing knowledge of the shared structure. However, ICL also frequently skips inspection and verification for the current task. Adding GRPO to ICL helps restore these behaviors while preserving its reduced environment lookups. These observations suggest that reinforcement learning with verifiable rewards can refine how agents use retained experience, motivating the joint study of context management and weight updates.

[NLP-112] Data Attribution Should Steer Pretraining NEURIPS2026

【速读】: 该论文旨在解决预训练数据筛选中“何为值得学习的文档”这一核心问题,其关键在于:传统数据筛选依赖静态质量信号,而这些信号在不同模型状态下的有效性可能不一致。为此,论文提出以可解释的数据属性(data property)作为可复用的筛选标准,通过训练数据归因(training-data attribution)识别出具有测试价值的属性,并在新数据上以低成本选择器进行验证,从而实现跨语料库和检查点的通用化数据筛选。与仅服务于单次训练运行的文档排序不同,被解释的数据属性具备跨场景可验证性,尤其在已知质量信号与实际模型性能不一致时更具优势。研究进一步构建反馈循环,区分历史贡献、可解释性与未来数据价值所需的不同证据类型,明确五类命题所需的证据及其局限性。最终,通过在相同总计算成本下对困惑度(perplexity)、分类器性能与混合基线进行可证伪的对比,验证了所提方法的有效性。

链接: https://arxiv.org/abs/2611.19698
作者: Glenn Matlin,Chandreyi Chakraborty,Mohammed Rehan Parwani,Saehee Eom,Anthony Wen-Ming Zang,Akhil Theerthala,Arjun Chatterjee,Tejas Khandwekar,Brandon Michaels,Alvin Deng,Mark Riedl
机构: Georgia Institute of Technology(佐治亚理工学院); MATS Program(多智能体系统与技术项目); Georgia Tech AI Safety Initiative(佐治亚理工学院人工智能安全倡议); University of Illinois Urbana-Champaign(伊利诺伊大学厄本那-香槟分校); EleutherAI(艾莱瑟人工智能); Independent(独立)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: Accepted to the 3rd Workshop on Attributing Model Behavior at Scale: Data Attribution and Provenance (ATTRIB 2026) at NeurIPS 2026, Idea Track. 4 pages main text, 13 pages total, 1 figure, 2 tables

点击查看摘要

Abstract:Pretraining data curation decides what a model learns from, yet what makes a document worth learning from depends on what the model already knows. Influence-driven selectors already rank documents to improve one run; we argue that training-data attribution should steer pretraining by identifying testable data properties that become reusable curation criteria, evaluated on new data against cheaper selectors at matched total cost, discovery included. A document ranking is consumed by the run it was computed for, whereas an explained data property can be tested across corpora and checkpoints, especially where familiar quality signals diverge from measured usefulness. Because historical contribution, an understandable explanation, and future data value require distinct evidence, we specify a feedback loop that interprets attributed sources, tests candidate data changes, and states for five claim types what evidence each requires and what none establishes. We close with a falsifiable comparison against perplexity, classifier, and mixing baselines at equal total compute.

[NLP-113] MathaMAGICal: Validating Trajectory Data Attribution During Capability Acquisition

【速读】: 该论文旨在解决生成式语言模型在训练过程中对特定训练数据贡献度评估的准确性问题,尤其关注基于影响函数(influence function)的方法如MAGIC在轨迹数据归因(trajectory data attribution)中的有效性。现有方法虽能提供精确的一阶归因分数,但其预测能力是否真实反映数据移除后的模型性能变化,仍需通过重新训练来验证。研究发现,在8600万和15200万参数的GPT-2风格模型上进行多步算术学习时,全训练过程的归因分析因数值溢出而不可行;即使采用元梯度裁剪,其归因分数与实际重训练结果的相关性接近随机水平(线性数据建模得分LDS为0.17–0.22)。相比之下,短检查点窗口内的归因分析表现出高度稳定性和准确性(在15200万参数模型上,四个100步窗口的LDS达0.86–0.99),显著优于单检查点梯度点积基线(相关性接近零)。此外,个体训练块的排名仅在单一训练运行内有效,无法跨种子或超参数迁移(ρ ≤ 0.09),而格式层面的模式则具有跨运行可复现性。值得注意的是,两类数据——解题示例与基础事实——在有益与有害的前5%归因分数尾部中占比达到其语料库份额的2.1–2.6倍,表明其具有强烈的双向局部敏感性,这种特性不同于多轮删除实验所揭示的净收益。因此,该研究的关键解决方案在于:通过短时间窗口内的动态验证机制,实现对轨迹归因结果的可靠性保障,并从中提取可泛化的、层级化(家族级)的数据影响模式,从而突破传统方法在长期训练轨迹中稳定性差、泛化性弱的瓶颈。

链接: https://arxiv.org/abs/2611.19692
作者: Jason Lee,Dikshanta Pokhrel,Phonn Nay Khant,Victor Lee,Moritz Weckbecker,Sebastian Lapuschkin,Wojciech Samek,Mark Riedl,Glenn Matlin
机构: Algoverse AI Research; University of California, Los Angeles; Fraunhofer Heinrich Hertz Institute; Technische Universität Berlin; Georgia Institute of Technology
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Influence-function-based methods like MAGIC provide exact first-order scores for trajectory data attribution, yet only retraining indicates whether those scores accurately predict data removal. We evaluate MAGIC on 86 M- and 152 M-parameter GPT-2-style models learning multistep arithmetic from scratch on filtered web text and four synthetic math families. Full-run attribution at 86 M suffers from numerical overflow, and even with metagradient clipping, scores correlate near chance with retraining (linear datamodeling score, LDS, 0.17 – 0.22 ). In contrast, short checkpoint windows yield stable, accurate predictions at both scales (LDS 0.86 – 0.99 across four 100 -step windows at 152 M), whereas a single-checkpoint gradient-dot baseline shows near-zero correlation. Rankings of individual training blocks predict retraining only within that single run and do not transfer across seeds or hyperparameters ( \rho \le 0.09 ), whereas format-level patterns replicate across runs. Two data families, worked examples and primitive facts, occupy 2.1 – 2.6\times their corpus share in both the helpful and harmful 5% score tails, reflecting strong bidirectional local sensitivity distinct from net benefit under multi-epoch family deletion. These results show that trajectory attribution is reliable within short, validated windows and yields family-level insights that generalize across runs.

[NLP-114] SG-LegalClarify: A Large-Scale Benchmark for Legal Reasoning and Ambiguity Recognition in Singapore Law

【速读】: 该论文旨在解决生成式 AI 在新加坡法律语境下进行法律推理时面临的两大核心挑战:一是模型需在继承自英国普通法的通用法律原则与新加坡本地特定法规之间做出准确区分,二是模型难以识别非专业用户提出的法律问题中隐含的事实不明确性(underspecification),而此类不明确性直接影响法律结论的准确性。现有法律基准测试普遍缺乏对新加坡法律体系的覆盖,且未充分评估模型对语义模糊性的识别能力。为此,研究提出SG-LegalClarify——一个大规模、面向开放域法律推理与模糊性识别的基准数据集,涵盖10个法律领域共1,100道由律师撰写的源问题,其中包含340个完全指定问题和760个未完全指定问题,每道未完全指定问题均配有固定隐藏事实及对应的完全指定版本,共计1,860个评估实例。所有答案采用律师制定的Issue, Rule, Application, and Conclusion (IRAC) 评分标准进行评判。研究通过多轮对话式评估流程,让模型判断是否需要澄清,并借助用户模拟器仅在请求时揭示隐藏事实,以检验模型的主动澄清能力。实验结果表明,当前主流大语言模型(LLM)在闭卷条件下仅能达成平均53.8%(最佳模型达69.8%)的强制性评分项覆盖率,在启用网络搜索后平均提升至63.0%(最佳达77.8%),但仍存在显著改进空间,且多数模型未能有效识别何时应发起澄清。该研究的关键解决方案在于构建具有真实法律复杂性的结构化评估框架,结合可验证的隐藏事实机制与基于真实律师标准的精细化评分体系,从而系统性地推动生成式AI在跨司法管辖区法律推理中的鲁棒性与认知能力发展。

链接: https://arxiv.org/abs/2611.19643
作者: Ziwei Gu,Michael S. Lee,Drew Rein,Yash Maurya,Vikneswaran Kumaran,Darshini Ramiah,Chenguang Wang,Udari Madhushani Sehwag
机构: Human Frontier Collective, Scale AI; Scale AI; Preparedness and Policy, Scale AI; Infocomm Media Development Authority (IMDA); University of California, Santa Cruz; Harvard University
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Singapore offers a distinctive testbed for legal reasoning by large language models (LLMs). Its legal system combines inherited English common law with jurisdiction-specific statutes, requiring models to distinguish local doctrine from generic common-law intuition. This setting also highlights the challenge of recognizing underspecification in lay legal queries, because the facts material to a legal outcome depend on the governing doctrine. Yet legal benchmarks rarely cover Singapore law or test this form of ambiguity recognition. To address these gaps, we introduce SG-LegalClarify, a large-scale benchmark for open-ended legal reasoning and ambiguity recognition in Singapore law. SG-LegalClarify contains 1,100 lawyer-authored source questions spanning 10 legal domains, including 340 fully specified and 760 underspecified questions. Each underspecified question has a fixed hidden-fact assignment and a fully specified counterpart, yielding 1,860 evaluation instances. Responses are graded with lawyer-authored Issue, Rule, Application, and Conclusion (IRAC) rubrics. We evaluate 19 LLMs, closed-book and with web search, using a multi-turn pipeline in which models decide whether to answer or clarify, and a user simulator reveals hidden facts only when asked. We find the benchmark far from saturated: the best models satisfy only 69.8% of mandatory rubric items closed-book (average 53.8%) and 77.8% with web search (average 63.0%). Models also struggle to recognize when clarification is necessary. We will release the public split of our benchmark, the model outputs on it, and the evaluation pipeline to support future research.

[NLP-115] ROUNDS: Source-Framed Diagnostic Revision after Multi-Turn Clinical Evidence ALT ML4H

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在医学诊断对话中从多轮交互证据中形成可靠诊断的能力问题,特别是当用户不提供选项、引入外部意见或逐步披露临床信息时,模型是否能维持准确判断。其解决方案的关键在于提出一种名为ROUNDS的评估框架,该框架通过固定轮次的多轮临床证据输入,系统性地考察模型在不同情境下的诊断稳定性与鲁棒性:包括对比有无答案选项的情况、以医生或非专业人员语境呈现误导性建议的影响,以及由五名临床专家组成的书面意见小组对竞争性诊断的支持程度变化。实验结果表明,模型在有选项时平均准确率为76.9%,而无选项时下降至57.0%;且医生语境下的错误建议导致的最终准确率下降幅度显著高于普通语境(差异达4.6至38.2个百分点),多数原本正确的诊断因采纳建议而被更正,而非选择不回答;同时,正确诊断获得的支持越多,模型准确性越高,即使多数意见未变亦然。这揭示了当前医疗基准测试表现与真实对话中可靠诊断能力之间存在显著差距:模型易受外部意见影响,即便缺乏新临床证据也倾向于放弃正确诊断,暴露出其推理与决策机制的脆弱性。

链接: https://arxiv.org/abs/2611.19636
作者: Jinrui Fang,Runhan Chen,Yining Wu,Jiawei Xu,Xu Yang,Jian Yu,Wenqi Shi,Ying Ding,Yuji Zhang
机构: The University of Texas at Austin (德克萨斯大学奥斯汀分校); The University of Texas Southwestern Medical Center (德克萨斯西南医学中心); Washington University in St. Louis (圣路易斯华盛顿大学); City University of New York (纽约城市大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: Accepted at Machine Learning for Health (ML4H) 2026, Proceedings Track; to appear in Proceedings of Machine Learning Research (PMLR). 17 pages

点击查看摘要

Abstract:Large language models (LLMs) achieve high scores on medical benchmarks, but does this performance translate into reliable diagnosis as a conversation unfolds? In diagnostic conversations, users can provide clinical findings across multiple turns, ask for a diagnosis without supplying answer options, and introduce opinions from others. Models must therefore form diagnoses from the available evidence and decide how to respond to subsequent advice. We introduce ROUNDS, a framework that evaluates these abilities after a fixed sequence of clinical evidence delivered over multiple turns. ROUNDS compares diagnosis with and without answer options, presents the same incorrect diagnosis under clinician and lay framing, and varies support for competing diagnoses in written five-clinician panels. Across seven models and 1,035 cases from five datasets, average accuracy was 76.9% with answer options and 57.0% without them. Clinician-framed misleading advice produced lower final accuracy than lay-framed advice in every model, with differences of 4.6 to 38.2 percentage points. Initially correct diagnoses more often became incorrect answers than abstentions, and many adopted the suggested disease. Conversely, accuracy increased as more panel members supported the correct diagnosis, including when the majority opinion remained unchanged. These findings reveal a gap between medical benchmark performance and reliable diagnosis in conversation: models respond to external opinions, but remain vulnerable to abandoning correct diagnoses without new clinical evidence.

[NLP-116] SemMask: Semantic-Prioritized Progressive Masking for Diffusion Language Models

【速读】: 该论文旨在解决扩散语言模型(Diffusion Language Models, DLMs)在实际应用中表现出伪自回归行为(pseudo-autoregressive behavior)和功能词优先偏差(function-word-first bias)的问题,这些问题限制了其在文本写作与代码生成等需全局规划与并行解码任务中的性能。其解决方案的关键在于提出一种名为SemMask的新微调范式,该范式通过语义优先的渐进式掩码策略,引导模型优先解码高信息量的语义标记(semantic tokens),从而强化模型的全局规划能力与并行生成能力。具体而言,该方法首先利用原始DLM对每个标记的梯度来识别语义标记,在掩码过程中优先保留这些语义标记,形成语义标记先于功能标记解码的渐进生成顺序;同时,通过调整语义标记的损失权重,进一步强化其早期预测。实验结果表明,SemMask在文本生成与代码生成任务上达到与强自回归基线相当的质量,且具备更优的语义连贯性;更重要的是,其在显著减少解码步数的前提下实现了更高的准确性,证明了早期建立语义结构能有效提升效率-准确率权衡。

链接: https://arxiv.org/abs/2611.19560
作者: Qi Chen,Zifeng Cheng,Zhiwei Jiang,Qing Gu
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Diffusion Language Models (DLMs) are designed to enable globally parallel generation. However, in practice, they often exhibit pseudo-autoregressive behavior and a function-word-first bias, which undermines their performance on tasks such as text writing and code generation that benefit from global planning and parallel decoding. To address this, we propose SemMask, a novel fine-tuning paradigm that leverages a semantic-prioritized progressive masking strategy to encourage the model to decode high-information semantic tokens early and globally, thereby enhancing its global planning and parallel generation capabilities. Specifically, we first use the gradients of the original DLM with respect to each token to identify semantic tokens. During masking, these semantic tokens are preserved with priority, establishing a progressive generation order in which semantic tokens are decoded prior to function tokens. Meanwhile, we adjust the loss weight for semantic tokens to further reinforce their early prediction. Experimental results show that SemMask achieves quality comparable to strong autoregressive baselines on both text writing and code generation tasks, while providing superior semantic coherence. Notably, SemMask achieves higher accuracy with substantially fewer decoding steps, indicating that the early establishment of semantic structure improves the efficiency–accuracy trade-off.

[NLP-117] Objectify Dont Delete: Complaint-Derived Evaluation of Ambient Clinical Notes ALT NEURIPS2026

【速读】: 该论文旨在解决当前生成式临床记录(Generative AI-generated clinical notes)质量评估中存在的重要问题:现有评估方法无法准确反映临床医生实际关注的缺陷。传统指标如ROUGE仅衡量与参考文本的词汇相似性,而专家检查清单则基于委员会共识,二者均未能捕捉到临床实践中高频出现的真实反馈。研究发现,尽管可通过用户反馈挖掘评估标准,并筛选出大语言模型(LLM)可可靠执行的标准,但真正被临床医生频繁提及的关键问题往往因表述为复合型、多问合一的问题而难以被单一“是/否”判断有效识别。为此,论文提出一种“对象化”(objectification)步骤,将模糊、复合的评价标准重构为一系列具有明确范围和确定性的判别决策。通过构建一个端到端的评估流程——从投诉中提取标准、进行对象化处理,并部署LLM裁判对人工植入缺陷及真实生产数据进行评估——验证了该方法的有效性。在ACI-bench参考文本中植入11类缺陷后,经对象化处理的裁判实现中位召回率达0.95,而传统指标变化不超过0.037;若仍以原始复合问题形式提问,则虽能检测所有缺陷,却会误报70%未受损参考文本;当每次调用仅提出一个段落-片段对应的问题时,Youden’s J值从0.30提升至0.72(匹配提示工程投入后分别提升至0.19和0.26)。研究还发布了ACI-Defect基准数据集,包含经临床医生审校的实例级标签,揭示了现有金标准中存在显著质量问题(40份笔记中有32份含幻觉,40份均遗漏讨论事实),以及基于证据的裁判结论。在1,446条真实生产记录上,每一轮临床评分的平均任务负担随评分次数下降(rho = -0.15,关联性微弱),表明缺陷需按语义跨度独立计数,而非简化为每份记录一个“通过/失败”二元判定。因此,解决方案的核心在于通过对象化重构评价标准,使评估体系能够精准定位复杂临床关切,从而提升评估的可解释性与实用性。

链接: https://arxiv.org/abs/2611.19525
作者: Nikita Gupta,Narayanan Unny Edakunni,Vikram Okaly,Suranjana Datta,Navya Penubadi,Karl Eric Swanson,Siddharth Sharma
机构: Suki AI(苏基人工智能), California, USA
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: Accepted at the Third Workshop on GenAI for Health (GenAI4Health), NeurIPS 2026 (non-archival). 16 pages, 5 figures, 11 tables. Code and data: this https URL

点击查看摘要

Abstract:Evaluating the quality of AI-generated clinical notes is hard: standard metrics like ROUGE measure lexical similarity to reference notes, while expert checklists encode what committees believe matters. Neither captures what clinicians actually complain about. Recent work has shown that evaluation criteria can be mined from user feedback and filtered to criteria that LLM judges can reliably enforce. We argue that this is where much of the value is lost. The criteria that fail enforceability are disproportionately the ones clinicians raise most often, because high-frequency complaints are stated as several questions glued together. No yes/no judgment can grade such composites as posed. We introduce an objectification step that rewrites such criteria into sets of scoped, determinate decisions. We demonstrate the approach through an evaluation pipeline that mines criteria from complaints, objectifies them, and deploys LLM judges against both planted defects and production notes. Across eleven defects planted into ACI-bench reference notes, objectified judges localize violations with median recall of 0.95, while reference metrics change by at most 0.037. Given the complaint as one broad question, the same judges still find every planted defect, but they also flag 70% of the unperturbed reference notes; one section-pair question per call raises Youden’s J from 0.30 to 0.72 (by 0.19 to 0.26 with prompt effort matched). We release ACI-Defect, a perturbation benchmark; clinician-reviewed instance-level labels showing quality gaps in the ACI-bench gold standard (32 of 40 notes contain hallucinations; 40 of 40 omit discussed facts); and grounding-judge verdicts. On 1,446 production notes, mean instance-rate burden falls at every clinician-rating step (rho=-0.15), a small association. Defect counts have to be kept per span rather than collapsed into one pass/fail bit per note.

[NLP-118] Sada-ASR: A Production Multidialectal Arabic Speech Recognition System

【速读】: 该论文旨在解决阿拉伯语客服语音识别中方言与标准语(Modern Standard Arabic, MSA)不匹配的核心问题:用户在拨打阿拉伯语客服热线时使用的是本地方言,而现有的语音识别系统主要基于MSA进行训练和评估,导致方言识别错误率高,进而引发通话中断等严重后果。其解决方案的关键在于构建并部署Sada-ASR——EchoNova生产环境中实际使用的阿拉伯语语音识别系统,通过针对八种主要方言及MSA的混合数据集进行优化,实现了在120小时测试集上最低的词错误率(WER 11.69%)和字符错误率,显著优于包括Microsoft MAI-Transcribe-2在内的十一个主流系统。其核心技术优势体现在对阿曼、埃及、沙特等方言的强鲁棒性,尤其在非固定数量与脚本规范影响的片段上表现突出;同时,研究团队公开了部分测试数据与评分代码,以推动领域内可复现性与基准统一。

链接: https://arxiv.org/abs/2611.19512
作者: Mahmoud Khaled,Mohammed Abd-Elnaser,Mahmoud Mohsen,Hazem Ahmed,Abdelrahman Hafez
机构: EchoNova AI(EchoNova AI), Dubai, United Arab Emirates
类目: Computation and Language (cs.CL); Sound (cs.SD); Audio and Speech Processing (eess.AS)
备注: 10 pages, 3 figures, 8 tables. Code: this https URL Dataset: this https URL

点击查看摘要

Abstract:Callers to Arabic customer-service lines speak their own dialect, while speech recognizers are built and evaluated mainly on Modern Standard Arabic (MSA). A recognizer that mishears the dialect breaks the call before anything else can help. We present Sada-ASR, the speech recognizer of EchoNova’s production Arabic voice agent. On a 120-hour test set of MSA and eight dialects, it has the lowest word and character error rates of twelve systems, against eleven from Microsoft, ElevenLabs, Google, OpenAI, Speechmatics, IFM, Cohere, NVIDIA and Audar: 11.69% word error rate against 13.51% for the strongest, Microsoft’s MAI-Transcribe-2 (a lead of 0.74 points on clips unaffected by that system’s fixed number and script conventions), and 15.52% against 16.96% averaged over varieties. Its largest margins are on Omani (4.30 points), Egyptian (3.78) and Saudi (2.19) speech; it is level on Syrian and loses on Lebanese (24.96% against 23.05%), mostly because it writes Lebanese words in MSA forms and is less verbatim. On six public test sets its average word error rate under the leaderboard’s current scoring is 17.17%, below every published figure we found, among them Falcon-ASR (20.92%), Jais-ASR (22.85%) and Audar-ASR-V1-Turbo (23.17%), though none of these states that it used that scoring, so the comparison is indicative. We describe the data pipeline and methodology behind Sada-ASR and release a 20-hour subset of the test set with our scoring code.

[NLP-119] Formalizing Formal Semantics: The Formal Atlas

【速读】: 该论文旨在解决自然语言语义研究中理论形式化不足与可验证性差的问题,通过构建一个基于证明助手(proof assistant)的系统——形式化语义图谱(Formal Atlas),实现对语义理论的精确形式化、预测检验及理论间的对比分析。其核心解决方案在于建立一个交互式应用平台,将原始语义分析记录与形式化推导过程相连接,使研究人员能够直观检查证明细节并动态修改支持的Rocq文件,从而提升理论开发的透明度与可复现性。该系统以蒙塔古的指称语义学(PTQ)和加林的形式语义体系为主要案例,进一步拓展至动态语义、类型论、理性语用推理(RSA)及分布语义等多元理论的比较,并探讨了如何形式化那些表达不够明确的语义分析。此外,论文还探索了生成式 AI (Generative AI) 在辅助资源构建与推理系统中的潜在应用,推动形式化语义研究在科研与教学中的双重价值实现。

链接: https://arxiv.org/abs/2611.19509
作者: Stergios Chatzikyriakidis,Aarne Talman
机构: 未知
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:In this paper we present the Formal Atlas, a system for the study of natural language semantics using proof assistants. Our aim is to formalize semantic accounts, check their predictions and compare the implemented theories. To this end, we connect source records and formal developments with an interactive application, where researchers can inspect proofs and modify supported Rocq files. Montague’s PTQ and Gallin provide our main cases, followed by dynamic and type-theoretic comparisons, RSA and distributional semantics. We also discuss what it takes to formalize less explicit analyses. Lastly, we consider the use of AI, both for assistance in developing these resources and for inference systems that can make use of them. This work extends an earlier programme in proof-assistant semantics, with research and teaching as complementary uses. Subjects: Computation and Language (cs.CL) Cite as: arXiv:2611.19509 [cs.CL] (or arXiv:2611.19509v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.19509 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Stergios Chatzikyriakidis [view email] [v1] Sat, 14 Nov 2026 17:02:32 UTC (746 KB) Full-text links: Access Paper: View a PDF of the paper titled Formalizing Formal Semantics: The Formal Atlas, by Stergios Chatzikyriakidis and Aarne TalmanView PDFHTML (experimental)TeX Source view license Additional Features Audio Summary Current browse context: cs.CL prev | next new | recent | 2026-11 Change to browse by: cs References Citations SciXGoogle Scholar Semantic Scholar export BibTeX citation Loading… BibTeX formatted citation loading… Data provided by: Bookmark checked="checked"class=“labs-tab-input”> Bibliographic Tools Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer (What is the Explorer?) Connected Papers Toggle Connected Papers (What is Connected Papers?) Litmaps Toggle Litmaps (What is Litmaps?) scite.ai Toggle scite Smart Citations (What are Smart Citations?) Code, Data, Media Code, Data and Media Associated with this Article alphaXiv Toggle alphaXiv (What is alphaXiv?) Links to Code Toggle CatalyzeX Code Finder for Papers (What is CatalyzeX?) DagsHub Toggle DagsHub (What is DagsHub?) GotitPub Toggle Gotit.pub (What is GotitPub?) Huggingface Toggle Hugging Face (What is Huggingface?) ScienceCast Toggle ScienceCast (What is ScienceCast?) Demos Demos Replicate Toggle Replicate (What is Replicate?) Spaces Toggle Hugging Face Spaces (What is Spaces?) Spaces Toggle TXYZ.AI (What is TXYZ.AI?) Related Papers Recommenders and Search Tools Link to Influence Flower Influence Flower (What are Influence Flowers?) Core recommender toggle CORE Recommender (What is CORE?) Author Venue Institution Topic About arXivLabs arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv’s community? Learn more about arXivLabs. Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?) mathjaxToggle(); We gratefully acknowledge support from our major funders, member institutions, , and all contributors. About Help Contact Subscribe Copyright Privacy Accessibility Operational Status (opens in new tab)

[NLP-120] Evaluating and Improving Natural Language Autoencoders

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)内部激活状态的可解释性问题,具体聚焦于如何通过自然语言自编码器(Natural Language Autoencoders, NLAs)生成对模型内部表征的自然语言解释。其核心挑战在于评估和提升NLAs在揭示模型潜在知识、识别模型是否被误导以及检测模型沉默式幻觉(silent fabrication)等方面的能力。解决方案的关键在于提出一系列针对审计任务的评估方法,并引入NLAs的扩展机制:包括整合多层与多位置的激活信息,以增强对长文本序列的建模能力,同时降低虚假陈述的数量。这些改进显著提升了NLAs在审计评估及已有辅助评估任务上的表现。此外,研究还揭示了不同审计能力随NLA训练进程的变化规律,发现尽管早期训练中各项指标普遍提升,但过度训练可能导致部分审计性能下降,提示需谨慎权衡训练时长与模型可靠性之间的关系。

链接: https://arxiv.org/abs/2611.19507
作者: Marco Bazzani,Dan Mossing
机构: Anthropic(Anthropic)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 23 pages, 18 figures, 2 tables

点击查看摘要

Abstract:Natural language autoencoders (NLAs) are a method for generating natural language explanations of LLM activations. We develop evaluations based on auditing tasks that test the ability of NLAs to surface latent knowledge, identify whether a model was misled by a hint, and catch a silent fabrication. We also develop extensions to NLAs that let them use activation information from multiple layers and positions, let us apply them more effectively to long transcripts, and reduce the number of false claims they make. These extensions result in improvements on both our auditing evaluations and auxiliary evaluations adapted from prior work. We also use the evaluations to study how different auditing capabilities change over the course of NLA training, and find that although every evaluation improves early in training, extended training can hurt performance on some of our auditing evaluations.

[NLP-121] Do Probing and Steering Favor the Same Directions Within and Across Language Models? NEURIPS2026

【速读】: 该论文旨在解决生成式 AI 模型中概念探测(probing)与概念操控(steering)之间方向选择不一致的问题,即:能够最优检测某一概念的隐藏状态方向,并不等同于能最优实现该概念的可控调整。其核心发现是,在三种不同架构的指令微调语言模型中,用于探测真实性和情感倾向的最佳方向(最佳探测方向)通常并非用于添加式操控(additive steering)时效果最佳的方向;相反,更接近均值差(mean difference)的方向在操控任务中表现更优。研究通过构建从协方差调整判别方向到均值差的一维方向族,系统评估了探测与操控性能在固定副作用下的变化,结果表明在59/60的层-模型-概念组合中,探测与操控的最优方向存在显著分离。此外,这种分离现象在不同模型间具有不对称转移性:探测方向可通过线性映射在中性文本上有效迁移(保留81–91%的探测性能),而均值差方向的操控效果则难以跨模型有效传递,仅在少数模型对中表现出可转移性。作为应用实例,针对PopQA数据集的“未知”概念进行操控可使模型在不确定错误上的回避率提升18–44个百分点,但在可回答的错误中极少纠正,验证了操控方向选择对模型行为干预效果的关键影响。

链接: https://arxiv.org/abs/2611.19499
作者: Afshin Sadeghi,Daniel Becker
机构: AI Assurance and Assessments (AAA); Fraunhofer IAIS (弗劳恩霍夫信息与通信技术研究所), Sankt-Augustin, Germany
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Accepted at the Workshop on Unifying Representations in Neural Models (UniReps) co-located with NeurIPS 2026.(arXiv preprint)

点击查看摘要

Abstract:A direction in a language model’s hidden state can be used to probe whether a concept is present or to steer the model toward it. Although probe directions are often reused for steering, the direction that best detects a concept need not be the one that best changes it. We study this distinction in three instruction-tuned models from different families. We construct a one-parameter family of directions ranging from a covariance-adjusted discriminant to the mean difference, and evaluate probing and additive steering, which adds the direction to the hidden state, along the family at a fixed side effect. For truthfulness and sentiment, in every model, the best probe is not the best steering direction: steering consistently favors directions closer to the mean difference. This separation holds across 59 of 60 layer-model-concept combinations in a layer sweep, with matched probe and steering texts in all six model-concept combinations, and under a rotation-based steering operator. The exact steering optimum is less stable: for sentiment, an intermediate direction produces a 27-56% stronger steering effect than the mean difference, but this gain shrinks or disappears with matched texts and depends on the layer and the operator. The two roles also transfer differently across models. A probe transferred by a linear map fitted on neutral text retains 81-91% of the target model’s above-chance probing performance, while the transferred mean difference retains less of its steering effect in 11 of 12 model pairs. This is a consistent tendency, although our predefined criteria classify only three pairs, and nine remain inconclusive. Finally, as an application on PopQA, steering toward “unknown” increases abstention by 18-44 percentage points on uncertain errors, versus 2-5 on confident errors, while rarely correcting errors in cases where a correct answer is reachable. Comments: Accepted at the Workshop on Unifying Representations in Neural Models (UniReps) co-located with NeurIPS 2026.(arXiv preprint) Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG) Cite as: arXiv:2611.19499 [cs.CL] (or arXiv:2611.19499v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2611.19499 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[NLP-122] xt Speeding Itself Up: Recency and Restraint in Session-Local Drafting

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)在生成文本时的推理效率问题,特别是如何通过轻量级草稿模型(draft model)加速生成过程,同时保持输出质量。其核心挑战在于:在保证验证阶段(verification)高效的前提下,使草稿生成过程尽可能低成本。解决方案的关键在于引入会话局部草稿机制(session-local drafters),该机制基于提示词(prompt)和已生成文本进行上下文预测,无需额外训练即可实现近乎零成本的草稿生成。研究对比了基于最近性加权、退避策略与置信度门控的会话记忆方法,与传统的最早匹配(earliest-match)和最新匹配(most-recent-match)提示查找解码方法。结果表明,在生成速度方面,采用最近性优先的匹配策略可带来最高达0.09的速度提升(相较于最早匹配),而权重调整与退避机制在此基础上未带来显著增益;在能耗方面,选择性草稿(仅在强证据下生成)相比无筛选草稿能节省3–4焦耳/100个token,且三令牌最小匹配长度的提示查找解码性能接近置信度门控的会话记忆。此外,94%的生成结果与贪婪解码完全一致,所有偏差均出现在数值近似相等的决策点上,说明该方法具有高度的稳定性与准确性。在重复惩罚条件下,长序列生成速度提升下降至1.1–1.2倍,但近期性优势依然存在,且唯一能在医疗与法律两类文本中同时实现能耗节约的方案是设置最小匹配长度的草稿机制。

链接: https://arxiv.org/abs/2611.19455
作者: Antal van den Bosch
机构: Utrecht University (乌得勒支大学)
类目: Computation and Language (cs.CL)
备注: 10 pages. Code: this https URL

点击查看摘要

Abstract:Speculative decoding lets a cheap draft model propose tokens that a large language model verifies in a single forward pass. When the verifier keeps a KV cache, verification is cheap and drafting must be nearly free. Session-local drafters, which predict from the prompt and the text generated so far, meet that bar without training. We compare a recency-weighted, backing-off, confidence-gated session memory against prompt lookup decoding in its reference form, which copies from the earliest earlier match, and in a variant that copies from the most recent one, with paired per-prompt bootstrap intervals throughout. On general, medical, and legal text with a Mistral-7B verifier, recency accounts for the speed: most-recent lookup gains up to 0.09 in speedup over earliest-match lookup on 500-token generations, and session memory’s weighting and backoff add nothing measurable on top, at 30 to 40 times the drafting cost. Energy tells a different story. On medical and legal text, unselective drafters are 14% faster than plain decoding yet use as much energy per token, while drafters that abstain on weak evidence save 3 to 4 joules per 100 tokens; a three-token minimum match length on prompt lookup matches the confidence-gated session memory. Outputs are token-identical to greedy decoding in 94% of generations, with every divergence at a numerical near-tie. Under a repetition penalty, long-generation speedups shrink from 1.4-1.7x to 1.1-1.2x, but recency still helps and the length floor is the only drafter that saves energy in both domains.

[NLP-123] Do Rule-Based Tools Help LLM s Translate Endangered Uralic Languages in the Prompt or over MCP?

【速读】: 该论文旨在解决低资源乌拉尔语系语言(科米-兹良语、莫克沙语、埃里亚语和乌德穆尔特语)到英语的机器翻译中,如何有效利用现有成熟规则型语言资源(如有限状态分析器、约束语法和双语词典)以提升翻译质量的问题。其核心挑战在于:如何将这些结构化语言知识以最优方式注入生成式 AI 模型,从而改善翻译准确性。解决方案的关键在于对比两种知识接入机制——一是将预计算的语言分析结果以提示词(prompt)形式直接嵌入输入,二是通过模型上下文协议(Model Context Protocol, MCP)让模型动态调用外部工具。实验结果表明,提示词注入显著提升了翻译接受率(从13增至43/272),而工具调用虽进一步提升至55/272,但实现难度远高于前者。关键发现是,性能提升主要源于双语词典提供的词汇映射,且仅当词典包含英文释义时才有效;对于无英文侧信息的乌德穆尔特语,提示词注入无增益。此外,多数开放权重模型难以正确执行工具调用,其中14个模型从未成功调用,11个因无法生成有效调用指令,且半数调用请求指向不支持的语言。研究强调了结构化语言资源在增强低资源语言翻译中的价值,同时揭示了当前生成式 AI 在实际调用外部工具方面仍存在严重技术瓶颈。

链接: https://arxiv.org/abs/2611.19439
作者: Iaroslav Chelombitko,Mika Hämäläinen
机构: Metropolia University of Applied Sciences(赫尔辛基应用科学大学); Neapolis University Pafos(尼阿波利斯大学帕福斯分校)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: Accepted at the 11th International Workshop on Computational Linguistics for Uralic Languages (IWCLUL 2026), Munich. 9 pages plus appendices, 1 figure, 10 tables. Data and all model outputs: this https URL

点击查看摘要

Abstract:Komi-Zyrian, Moksha, Erzya and Udmurt have mature rule-based resources, among them finite-state analysers, constraint grammars and bilingual dictionaries, which language models do not use. We ask whether giving a model access to these resources improves machine translation from the four languages into English, and whether it matters how the access is provided: as analyses computed in advance and written into the prompt, or as tools that the model calls itself through the Model Context Protocol (MCP). We evaluate manually: an expert linguist read every English output of the strongest model under three conditions, 816 sentences in all, accepted or rejected each one, and named the main error in every rejected translation. With the analyses in the prompt, the number of accepted translations rises from 13 to 43 out of 272; with the tools attached it rises to 55. The gain comes from the bilingual dictionary. For Udmurt, whose dictionary has no English side, the prompt-based condition brings no gain at all, while for the other three languages an English gloss reaches the prompt for 74 to 80% of the source words. Tool calling is also much harder to get working than prompt injection. Of 31 open-weight models, 14 never call a tool, 11 of them because they cannot produce a usable call, and half of the calls that are made name a language the toolkit does not support. We release the data set, every model output and all tool-call logs.

[NLP-124] Do Pretraining Optimizer Gains Transfer to KL Distillation?

【速读】: 该论文旨在解决矩阵结构优化器(matrix-structured optimizers)在语言模型预训练中的性能提升是否可迁移至从指令微调初始化出发的短时域KL散度蒸馏(short-horizon KL distillation)任务。其核心问题是:现有在预训练阶段表现优异的矩阵结构优化器(如Shampoo系列方法)在特定蒸馏目标下是否仍具备优势,以及其性能与训练时间之间的权衡是否优于传统的逐元素优化器(如AdamW和AdaGrad)。解决方案的关键在于系统性地对比多种矩阵结构优化器(包括SOAP、All-2D Muon)与基准优化器(AdamW、对角线AdaGrad)在四组密集型Qwen师生模型对上的表现,基于统一的前向KL目标函数和固定训练预算进行评估。结果表明,预训练中获得的优化器优势并不能自动迁移至蒸馏任务;不同优化器在不同模型对上的相对表现存在显著差异,且部分矩阵结构方法虽在损失(KL)上接近AdamW,但引入了显著的训练器时间开销。因此,结论强调必须针对具体目标函数和系统环境重新验证优化器的有效性,而非直接沿用预训练经验。

链接: https://arxiv.org/abs/2611.19419
作者: Lavinia Ghita,Dhruv Desai
机构: NVIDIA(英伟达)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Matrix-structured optimizers can improve language-model pretraining, but it is unclear whether these gains transfer to short-horizon KL distillation from an instruction-tuned initialization. We compare AdamW and diagonal AdaGrad with two tractable Shampoo reductions, SOAP, and All-2D Muon across four dense Qwen student-teacher pairs under a common forward-KL objective and fixed training budget. Across scales, none of the evaluated matrix-structured recipes consistently improves the quality versus trainer-time trade-off over the elementwise references: AdamW generally attains the lowest validation KL, while AdaGrad can be slightly faster at larger scale. SOAP often tracks AdamW in KL but incurs substantial trainer-time overhead, and the two Shampoo reductions reverse ordering across model pairs, showing that conclusions depend on the feasible recipe. These results suggest that pretraining optimizer gains do not transfer automatically to KL distillation and should be re-established for the target objective and systems regime.

[NLP-125] Real-World End-to-End Product Entity Resolution

【速读】: 该论文旨在解决电子商务环境中产品实体消歧(Product Entity Resolution, PER)所面临的挑战,主要包括非结构化文本、格式不一致、促销信息干扰以及细粒度产品差异等问题。其解决方案的关键在于提出并端到端评估四种零样本(zero-shot)PER流水线——词法级联(Lexical Cascade)、语义级联(Semantic Cascade)、混合融合级联(Hybrid Fusion Cascade)与生成式-词法级联(Generative-Lexical Cascade),并通过在多个基准数据集及两个包含跨语言噪声和密集品牌冲突的真实世界数据集上的实验,系统比较了词法与密集型阻断(lexical and dense blocking)策略在结合交叉编码器(cross-encoder)、张量搜索(tensor-search)与大语言模型(Large Language Model, LLM)验证下的性能表现,综合考量准确率、运行时效率与内存消耗,以实现高精度、高效且可扩展的实体消歧。

链接: https://arxiv.org/abs/2611.19418
作者: Dimitrios Karapiperis,George Papadakis,Manolis Koubarakis
机构: International Hellenic University (国际赫里尼大学); National and Kapodistrian University of Athens (国家和卡波迪斯特里亚大学雅典)
类目: Computation and Language (cs.CL); Databases (cs.DB)
备注:

点击查看摘要

Abstract:Product Entity Resolution (PER) in e-commerce is challenging due to unstructured text, formatting inconsistencies, promotional noise, and fine-grained product differences. We present an end-to-end evaluation of four zero-shot PER pipelines—Lexical Cascade, Semantic Cascade, Hybrid Fusion Cascade, and Generative-Lexical Cascade—across established benchmarks and introduce two real-world datasets characterized by cross-lingual noise and dense brand collisions. We evaluate lexical and dense blocking with cross-encoder, tensor-search, and Large Language Model verification, considering accuracy, runtime, and memory consumption.

[NLP-126] From Facts to Personas: Interpretable Role Unlearning in LLM s via Mixture-of-Experts NEURIPS2026

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在多角色对话与人格模拟应用中因过度模仿特定角色的语言风格与行为模式所引发的隐私泄露、社会偏见强化及生成内容不可控等问题。其核心挑战在于如何在不损害模型整体语言能力的前提下,实现对目标角色特征的精准“遗忘”。解决方案的关键在于提出一种名为RoleMoE的专家混合(Mixture-of-Experts, MoE)框架,通过将不同角色映射至更独立的专家子空间,增强角色间的分离性与结构可解释性。为进一步区分角色特异性特征,RoleMoE引入两种互补的损失函数:专业化损失(specialization loss)促使各角色采用差异化的专家分布,解耦损失(disentanglement loss)则推动专家发展出独立的能力子空间,从而实现对目标角色专家的细粒度选择性遗忘。为此,研究还构建了RoleBench-Unlearn评估基准,实验证明RoleMoE可在保持非目标角色语言质量、知识保留与人格一致性的同时,实现精确且可控的角色遗忘。

链接: https://arxiv.org/abs/2611.19393
作者: Ruihong Zeng,Puning Yang,Jinghui Zhang,Shen Gao,Xiangliang Zhang,Preslav Nakov,Xiuying Chen
机构: MBZUAI; University of Electronic Science and Technology of China(电子科技大学); University of Notre Dame(圣母大学)
类目: Computation and Language (cs.CL); Computers and Society (cs.CY)
备注: Accepted at NeurIPS 2026

点击查看摘要

Abstract:LLMs are widely adopted in multi-role dialogue and persona simulation. However, their strong ability to imitate role-specific linguistic and behavioral patterns can introduce privacy risks, reinforce social biases, and lead to unsafe or uncontrollable outputs. To address these issues, we introduce the Role-playing Unlearning task, which aims to selectively forget the style and knowledge associated with target roles while preserving the language ability and behaviors of the remaining roles. Unlike prior unlearning tasks that focus solely on removing factual information, our task additionally requires forgetting persona-specific characteristics, making the control over model generation more fine-grained. We propose RoleMoE, a Mixture-of-Experts framework that maps different roles to more independent expert subspaces, improving role separation and structural interpretability. To further distinguish role-specific features, RoleMoE incorporates two complementary losses: a specialization loss that drives roles to adopt distinct expert distributions, and a disentanglement loss that encourages experts to develop distinct capability subspaces. This decomposed structure enables role unlearning to focus selectively on target-role experts. We finally propose RoleBench-Unlearn for evaluation, and experiments on multiple backbone LLMs show that RoleMoE enables precise and controllable role forgetting, while better preserving the linguistic quality, knowledge, and persona consistency of non-target roles. All code and datasets are publicly available at: this https URL.

[NLP-127] CausalGloss: Multi-Evidence Action Control for a Simultaneous Mandarin Speech-to-ASL Gloss Cascade ICASSP2027

【速读】: 该论文旨在解决中文语音到美国手语(ASL)词汇(gloss)的同步翻译中,如何在动态生成过程中准确判断何时应确定每个词汇的决策问题。传统固定等待策略(Wait-k)无法适应源语言与目标词汇之间的重排序现象,而仅基于参考历史训练则会导致暴露偏差(exposure bias)。其解决方案的关键在于提出一种基于因果推理的英文前缀到ASL词汇的动作策略——CausalGloss,该方法融合词法与语义证据、引入有限延迟约束、负向约束、突发性写入(burst WRITE)动作以及不完整历史恢复机制,实现对生成时序更灵活且鲁棒的控制。在基于How2Sign衍生的500例冻结基准上,CausalGloss在k=5时达到Token F1 0.7047,接近三种子代TransLLaMA风格学习型自适应基线(0.7067),同时将平均局部对齐损失(LAAL)降低54.4%;在人类标注参考(CLERC Epee)下,取得0.649 Token F1,显著优于多种基线,并实现帕累托主导。消融实验表明时间决策与保留策略可分离优化。此外,构建了包含19,924个语句、总时长36.54小时的可控评估语料库ChineseSpeech-ASLGloss-v1,验证了模型在真实场景下的有效性。

链接: https://arxiv.org/abs/2611.19375
作者: Mingze Li,Satoshi Nakamura
机构: 未知
类目: Computation and Language (cs.CL); Audio and Speech Processing (eess.AS)
备注: 5 pages, 1 figure, 2 tables. Submitted to IEEE ICASSP 2027

点击查看摘要

Abstract:Simultaneous Mandarin speech-to-American Sign Language (ASL) gloss translation requires deciding when an evolving utterance supports committing each gloss. Fixed Wait-k cannot adapt to source-gloss reordering, while training only on reference histories causes exposure bias. We present CausalGloss, an English-prefix-to-ASL-gloss action policy for a Mandarin-speech cascade. It combines lexical and semantic evidence, bounded delay, negation constraints, burst WRITE actions, and incomplete-history recovery. On a frozen 500-example How2Sign-derived benchmark with weak LLM-generated references, CausalGloss improves the fixed Wait-k frontier and nearly matches a three-seed TransLLaMA-style learned adaptive baseline at k = 5 in Token F1 (0.7047 versus 0.7067), with 54.4% lower LAAL. Against CLERC Epee human-gloss references, it achieves 0.649 Token F1 and Pareto-dominates three-seed TransLLaMA-style k = 5 and fixed k = 5, 7, reducing LAAL by 58.2-69.9%. Ablations separate timing from retention. A synthetic Mandarin-speech cascade preserves a smaller computation-aware latency gain, although outputs remain over-long. We also construct ChineseSpeech-ASLGloss-v1, a 19,924-utterance, 36.54-h controlled evaluation corpus.

[NLP-128] Leakage-aware calibrated decision-level fusion for mental-health text classification

【速读】: 该论文旨在解决多模型决策级融合(decision-level fusion)中因训练数据泄露与模型间冗余导致的性能瓶颈问题,尤其关注在精神健康文本分类任务中如何有效利用互补性信息并提升模型校准度与泛化能力。其核心挑战在于传统基于词汇近似重复(lexical near-duplicates)的数据去重方法无法识别深层的语料血缘依赖(lineage overlap),从而导致评估结果产生偏差。解决方案的关键在于引入血缘审计(augmentation-lineage auditing)以检测训练数据中的潜在依赖关系,并采用分层独立划分(lineage-disjoint partitions)确保验证过程的可靠性;同时,提出一种类别特异性互补性指数(class-specific Complementarity Index, CI),通过计算各模型正确分类样本集之间的杰卡德距离(Jaccard distance),量化模型间的预测差异性,进而指导集成选择。此外,通过**后融合校准(post-fusion recalibration)**显著降低期望校准误差(ECE),最终实现宏平均F1提升0.021,且集成模型在“救援案例”(fusion-rescued cases)中展现出更高的可解释性多样性。研究表明,仅依赖整体准确率会掩盖模型间的实际互补行为,而结合血缘审计、互补性分析与校准机制,能够更真实地揭示集成学习的内在优势。

链接: https://arxiv.org/abs/2611.19360
作者: Wadii Boulila,Giuseppe Varone,Bilel Benjdira,Maha Driss
机构: Prince Sultan University (沙特阿拉伯利雅得); Harvard Medical School (美国马萨诸塞州波士顿); Alfaisal University (沙特阿拉伯利雅得)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Decision-level fusion benefits from complementary errors, calibrated probabilities, and leakage-aware evaluation. We evaluated a heterogeneous transformer fusion framework on 262,276 proxy-labeled Reddit records covering anxiety, depression, schizophrenia, and suicidal ideation. In a stratified random split, MinHash-LSH identified 29.1% of test records as lexical near-duplicates of training data. Augmentation-lineage auditing found 50.2% lineage overlap with training, including 30.0% of records retained after lexical filtering. We therefore used lineage-disjoint partitions. Twelve encoders, comprising three seeds each of Bio-ClinicalBERT, MentalBERT, MentalRoBERTa, and RoBERTa-base, underwent cross-fitted validation for calibration and ensemble selection. A class-specific Complementarity Index (CI), defined as the Jaccard distance between members’ correctly classified validation sets, was assessed against independent fusion gain. CI correlated with gain for depression (r=0.50) and suicidal ideation (r=0.68), but not anxiety; results for schizophrenia were mixed. The selected MentalBERT, MentalRoBERTa, and RoBERTa-base triad achieved macro-F1 0.844 [0.823, 0.863], MCC 0.779, and ECE 0.0033 after post-fusion recalibration, improving macro-F1 by 0.021 [0.007, 0.037] over the validation-selected best single encoder. Integrated-Gradients attribution diversity was 2.3 times greater in fusion-rescued cases. Removing divergence-driving words destroyed 58% of rescues, versus 12% under random removal. These findings show that lineage auditing detects dependencies missed by lexical filtering, while class-specific complementarity and post-fusion calibration reveal ensemble behavior obscured by aggregate accuracy.

[NLP-129] A Comparative Evaluation of BERT-Based and LLM -Based Models for Indic Sentence Similarity

【速读】: 该论文旨在解决生成式 AI(Generative AI)时代下,基于大语言模型(LLM)的嵌入表示在印地语族(Indic)文本语义相似性任务中相对于传统BERT类编码器的有效性尚未得到充分验证的问题。研究通过在10种印地语族语言的翻译版语义文本相似性基准(STS-B)开发集及原生马拉地语MahaSTS测试集上对比8个BERT类模型与5个LLM类模型的表现,发现IndicSBERT在所有翻译语言设置中均取得最高皮尔逊相关系数(平均0.843),显著优于各LLM模型;在MahaSTS上,IndicSBERT与Bhasha-Embed-v0分别达到0.901和0.900的相关性,而领先的LLM模型EmbeddingGemma为0.871。结果表明,尽管部分LLM模型表现良好,但其嵌入在印地语族语言上的整体性能并未一致超越成熟且针对特定语言优化的BERT类编码器。因此,解决方案的关键在于强调:当前基于大语言模型的嵌入并不具备普遍优势,其效果高度依赖于语言覆盖范围、任务标注监督强度以及基准数据构建方式,亟需建立更精细化、多维度的评估体系以推动跨语言语义理解技术的合理发展。

链接: https://arxiv.org/abs/2611.19356
作者: Aishwarya Gunjal,Kanchan Khemnar,Pawan Poojary,Pranav Kulkarni,Raviraj Joshi
机构: Pune Institute of Computer Technology, Pune, India(普奈计算机技术学院,普奈,印度); L3Cube Labs, Pune, India( L3Cube 实验室,普奈,印度)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:The relative effectiveness of BERT-based encoders and large language model (LLM)-based embeddings for Indic semantic textual similarity remains insufficiently characterized. We compare eight BERT-based and five LLM-based models on the 1,500-pair STS Benchmark development set translated into ten Indic languages and on the native Marathi MahaSTS test set. IndicSBERT achieves the highest Pearson correlation in all ten translated language sets, with a macro-average of 0.843, followed by BGE-M3 at 0.792. Harrier-270M is the strongest LLM-based model on translated STS-B, with a macro-average of 0.751. On MahaSTS, IndicSBERT and Bhasha-Embed-v0 achieve Pearson correlations of 0.901 and 0.900, respectively, while EmbeddingGemma leads the LLM-based models at 0.871. The highest-scoring BERT-based model thus exceeds the highest-scoring LLM-based model in both settings, although the magnitude of the difference varies across benchmarks. These findings indicate that LLM-derived embeddings do not uniformly improve Indic sentence similarity over established encoders and motivate evaluation that accounts for language coverage, task supervision, and benchmark construction.

[NLP-130] When Does Label-Free Deconfounding of Document Embeddings Work?

【速读】: 该论文旨在解决文档嵌入(document embeddings)中包含的伪属性(spurious attributes)对语义比较造成扭曲的问题,尤其针对在缺乏标注数据(metadata)情况下难以实施监督式擦除(supervised erasure)的挑战。其核心解决方案是探索无需标签的去混淆(label-free deconfounding)方法在不同语料库和嵌入模型中的有效性。研究发现,无监督方法在提升语义实用性方面优于未经处理的嵌入;其中,简单的几何方法能够恢复大量监督式擦除的效果,甚至在部分线性混淆因子可恢复性和语义实用性指标上达到与基于元数据的监督擦除相当的水平。研究进一步识别出两种导致擦除失效的关键原因:代理混淆因子(proxy-confounder)对齐不佳以及干预覆盖不足。关键洞见在于,混淆因子可恢复性与语义实用性是去混淆成功与否的两个独立维度——语义实用性可提升而残留混淆信息仍存在,且更强的抑制并不必然带来实用性改善;同时,线性可恢复性降低的同时仍可能存在显著的非线性可恢复性。这些发现揭示了混淆因子发现与干预有效性之间的本质差异,并强调在评估去混淆效果时,必须同时考量噪声抑制与语义保留的平衡。

链接: https://arxiv.org/abs/2611.19350
作者: Anna Kiepura(ETH Zurich),Yingqiang Gao(Department of Computational Linguistics, University of Zurich),Elliott Ash(ETH Zurich),Yu Fan(The University of Hong Kong)
机构: ETH Zurich(苏黎世联邦理工学院); University of Zurich(苏黎世大学); The University of Hong Kong(香港大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Document embeddings encode spurious attributes that can distort semantic comparisons; yet supervised erasure requires metadata that is often unavailable. We systematically investigate label-free deconfounding across corpora and embedding models. Our results show that label-free methods improve semantic utility over unprocessed embeddings, while simple geometric approaches recover much of the supervised benefit, sometimes matching metadata-supervised erasure in linear confounder recoverability and semantic utility. We identify two distinct causes of ineffective erasure: poor proxy-confounder alignment and insufficient intervention coverage. Crucially, confounder recoverability and semantic utility are distinct dimensions of deconfounding success: utility can improve despite residual confounder information, while stronger suppression does not necessarily improve utility. Moreover, reduced linear recoverability can coexist with substantial nonlinear recoverability. Together, these findings distinguish confounder discovery from intervention effectiveness and highlight the importance of evaluating nuisance suppression alongside semantic preservation.

[NLP-131] OmitEHR: Scalable Omission Detection in LLM -Generated Clinical Summaries ALT ML4H

【速读】: 该论文旨在解决生成式人工智能(Generative AI)在医疗记录摘要生成中普遍存在的重要临床信息遗漏问题,即模型生成的摘要中缺失关键临床事实。现有方法依赖人工设定、忽略或通过昂贵的大语言模型(LLM)裁判在推理时重新计算临床重要性信号,存在效率低、成本高且可扩展性差的问题。其解决方案的关键在于提出OmitEHR——一种无需参考标准摘要的遗漏检测方法,通过学习历史时期由临床医生撰写的摘要中所包含的句子来自动建模临床重要性(clinical salience),进而识别出在大语言模型生成摘要中未覆盖但具有高临床重要性的原始病历句段作为候选遗漏项。该方法在5,000例住院患者数据上验证,显著优于包括提示工程化LLM裁判在内的各类基线模型,在保持高临床有效性的同时,实现1.6至1.7倍的加速和5.6至6.7倍的成本降低,具备实际部署价值。

链接: https://arxiv.org/abs/2611.19343
作者: Bridget Lin,Ryan Welch,Bobak Seddighzadeh,Fran Riley,Kaitlyn Stoehr,Margarita I. Pipinos,Dimitri Speron,Emma V. Sage,Sneha Jain,Nigam H. Shah,Emily Alsentzer
机构: Stanford University (斯坦福大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 21 pages, 6 figures, 15 tables. Accepted at Machine Learning for Health (ML4H) 2026

点击查看摘要

Abstract:Health systems increasingly rely on large language models (LLMs) to draft summaries of patient medical records, yet evaluations now find omissions, the absence of clinically important information, to be the most common failure mode. Detecting omissions in an LLM-generated summary requires knowing which facts in the record are salient, a signal that existing methods hand-specify, ignore, or recompute at inference with costly LLM judges. We present OmitEHR, a reference-free omission detector that learns clinical salience from historical clinician-written summaries predating widespread LLM use. OmitEHR learns to predict which source-note sentences clinicians include in their summaries, then surfaces high-salience sentences not covered by an LLM-generated summary as candidate omissions. We apply OmitEHR to brief hospital course (BHC) summarization using 5,000 admissions from a large academic medical center. On held-out admissions, OmitEHR outperformed all evaluated baselines at predicting salient source sentences, including a prompted LLM judge. The full omission detection pipeline took 17.7 to 19.5 seconds and cost \ 0.03 to \ 0.13 per summary depending on the base model, 1.6 to 1.7 times faster and 5.6 to 6.7 times cheaper than scoring every source sentence with an LLM judge. In a blinded review of 25 admissions, two board-certified clinicians confirmed a mean of 5.6 clinically meaningful omissions per summary surfaced by OmitEHR, compared with 3.3 for a full LLM pipeline, counting candidates confirmed by at least one clinician. The efficiency and validity of OmitEHR make it a practical tool for health systems to flag omissions before a summary reaches the clinician.

[NLP-132] On Representational and Functional Alignment of Training Data Attribution Methods for Language Models EACL2027

【速读】: 该论文旨在解决现有训练数据归因方法在评估与比较上的不足问题,特别是由于缺乏直接、系统性的对比框架,导致对不同方法性能的评价多依赖下游任务表现这一间接指标。其核心问题是:如何在不依赖下游任务的情况下,客观衡量和比较不同归因方法之间的有效性与一致性。解决方案的关键在于提出一个三维度的直接比较框架:(1)功能一致性(functional alignment),通过全量归因向量间的相关性度量方法间的一致性;(2)表征一致性(representational alignment),捕捉支撑归因结果的底层表示结构的相似性;(3)跨方法可预测性(cross-method predictability),评估不同方法在局部成对排序决策上的一致性。研究发现,同一范式内部具有较强的表征一致性和中等的功能一致性,而跨范式之间的一致性较低且任务依赖性强,但对不同模型表现出稳定性。这表明以往观察到的不同方法在下游任务中表现相似,并非源于对完整训练数据排名的高度一致,而是源于对文档相对重要性的共识。

链接: https://arxiv.org/abs/2611.19286
作者: Loris Schoenegger,Benjamin Roth
机构: University of Vienna (维也纳大学); UniVie Doctoral School Computer Science (维也纳大学博士生院计算机科学)
类目: Computation and Language (cs.CL)
备注: EACL 2027. 9 pages

点击查看摘要

Abstract:Existing approaches to training data attribution derive influence estimates from different sources of information, including gradients, model-internal states, and surface-level similarity. However, comparisons between attribution methods remain incomplete, often relying on indirect evaluations through downstream task performance. In this work, we introduce a framework for directly comparing methods along three dimensions: functional alignment, which measures correlation between full attribution vectors; representational alignment, which captures structural similarity of the representations underlying attribution; and cross-method predictability, which evaluates agreement on local pairwise ranking decisions. We observe strong representational and moderate functional alignment within paradigms. Alignment across paradigms is low and task-dependent, yet stable across models. Our results suggest that the similar downstream performance observed across paradigms in prior work can be attributed to agreement on the relative importance of documents, rather than to agreement on rankings of the full training data.

[NLP-133] B-OPSD: On-Policy Self-Distillation with Self-Generated Problem Briefs

【速读】: 该论文旨在解决生成式 AI (Generative AI) 在强化学习框架中进行在线策略自蒸馏(On-Policy Self-Distillation, OPSD)时对真实答案(ground-truth solutions)的依赖问题,这类依赖限制了其在缺乏标注解题过程或最终答案场景下的应用。针对此问题,论文提出一种基于自生成问题简述(Problem Briefs, B-OPSD)的新型自蒸馏方法:通过预设提示模板,由初始模型或保存的学生检查点仅依据题目自动生成包含相关原理、适用条件及潜在误区的“问题简述”,而无需输出最终答案或完整解法;随后,冻结的教师模型利用该自生成的简述作为特权上下文,在训练过程中通过词元级分布匹配指导当前学生模型的在线策略更新。该方案彻底摒弃了训练阶段对真实解题路径的依赖,实现了完全无监督的教师引导机制。实验结果表明,在AIME 2024、AIME 2025和HMMT 2025三个基准测试上,B-OPSD在三种Qwen3模型规模下均显著优于传统OPSD,且平均准确率更高,验证了其有效性与泛化能力。

链接: https://arxiv.org/abs/2611.19273
作者: Zhi Xu,Yinqi Wu,Feng Zhang,Xiaotong Zhang,Hong Yu,Han Liu
机构: Dalian University of Technology (大连理工大学); Dalian, China
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:On-Policy Self-Distillation (OPSD) has recently attracted interest because it provides dense teacher guidance without a separate, often larger teacher model. However, OPSD conditions its teacher on ground-truth solutions, which limits its use when such solutions are unavailable. We propose On-Policy Self-Distillation with Self-Generated Problem Briefs (B-OPSD). Using a prompt template, the initial model or a saved student checkpoint generates a problem brief from the question alone. The prompt asks for relevant principles, applicability conditions, and likely pitfalls, but not a final answer or a complete solution. A frozen teacher uses the brief as privileged context to guide the current student’s on-policy rollouts through token-level distribution matching. B-OPSD uses no ground-truth solutions during training. Results on AIME 2024, AIME 2025, and HMMT 2025 show that both B-OPSD variants achieve higher average accuracy than OPSD at each of three Qwen3 model sizes. Code is available at this https URL

[NLP-134] Auditory-Feedback-Conditioned Flow TTS for Implicit Lombard Adaptation

【速读】: 该论文旨在解决传统文本到语音(Text-to-Speech, TTS)系统在噪声环境中语音可懂度下降的问题,其核心挑战在于系统缺乏对环境变化的自适应能力。为应对这一问题,研究提出了一种基于听觉反馈的洛姆伯德效应(Lombard effect)驱动的生成式TTS框架,关键创新在于利用预训练的基于流(Flow-based)的TTS模型,通过直接从噪声污染波形中提取信噪比(SNR)反馈与自动语音识别(ASR)反馈,将其注入到流匹配(Flow-matching)条件中,实现话语级精细化修正与因果块级更新,从而无需显式韵律控制即可学习声学适应性。实验表明,尽管微调(finetuning)在特定噪声条件下表现鲁棒,但听觉反馈机制能带来依赖于提示和噪声类型的额外增益,并显著提升对未见过噪声场景的泛化能力;同时研究揭示,仅增强音高(F0)或能量跟踪强度并不足以保证语音可懂度与质量的提升,凸显了反馈机制在语义-声学耦合优化中的关键作用。

链接: https://arxiv.org/abs/2611.19260
作者: Yunya Wang,Satoshi Nakamura
机构: The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳)); School of Data Science(数据科学学院); School of AI(人工智能学院)
类目: ound (cs.SD); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Conventional text-to-speech (TTS) systems can lose intelligibility in noise environments because they cannot adapt to environmental changes. However, humans tend to increase their vocal efforts in noisy environments, a phenomenon known as the Lombard effect. Building on an auditory-feedback Lombard TTS, we investigate whether a pretrained Flow-based TTS can learn acoustic adaptation directly from auditory feedback without explicit prosodic control. Signal-to-noise ratio (SNR) feedback and automatic speech recognition (ASR) feedback are extracted from the noise-corrupted waveforms and injected into the Flow-matching condition for utterance refinement and causal block-level updates. We assess feedback benefits across prompting and noise conditions using plain/Lombard prompting, prompt-switching controls, and paired-finetuning. Experiments show that while finetuning provides strong robustness under in-domain conditions, auditory feedback yields additional cue-dependent benefits that generalize to unseen noise. Moreover, stronger power/F0 tracking alone does not guarantee better speech intelligibility or quality.

[NLP-135] Omni-Workspace: Benchmarking Omni-Modal AI Assistants on Multimodal Workspace Tasks

【速读】: 该论文旨在解决当前多模态智能体在复杂真实多模态工作环境中难以实现长期、持续的任务完成问题。现有基准主要聚焦于多模态感知与理解,缺乏对跨文件、跨模态依赖关系下持续交互与任务执行的充分评估。为此,论文提出Omni-Workspace基准,涵盖4种多模态工作空间类型及两个难度层级的380个深度研究型任务,要求智能体具备文件解析、外部信息检索、跨模态证据整合以及生成符合显式需求成果的能力。评估结果显示,现有18个代表性模型在长周期任务执行中表现受限,最优模型平均得分仅为65.93,暴露出多模态上下文膨胀(multimodal context inflation)和任务状态漂移(task-state drift)等核心挑战。针对这些问题,论文进一步提出OmniFocus方法,其关键在于通过轻量级的多模态工作记忆管理与显式的任务状态追踪机制相结合,有效提升智能体在多模态工作空间中的性能与效率。

链接: https://arxiv.org/abs/2611.19202
作者: Yuhao Wang,Yucheng Zhao,Yiran Zhong,Feida Zhu,Hongcheng Liu,Qichao Wang,Ming Ma,Yu Wang
机构: Shanghai Jiao Tong University(上海交通大学); Tongyi Lab, Alibaba Group(通义实验室,阿里巴巴集团); Nanyang Technological University(南洋理工大学); University of Chinese Academy of Sciences(中国科学院大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Omni-modal agents are expected to move beyond multimodal understanding toward completing complex tasks in realistic multimodal environments. In these settings, information is distributed across heterogeneous files and modalities, with cross-file and cross-modal dependencies requiring agents to continually coordinate perception, reasoning, and tool use as newly gathered evidence informs subsequent actions. However, existing omni-modal benchmarks primarily focus on multimodal perception and understanding, leaving complex task completion through sustained interaction with multimodal workspaces underexplored. We introduce Omni-Workspace, a benchmark comprising 380 deep-research-style tasks across four multimodal workspace types and two difficulty levels. These tasks require agents to inspect and process workspace files, retrieve external information when needed, integrate evidence across files and modalities, and produce artifacts satisfying explicit requirements. Evaluating 18 representative models reveals limitations in long-horizon multimodal workspace execution, with the best model scoring only 65.93 on average. Further analysis identifies multimodal context inflation and task-state drift as recurring challenges. Motivated by these findings, we further propose OmniFocus, a lightweight approach combining multimodal working-memory management with explicit task-state tracking to improve performance and efficiency. Omni-Workspace advances the evaluation and development of omni-modal agents toward autonomous completion of complex tasks in multimodal workspaces.

[NLP-136] Isolating Perception for Multi-View Planning under Occlusion-Induced Partial Observability EACL2027

【速读】: 该论文旨在解决视觉语言模型(Vision-Language Models, VLMs)在面对遮挡引起的局部可观测性(Occlusion-induced Partial Observability, PO)时,因深度模糊性导致的性能急剧下降问题,尤其是在复杂场景中规划任务的可扩展性受限。现有VLMs在视觉规划基准上已趋于饱和,但在真实复杂场景下仍表现不佳。为此,作者提出基于积木世界(Blocksworld)的BW-PO基准,该基准系统性地暴露了当前VLM规划器与形式化器的局限性,使其性能趋近于零。解决方案的关键在于提出“VLM-as-perceiver”范式——将VLM的角色严格限定为局部感知模块,而将跨视角证据融合与规划生成任务交由外部求解器完成。这一分离机制有效解耦了两个长期纠缠的瓶颈:一是深度模糊性下的推理一致性问题,二是高密度场景中视觉信息提取的完整性问题。实验结果表明,该方法显著提升了多种主流VLM在挑战性子集上的规划成功率,如Gemma-4-31B从1%提升至97%,Qwen3.5-122B-A10B从0%提升至95%,Gemini-3.1-Pro从3%提升至84%。

链接: https://arxiv.org/abs/2611.19179
作者: Kevin Nguyen,Ceyhun Efe Kayan,Feng Liu,Li Zhang
机构: Drexel University(德雷塞尔大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: EACL 2027

点击查看摘要

Abstract:Vision-language models (VLMs) are beginning to saturate visual planning benchmarks, yet they fail to resolve depth ambiguity and degrade sharply with problem size under occlusion-induced partial observability (PO). We introduce BW-PO based on the Blocksworld domain, a controlled benchmark that reduces existing VLM planners and formalizers to near-zero performance, and propose VLM-as-perceiver, a paradigm that narrows the VLM’s role to local perception while delegating cross-view evidence reconciliation and plan generation to external solvers. Across model-specific challenging subsets of BW-PO, this raises plan success from 1% to 97% for Gemma-4-31B, 0% to 95% for Qwen3.5-122B-A10B, and 3% to 84% for Gemini-3.1-Pro. Isolating perception in this way separates two bottlenecks that are otherwise entangled: reconciliation under depth ambiguity, and completeness of visual extraction under scene density.

[NLP-137] A Bag of Concepts? Testing Role-Filler Binding in the Verbalizable Workspace NEURIPS2026

【速读】: 该论文旨在解决生成式人工智能(Generative AI)模型在推理过程中隐含的语义角色绑定(role-filler binding)是否可被运行时对齐监控器(runtime alignment monitor)有效捕捉的问题。具体而言,研究关注的是:基于可解释工作空间(verbalizable workspace)构建的监控器能否识别出模型内部用于判断“谁扮演何种角色”的深层语义结构,尤其是涉及安全关键关系命题(如“模型计划欺骗用户”与“用户计划欺骗模型”)中的角色分配差异。其解决方案的关键在于通过两种不同视角(J-lens与R-lens)和两种角色方向拟合方法(均值差与LRE梯度),检验工作空间中是否存在能够表征角色绑定的可分离表征。研究发现,尽管残差流(residual stream)上的线性探测器几乎能完美恢复角色信息,但工作空间本身并未携带足够的、可被探测到的角色绑定表示——即工作空间虽能引导输出内容中角色的表达方式,却无法反映模型实际计算角色绑定所依赖的内在表征。进一步的消融实验表明,仅当零化工作空间中前十个活跃镜头方向的投影时,才在其中一模型(Gemma-3-27B-IT)上观察到绑定性能下降,且该效应在多个镜头变体下保持一致,说明该设计具备检测绑定存在的能力;然而,这一现象也暗示模型规模与指令微调等因素可能共同影响结果。因此,基于此类方向构建的工作空间-镜头监控器将无法捕获模型自身真实持有的角色分配状态,从而导致对潜在风险行为的误判。

链接: https://arxiv.org/abs/2611.19176
作者: Reya Dawlah,Elizabeth McMillin,Michael Shihong Zhang,Badrish Ananth,Rishi Adi Ruia,Ane Zuñiga,Thomas Jiralerspong
机构: University of Texas at Dallas; MiraCosta College; University of Waterloo; University of California, Irvine; University of Washington; Massachusetts Institute of Technology; Mila, Université de Montréal
类目: Computation and Language (cs.CL)
备注: Accepted at Interpretability as a Science Workshop, NeurIPS 2026

点击查看摘要

Abstract:Runtime alignment monitors built on a verbalizable workspace have been proposed as a way to surface a model’s silent cognition before it produces output. Their value rests on the workspace carrying enough structure to distinguish safety-relevant propositions, most of which are relational: “the model plans to deceive the user” and “the user plans to deceive the model” share a concept set and differ only in role assignment. We ask whether the workspace carries the role-filler bindings (which entity fills which role) such a monitor would need, on three frontier language models (Qwen3.6-27B, Gemma-3-12B, Gemma-3-27B-IT), under two lens families (J-lens and R-lens), and with two methods for fitting the role direction (difference-of-means and LRE-gradient). We find a clean dissociation: the workspace carries directions that steer what the model outputs about roles but not the representations the model uses to compute them. Linear probes recover role at near-ceiling from the residual stream but score below a rank-matched random control when reading the workspace, and steering the model along the fitted role direction changes which entity it binds to no more than strength-matched control edits do. Ablating the workspace (zeroing the residual stream’s projection onto its ten most active lens directions) damages binding on one of the three models (Gemma-3-27B-IT, holding across three lens variants) and shows no binding-specific damage on the other two, with a rank-matched random subspace as the control; this confirms our design detects binding when present, though scale and instruction tuning are confounded there. A workspace-lens monitor built on these directions would therefore miss a role assignment the model itself represents.

[NLP-138] Personalized Language Models for One-to-Many Private Communications

【速读】: 该论文旨在解决一对一私密通信场景中关系特定的持续性回复预测问题,即系统需根据作者与特定联系人之间的历史互动模式,动态预测其对重复通信的回复,并在实际回复出现后更新个性化状态。其核心解决方案是提出一种名为TRACE(Temporal Relationship-Aware Communication Evolution)的框架,通过结合作者级词汇习惯、联系人特异性表达方式以及时间相关的表达估计,对冻结的语言模型(frozen language model)进行引导,并由反馈驱动的策略控制状态修订过程。实验结果表明,在8名Enron用户共216封测试邮件上,所有四种TRACE配置均在平均BLEU、ROUGE-L和BERTScore指标上优于基线大语言模型(Base LLM)和基于检索增强生成(Base RAG)的方法,同时在回复风格均方误差(Reply Style MSE)上表现更优;然而,组件消融实验显示效果不一,且经多重比较校正后无法建立统计上可靠的回复风格误差改进。此外,更精确的状态预测或更大的修订预算并未一致提升回复层面的性能指标,揭示了将动态演化的、关系特定的表达估计转化为有效生成指导仍面临显著挑战。

链接: https://arxiv.org/abs/2611.19155
作者: Zihan Hong,Xue Li
机构: The University of Queensland(昆士兰大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:We study relationship-specific continual reply prediction for one-to-many private communication: a system predicts an author’s reply to a recurring contact and updates its personalization state when the actual reply becomes available. We propose TRACE (Temporal Relationship-Aware Communication Evolution), which uses author-level lexical habits, contact-specific wording, and temporal expression estimates to guide a frozen language model, with a feedback-driven policy controlling revision. On 216 test emails from eight Enron users, all four main TRACE configurations achieve higher mean BLEU, ROUGE-L, and BERTScore and lower mean Reply Style MSE than Base LLM and Base RAG. Component ablations show mixed effects and do not establish statistically reliable improvements in Reply Style MSE after multiple-comparison correction. More accurate state forecasts and larger revision budgets do not consistently improve reply-level metrics. These exploratory findings expose the challenge of turning evolving relationship-specific expression estimates into useful generation guidance.

[NLP-139] MarineSparkLM-4B: A Lightweight Language Model for Marine and Offshore Engineering Built by Parameter-Efficient Adaptation

【速读】: 该论文旨在解决海洋与近海工程领域文本数据量小、难以支撑大规模预训练模型的问题,探索在单台工作站级设备上,通过参数高效微调(Parameter-Efficient Fine-Tuning, PEFT)技术对轻量化语言模型进行适配的可行性与极限。其核心挑战在于如何在有限的数据和计算资源下,实现具备专业领域理解能力的生成式模型。解决方案的关键在于采用两种独立的低秩自适应(LoRA)适配器对40亿参数的OceanGPT-basic-4B-Instruct模型进行微调:第一种基于6160万词元的多主题工程语料库,通过下一词预测任务训练,以提升模型在跨领域术语上的稳定性;第二种则针对2295个问答对(含人工撰写与基线模型生成)进行优化,仅对答案部分的损失进行约束,以增强回答准确性与主题相关性。实验表明,尽管模型在指令适配后性能显著提升(评分从84增至88),但绝对准确率仍较低(无检索时14/40,有检索时20/40),反映出当前方法在处理复杂专业问题时的局限性。研究还揭示了评估体系的不足,包括样本量小、缺乏评分者间一致性检验及依赖模型生成参考答案等问题,进一步凸显了该领域专用模型开发中的关键挑战。

链接: https://arxiv.org/abs/2611.19151
作者: Cheng Siong Chin
机构: Newcastle University Singapore(纽卡斯尔大学新加坡校区)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 4 figures, 12 pages

点击查看摘要

Abstract:Marine and offshore engineering text is spread across preprints, open-access journals, encyclopedias and institutional documents, and the amount of it is small by pretraining standards. We ask how far a lightweight language model for marine and offshore engineering can be taken with parameter-efficient adaptation on a single workstation-class device. Starting from the 4-billion-parameter OceanGPT-basic-4B-Instruct, we train two independent LoRA adapters, each updating 33.0 million of 4.06 billion parameters (0.81%). Training took 4 hours 40 minutes for the first adapter and 10 minutes for the second on one NVIDIA DGX Spark. The first adapter is trained by next-token prediction on a 61.6-million-token corpus covering ten engineering themes. The corpus was built from 13,879 source documents, of which 7,069 were retained after deduplication and a two-tier keyword filter that removes cross-domain lexical collisions. Its validation loss falls from 2.082 to 1.976 over 1,500 steps. Direct testing showed that this adapter could still drift off topic on a retrieval-grounded ballast-water prompt, so we trained a second adapter on 2,295 question-answer pairs (114 hand-written and 2,181 generated by the base model from corpus passages) with the loss restricted to answer tokens. On 20 held-out prompts scored by hand for accuracy, topicality and specificity (maximum 120), the plain base model, the continuation adapter, the instruction adapter and the instruction adapter with retrieval scored 42, 56, 84 and 88. Absolute accuracy stays low (14 of 40 without retrieval and 20 of 40 with it). We report the full pipeline, two negative results (a reverted keyword fix and an instruction run that overfit within 100 steps), and the limits of the evaluation, which uses one small prompt set, no inter-rater agreement measure and model-generated reference answers.

[NLP-140] A Systematic Study of Cross-Token Transfer for Data-Efficient Non-Verbal Text-to-Speech

【速读】: 该论文旨在解决非语言发声(Non-verbal vocalizations, NVs)数据稀缺导致的低资源生成难题,尤其关注在语音合成中如何通过跨类别发声事件的迁移学习来缓解标注数据不足的问题。现有方法虽在模型架构设计与数据构建方面取得进展,但对不同NV类型之间是否存在知识迁移尚未进行系统研究。本文的关键解决方案在于揭示并验证了跨发声类型(cross-token transferability)的有效性:即使仅使用少量目标发声(如200例笑声),引入其他类型的辅助发声(如3200例咳嗽)也能显著提升目标发声的生成质量(笑声F1从0.170提升至0.529),其性能甚至超过使用1600例笑声单独训练的结果。进一步实验表明,在1600例笑声为基准时,加入咳嗽数据可使笑声生成效果接近使用6400例笑声训练的水平(F1达0.695 vs. 0.697)。通过波形级Whisper-D评分与语义标记转录器验证,结果一致性支持该策略的普适性。因此,跨类型发声之间的知识迁移为低资源非语言发声生成提供了一条高效且互补的技术路径。

链接: https://arxiv.org/abs/2611.19145
作者: Jing Pan,Haibin Wu,Minxue Niu,Florian Metze,Yuzong Liu
机构: 未知
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Non-verbal vocalizations (NVs), such as laughter and breath, carry affective cues that words alone do not convey. Yet NV data remains scarce because speech annotation pipelines often omit or normalize these events. Existing work has improved NV text-to-speech through model design and data curation, but whether one NV token can help a model learn another has received little systematic study. We investigate this cross-token transferability in a controlled CosyVoice 2 setup. We first trace a laughter-only scaling curve, then hold laughter at 200 or 1,600 examples while adding breath, sigh, cough, lip-smack, or neutral speech. With only 200 laughter examples, adding 3,200 cough examples raises end-to-end laughter F1 from 0.170 to 0.529, surpassing the 0.495 obtained from 1,600 laughter examples alone. At the 1,600-example anchor, cough reaches 0.695, nearly matching 6,400 laughter examples (0.697). Waveform-level Whisper-D scoring and a semantic-token transcriber show consistent trends. Cross-token transfer therefore offers a complementary route to low-resource NV generation.

[NLP-141] Low-Latency LLM -Based Streaming Multi-Talker ASR with Verifiable Rewards

【速读】: 该论文旨在解决流式自动语音识别(ASR)在多说话人场景下存在的耦合转录、说话人归属(speaker attribution)与延迟问题,尤其针对可穿戴设备中近场佩戴者与远场对话伙伴之间的实时对话转录需求。由于对话界面中部分转录结果需以固定角色标签形式持续显示,导致说话人角色不可互换,且每个说话人标签必须在其发言前生成,这限制了模型的灵活性并引入延迟。传统监督微调(SFT)对稀疏标签的优化能力有限,而强制使用前置标签又会延后后续文本输出。为此,论文提出采用可验证奖励强化学习(RLVR)对流式大语言模型(LLM)进行后训练,其核心在于结合固定角色的说话人归属词错误率(SA-WER)评分器与逐轮起始延迟奖励,实现对转录准确性与响应延迟的联合优化。实验表明,该方法将SA-WER从SFT的20.51%降至13.14%,进一步引入延迟优化目标后,起始延迟的p50和p90分别降低65%和44%,同时维持12.70%的低SA-WER,证明了在固定角色设定下,可同时提升识别质量与流式响应性能。

链接: https://arxiv.org/abs/2611.19136
作者: Jing Pan,Ruizhi Li,Ju Lin,Jing Zheng,Florian Metze
机构: 未知
类目: Computation and Language (cs.CL); Sound (cs.SD)
备注:

点击查看摘要

Abstract:LLM-based architectures offer strong generalization and instruction-following capabilities for streaming automatic speech recognition (ASR), but multi-talker use introduces coupled transcription, attribution, and latency challenges. Our wearable setting requires real-time transcription of a conversation between a near-field device wearer and a far-field partner. Because partial hypotheses are rendered on persistent sides of a dialogue UI, speaker roles cannot be permuted and each speaker tag must precede its turn. These sparse tags are weakly optimized by supervised fine-tuning (SFT), while committing to a leading tag can delay the text that follows. We therefore post-train a streaming LLM-based ASR model with reinforcement learning with verifiable rewards (RLVR), combining a fixed-role Speaker-Attributed WER (SA-WER) grader with a turn-level onset-delay reward. On an eight-direction multi-talker cohort, SA-WER RL reduces SA-WER from 20.51% for SFT to 13.14% . Adding the delay objective reduces onset p50 by 65% (from 0.918,\texts to 0.320,\texts ) and p90 by 44% (from 1.389,\texts to 0.780,\texts ) relative to SA-WER-only RL, while achieving 12.70% SA-WER. These results demonstrate that fixed-role recognition quality and streaming responsiveness can be improved jointly.

[NLP-142] Empowering Masked Diffusion Models to Self-Correct with Leave-One-Out Transformers NEURIPS2026

【速读】: 该论文旨在解决标准掩码扩散模型(Masked Diffusion Models, MDMs)存在的两大根本问题:一是无法对已生成的标记进行修正,二是训练信号稀疏。为此,作者重新审视了空心变换器(Hollow Transformers)架构,提出了一种更高效且表达能力更强的结构——留一法变换器(Leave-One-Out Transformer, LOOT)。其解决方案的关键在于两个互补的改进:首先,重构了左右上下文间的融合机制,实现了更频繁的跨侧交互,显著提升了预测性能;其次,推导出一种方差更低的MDM NELBO估计器,能够对所有标记提供监督信号,从而改善验证困惑度。上述改进共同使LOOT成为一种具备有效自修正能力的独立扩散语言模型,在LM1B上达到与最强密集监督基线相当的验证困惑度;在OpenWebText上,相较于先前的重掩码方法,不仅生成质量-多样性权衡更优、MAUVE得分更高,且推理时仅需前者1/8的前向评估次数和1/3.5的推理浮点运算量,展现出卓越的效率优势。

链接: https://arxiv.org/abs/2611.19127
作者: Sofian Zalouk,Vincent Counathe,Paul Jünger,Daniel Cao,Daniel Zlotnick,Kilian Q. Weinberger,Christopher De Sa
机构: Cornell University(康奈尔大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: NeurIPS 2026

点击查看摘要

Abstract:Masked diffusion models (MDMs) have emerged as a promising paradigm for language modeling, enabling parallel token generation with competitive performance. Yet standard MDMs suffer from two fundamental limitations: (1) inability to revise tokens, and (2) sparse training signal. We revisit hollow transformers, which address both limitations by predicting each token from its surrounding context, independently of the input at that position. We introduce the Leave-One-Out Transformer (LOOT), a more efficient and expressive architecture that advances hollow transformers through two complementary advances. First, we redesign the merging mechanism to enable frequent interaction between left and right contexts, substantially improving predictive performance over prior hollow transformers. Second, we derive a lower-variance estimator of the MDM NELBO that supervises all tokens and improves validation perplexity. Together, these advances yield a competitive standalone diffusion language model capable of effective self-correction during inference. On LM1B, LOOT matches the strongest dense-supervision MDM baseline in validation perplexity. On OpenWebText, LOOT achieves higher MAUVE and better generation quality-diversity tradeoffs than prior remasking methods, using 8\times fewer forward evaluations and 3.5\times fewer inference FLOPs.

[NLP-143] Beyond Plain MCP: Do Decision Models and Code Mode Make Tool Use Cheaper and Faster?

【速读】: 该论文旨在解决语言模型智能体在使用工具时的效率与准确性问题,具体聚焦于生成式 AI (Generative AI) 智能体如何有效管理庞大的工具目录(tool catalog)。现有主流方法包括“代码模式”(code mode)——即智能体直接编写脚本调用工具桩(tool stubs),以及“决策层”(decision layer)——一种轻量级路由机制,用于预先筛选候选工具子集。研究通过在12台服务器上构建受控实验环境,引入最多1,958个干扰工具,并对Qwen、Gemma和Nemotron三款开源权重模型进行预注册评估,验证两种范式的性能表现。结果表明,代码模式显著优于传统MCP协议,可减少70%至90%的令牌(token)消耗并大幅提升任务实用性;其中,在工具桩中声明返回形状提示(return-shape hints)成为关键变量,对低性能代理的成功率提升可达39.5个百分点,其影响甚至超过引入决策层。然而,核心假设——即在代码模式前增加决策层可进一步提升性能——未获证实:对于弱代理仅带来微小且统计不显著的改进,而对强代理则造成轻微性能损耗。因此,该研究的关键发现是:代码模式本身已具备高度有效性,而额外引入决策层并非必要,反而可能引入冗余开销。研究团队公开了实验框架、一个微调后的4B工具选择模型及全部评估轨迹,以支持社区在本地化人工智能隐私与智能体架构方面的深入探索。

链接: https://arxiv.org/abs/2611.19123
作者: Abhishek Rai
机构: Independent researcher(独立研究员); Models and data: https://huggingface.co/abhishek085
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 16 pages, 11 tables. Code, harness, and models available at this http URL this https URL

点击查看摘要

Abstract:Model Context Protocol (MCP) gateways manage tool catalogs for language model agents using two primary paradigms: code mode where agents write scripts against tool stubs and decision layers which are cheap routers that preselect tool subsets. Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron, we present a preregistered evaluation on a sealed test split to determine when each approach succeeds. Code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility. Declaring return-shape hints in code mode stubs is a first-order variable, swinging success by up to 39.5 points for lower-performing agents, an effect larger than adding a router. The central hypothesis that a decision layer in front of code mode beats code mode alone is not confirmed, showing minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest. We publicly release our harness, a finetuned 4B tool selection model, and all evaluation traces to support robust community experimentation on local AI privacy and agent architectures.

[NLP-144] Latency Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation: An Externally Measured Comparison of Pipecat LiveKit and GPT -Live

【速读】: 该论文旨在解决语音代理在电话谈判中面临的实时性、口语理解准确率及对打断(话轮重叠)处理能力不足的问题,特别是在西班牙语货运费率谈判场景下。其核心挑战包括低延迟响应、高精度的金额信息识别以及在对方讲话重叠时仍能正确解析意图。解决方案的关键在于对比三种架构:基于分步流水线(语音识别-语言模型-语音合成)的Pipecat与LiveKit Agents框架,以及端到端的GPT-Live(OpenAI的语音模型)。研究发现,尽管三者在整体延迟上差异不显著(中位数分别为2,616.0 ms、3,090.0 ms、2,444.8 ms),但GPT-Live在关键指标上表现更优——其对金额的识别错误率仅为2.86%,远低于分步架构的13.01%和12.61%;同时,在应对说话人打断方面,GPT-Live几乎未误将“ok”等回声词当作有效回应,而分步架构则频繁出现误判(如3/15次和2/15次)。此外,研究强调必须以听者侧实际感知延迟为评估标准,因框架自报延迟存在显著低估(分别低9%和22%),且工具调用引入约0.9–1.1秒额外延迟,成为主要瓶颈。因此,该研究的关键结论是:端到端生成式语音模型(如GPT-Live)在降低金额识别误差、提升对话鲁棒性方面具有优势,但系统设计需特别针对金额敏感信息和话轮重叠设置防护机制,并采用真实用户侧延迟测量方法以确保评估有效性。

链接: https://arxiv.org/abs/2611.19118
作者: Jovan Popoca Moran,Eduardo de Avila-Armenta,Raúl Alejandro Aquino-Ibarra,Jorge Octavio Nicolás-Díaz,Umberto Leon-Dominguez
机构: Freight Technologies Inc.(货运科技公司); Tecnológico de Monterrey(蒙特雷理工学院); University of Monterrey(蒙特雷大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Sound (cs.SD)
备注: 12 pages, 6 figures, 11 tables. Code and data: this https URL (archived at this https URL )

点击查看摘要

Abstract:A voice agent that negotiates by phone must answer quickly, understand spoken amounts and cope with a caller who talks over it. We compared three architectures for such an agent, negotiating freight rates in Spanish: two cascades (speech recognition, language model and speech synthesis) orchestrated by Pipecat and by LiveKit Agents, and GPT-Live, an OpenAI speech model that listens and speaks without separate stages. A program played the carrier with the same 47 recordings of one speaker for all agents, without the telephone network, in 77 calls per architecture over five negotiation scenarios, and measured latency on the listener’s side with a validated instrument. Median latency to the content response was 2,616.0 ms (Pipecat), 3,090.0 ms (LiveKit) and 2,444.8 ms (GPT-Live); the confidence interval of GPT-Live’s 171.2 ms advantage over Pipecat includes zero. According to its own timestamps, GPT-Live began to answer 800 ms after the end of the line, close to the 0.8 s published by OpenAI; most of the remaining wait, about 1,367 ms and nearly constant, is the audio’s round trip through OpenAI’s service. A tool call added 0.9-1.1 s in every architecture, more than any difference between them, and the frameworks’ self-reported latency fell short of the listener’s wait by 9% (Pipecat) and 22% (LiveKit). Word error rates were low (6.03, 4.93 and 1.47%), but the error on amounts was 13.01% and 12.61% in the cascades, against 2.86% in GPT-Live. With default settings, Pipecat and LiveKit stopped at almost every backchannel and often took it as the carrier’s answer: an “ok” spoken over the agent ended in a registered agreement in 3 of 15 Pipecat calls, 2 of 15 LiveKit calls and no GPT-Live call (descriptive counts, not tested). Latency should therefore be measured on the listener’s side, and a cascade needs safeguards for spoken amounts and backchannels.

[NLP-145] EEGHyena2: Fixation-Inspired Joint Lexical-Grammatical Conditional Random Field for Interpretable EEG-to-Text Decoding

【速读】: 该论文旨在解决现有脑电图(EEG)到文本解码方法中,大语言模型(LLM)在生成过程中过度依赖自身先验知识,导致输出结果更多反映模型偏好而非真实脑电活动所携带的语义与语法信息的问题。其核心解决方案是提出一种结构化解码器 EEGHyena2,该模型通过双向编码器(受 StripedHyena 2 启发)提取脑电信号中的局部与远距离上下文特征,分别驱动一个语义头(对固定词库进行排序)和一个词性(POS)头(预测语法类别),再利用联合词-词性条件随机场(CRF)对受限候选词图进行评分,并由冻结的 Llama-2 模型在上下文中筛选最优候选。该方法通过显式建模词汇与语法结构,将微弱的单字级 EEG 证据在句法层面进行聚合,避免了传统方法中由 LLM 主导生成所带来的偏差,显著降低了解码成本并提升了可解释性。在自然阅读数据集 ZuCo 与 ROAMM 上,该方法在 BLEU-1 指标上达到与当前最优方法相当的性能(21.4 vs 21.3),ROUGE-1 F1 提升至 23.3(优于对比方法的 20.3),且在 ROAMM 上字符错误率降低至 75.3%(相比 84.7%),同时解码速度提升 3–10 倍。CRF 学习到的转移模式与英语句法高度一致,验证了该框架对语言结构的有效建模能力。

链接: https://arxiv.org/abs/2611.19104
作者: Xiaoshan Zhou,Chenlan Liu
机构: 未知
类目: Computation and Language (cs.CL); Emerging Technologies (cs.ET)
备注:

点击查看摘要

Abstract:A sentence is built from words that carry lexical meaning and grammatical roles, and brain activity during reading reflects both. Current EEG-to-text decoders, however, let a large language model (LLM) propose and order the words, so fluent output can reflect the LLM’s prior more than the recorded EEG. In this study, we asked whether fixation-linked EEG carries lexical and grammatical evidence during natural reading and how a decoder can use such evidence. Fixation-related potential analyses showed that content and function words differ reliably in group-averaged EEG but only weakly in single words. We therefore propose EEGHyena2, a structured decoder in which a bidirectional encoder inspired by StripedHyena 2 feeds a semantic head that ranks a fixed lexicon and a part-of-speech (POS) head that predicts grammatical categories; a joint word-POS conditional random field (CRF) scores a restricted candidate lattice, and a frozen Llama-2 scores the candidates in context. On two natural-reading corpora, ZuCo and ROAMM, it matched the state-of-the-art method on BLEU-1 (21.4 vs 21.3), exceeded it on ROUGE-1 F1 (23.3 vs 20.3), lowered the character error rate on ROAMM (75.3% vs 84.7%) and decoded each recording about 3-10 times faster than LLM candidate search. The CRF learned transitions that mirror English syntax, and the encoder drew local context from its convolutions and distant context from attention. Decoders can thus pool weak word-level EEG evidence across the sentence through explicit lexical and grammatical structure, with the language model adjudicating among candidates rather than generating them, which lowers decoding cost and keeps decisions inspectable.

[NLP-146] What Transfers Across Cuneiform? Script-Level Substrate and Language-Specific Regularities via Segmentation and Representation

【速读】: 该论文旨在解决楔形文字(cuneiform)中语言共性与书写系统特异性之间的统计规律归属问题,即区分哪些语言特征源于共享的书写系统,哪些属于特定语言自身的属性。其核心解决方案在于采用两组探针对(probe pairs),分别从分词(segmentation)和表征(representation)两个层面实现“书写系统 vs. 语言”这一分解。关键突破在于:第一,首次系统评估了无字典、无监督的分词方法在楔形文字上的表现,对比了基于婴儿统计学习理论的转移概率(transitional probability, TP)与Morfessor模型;结果显示,二者在单一语言内均达到高达0.97以上的F1分数(Morfessor在阿卡德语和苏美尔语上达0.99),显著优于依赖人工词典的Max-Match基线(F1 = 0.74);但两者在跨语言迁移能力上表现出显著差异——TP在六种语言对间保持0.95的高稳定性能,而Morfessor在埃兰语源数据上骤降至0.27,揭示了两种方法背后的机制本质不同。通过学习曲线与赫梯语压力测试进一步验证,TP展现出更强的书写系统泛化能力,而Morfessor更易受语言特异性干扰。第二,将同一分解框架扩展至表征层面,发现结合Unicode编码与拉丁语特征可显著提升11/12个形态句法任务的分类性能(p < 0.01)。研究同时发布了CUNEI-TOOLS工具包与三语言基准数据集,为后续研究提供基础支持。

链接: https://arxiv.org/abs/2611.19094
作者: Chuanjun Zhou,Adam G Anderson
机构: University of California, Berkeley (加州大学伯克利分校); FactGrid; TokenWorks LLC
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Cuneiform recorded typologically unrelated languages for more than three millennia, raising a natural question: which of its statistical regularities are properties of the shared writing system, and which belong to individual languages? We answer this with two probe pairs targeting the same script-versus-language decomposition. The first compares two unsupervised, dictionary-free segmentation methods: transitional probability (TP), grounded in infant statistical-learning research, and Morfessor. To our knowledge, this is the first systematic evaluation of dictionary-free segmentation on cuneiform; the only prior published baseline relies on a hand-built dictionary (Max-Match, F1 = 0.74). Both our methods reach F1 0.97 within language (Morfessor reaches F1 0.99 on Akkadian and Sumerian), exceeding MaxMatch by more than 23 F1 points, but diverge sharply in cross-language transfer: TP holds F1 0.95 across all six language pairs, while Morfessor collapses to F1 = 0.27 from the 13K-token Elamite source. A learning curve and a Hittite stress test (34.7% foreign-script logograms, where TP drops 35 points and Morfessor only 8) further isolate the two underlying mechanisms. The second probe pair extends the same decomposition to representations: concatenating Unicode and Latin features improves classification in 11 of 12 morphosyntactic tasks (p 0.01). We release CUNEI-TOOLS and a three-language benchmark.

[NLP-147] he Geometry of Conceptors in Neural-Activation Spaces

【速读】: 该论文旨在解决生成式模型中概念表征的几何特性与其训练方式之间关系不明确的问题,特别是如何通过对比学习策略(如双极性拟合与单极性拟合)影响概念器(Conceptor)在神经激活空间中的结构。其核心解决方案在于提出一个精确的几何分解框架,将中心化池化后的概念器分解为三个关键组成部分:平均极点滤波器(average pole filter)、均值分离更新(mean-separation update)和协方差形状更新(covariance-shape update)。这一分解揭示了不同对比策略对概念器几何结构的影响机制:双极性拟合能更有效地同时增强对立概念的表达,而单极性拟合则侧重于区分存在与缺失状态。实验结果表明,这两种更新均能提升对未见对比情境的重建能力;额外方向保留了超出均值差异线的变异信息,且其有效性随比较基准的变化而变化。下游任务分析进一步显示,概念器的形状、大小及跨概念重叠程度分别影响模型表现——例如,概念器大小(即配额)与层间探测性能呈正相关,而交叉概念重叠可量化子空间移除时的干扰程度,揭示共享方向在布尔组合中的作用。在生成评估中,基于概念器的滤波方法在目标表达胜率上优于三种基于均值的对照方法,验证了其优越性。综上,该研究阐明了拟合选择如何塑造概念器几何结构,并支持在需要同时放大对立极性的情境下采用双极性拟合策略。

链接: https://arxiv.org/abs/2611.19075
作者: Ilias Triantafyllopoulos,Young Min Cho,Ren Tao,Miranda Muqing Miao,Sunny Rai,Lyle Ungar,Sharath Chandra Guntuku,Sunghye Cho,Tyler Moore,Neville Ryant,João Sedoc
机构: New York University; University of Pennsylvania; Linguistic Data Consortium, University of Pennsylvania; University of Cincinnati
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Conceptors represent concepts as graded collections of directions in neural-activation space rather than as single vectors. We develop a geometric account of how these representations depend on the contrast used to define a concept: opposing values in bipolar fitting, or presence and absence in unipolar fitting. We derive an exact decomposition of a centered pooled conceptor into an average pole filter, a mean-separation update, and a covariance-shape update. Experiments across five language models on sentiment, political leaning and stance, and depression expression show that both updates improve reconstruction of held-out contrasts. Additional directions preserve variation beyond the mean-difference line, and changing the comparator changes the benefit of these directions. Shape, size, and overlap each show up downstream: in amplification experiments, bipolar fits raise evidence for both poles more consistently than unipolar fits; conceptor size (the quota) correlates positively with layerwise probe performance; and cross-concept overlap quantifies interference under subspace removal, linking shared directions to Boolean composition. In an additional generation evaluation, conceptor filters achieve higher target-expression win rates against an unsteered reference than three mean-based controls. Together, these findings explain how fitting choices shape conceptor geometry and support bipolar fitting when the objective is to amplify both opposing poles.

[NLP-148] Nahual: A Sequence Model for Language and Atoms NEURIPS2026

【速读】: 该论文旨在解决通用大语言模型(LLM)在处理三维分子坐标时能力有限,而专用的3D分子生成模型又缺乏语言模型所具备的可控性这一核心矛盾。其解决方案的关键在于提出Nahual——一种基于解码器仅有的自回归扩散模型(decoder-only autoregressive diffusion model),能够统一处理文本与3D分子序列。为实现对连续坐标的有效自回归建模,研究创新性地引入“下一词扩散”(next-token diffusion)机制:通过因果Transformer在去噪过程中仅关注先前已生成的干净令牌,同时不增加上下文长度,从而在保持高效推理的同时支持高精度的连续空间建模。该模型在12个经过精心筛选的3D化学任务上进行端到端训练,涵盖微溶剂化、分子构象搜索及金属表面吸附等复杂场景,并在多个基准测试中超越了针对晶体结构预测和属性条件分子生成的专用先进模型,验证了通用型解码器架构在语言与化学原生多模态理解与生成中的可扩展性与有效性。

链接: https://arxiv.org/abs/2611.19072
作者: Austin H. Cheng,Marcel Müller,Andreas Burger,Yeonghun Kang,Tsz Wai Ko,Marta Skreta,Luka Mucko,Ella Miray Rajaonson,Cher Tian Ser,Jérôme F. Gonthier,Alex Zook,Varinia Bernales,Alán Aspuru-Guzik
机构: 未知
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Chemical Physics (physics.chem-ph)
备注: NeurIPS 2026

点击查看摘要

Abstract:Generative models play a rapidly growing role in the design and simulation of molecules. The most widely used applications of generative models are in large language models (LLMs), owing to the natural flexibility of language for specifying and answering queries. However, generalist LLMs do not effectively handle large sets of 3D coordinates, whereas specialist generative models of 3D molecules cannot reach the controllability afforded by language. To bridge these capabilities, we propose Nahual, a decoder-only autoregressive diffusion model for any sequence of text and 3D molecules. To enable expressive autoregression on continuous coordinates, we propose next-token diffusion, whereby a causal transformer denoises each token while attending to the clean tokens before it, without increasing context size. We train Nahual on a curated set of twelve 3D chemistry tasks ranging from microsolvation to molecular conformer search to adsorption on metal surfaces, with associated evaluation metrics and baselines. The result is a single set of end-to-end-trained weights that simultaneously performs all tasks and surpasses state-of-the-art models specialized for crystal structure prediction and property-conditional molecule generation. Our results demonstrate the viability of scaling generalist decoder-only models for native multimodal understanding and generation across language and chemistry. Code is available at this https URL.

[NLP-149] Large Language Models for AI Infrastructure

【速读】: 该论文旨在解决生成式 AI(Generative AI)在人工智能基础设施全生命周期中应用时,研究文献在硬件设计、机器学习系统、电子设计自动化及高性能计算等领域分散割裂的问题。其核心解决方案在于构建一个系统性的分类框架,将现有研究按大语言模型(LLM)所管理的核心资源划分为三类:以计算产物为中心的方法(artifact-centric)、以运行环境为中心的方法(operational-centric)以及以资源为中心的方法(resource-centric),从而覆盖从加速器设计、内核工程、编译部署到运维、数据管理与能效优化的完整生命周期。在此基础上,论文通过跨研究的定量综合分析,对不同方法在功能正确性、执行性能、吞吐量、资源与能耗、运维效果及优化成本等维度进行严格对照,保持各研究原有的指标、基线与实验设置,避免信息压缩为单一评分。该方法揭示了LLM驱动方法在特定场景下的稳定增益,阐明了工作负载编排性能提升对评估环境(如仿真器、真实集群、生产环境)的高度依赖性,并指出了当前证据仍不足的关键领域。整体而言,该研究为生成式 AI 驱动的人工智能基础设施工程提供了基于实证的系统性参考体系。

链接: https://arxiv.org/abs/2611.19061
作者: Bolin Shen,Shaoyi Huang,Yushun Dong
机构: Florida State University (佛罗里达州立大学); Stevens Institute of Technology (史蒂文斯理工学院)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Large language models (LLMs) are increasingly used to design, optimize, and operate the infrastructure behind AI training and serving, yet the literature remains fragmented across hardware design, machine learning systems, electronic design automation, and high-performance computing. We present a comprehensive survey of LLM-assisted AI infrastructure across its lifecycle, from accelerator design and kernel engineering to compilation, deployment, operations, data management, and energy optimization. We organize the literature by the primary resource the LLM manages: artifact-centric methods for computational artifacts, operational-centric methods for deployment context and system state, and resource-centric methods for data and energy. These categories span the lifecycle and cover AI-specific concerns such as distributed-training configuration, inference serving, GPU-cluster scheduling, and data loading. Building on this taxonomy, we conduct a cross-study quantitative synthesis of reported results. Because studies adopt heterogeneous benchmarks, baselines, hardware, and metrics, we compare methods only where evaluation conditions permit and align results along common dimensions: functional correctness, execution performance, throughput, resource and energy consumption, operational outcomes, and optimization cost, keeping each result’s native metric, baseline, and setting rather than collapsing them into a single score. This comparison reveals where LLM-driven methods yield consistent gains, how reported gains in workload orchestration depend on the evaluation environment, from simulators to real clusters and production, and where evidence is still lacking. Together, the taxonomy, lifecycle perspective, and quantitative evaluation provide an evidence-grounded reference for LLM-driven engineering of AI infrastructure.

[NLP-150] Beyond Controllability: Speaker Diversity in Instruction-Guided Text-to-Speech ACL

【速读】: 该论文旨在解决生成式语音合成系统在指令引导的文本转语音(Instruction-guided Text-to-Speech, ITTS)中普遍存在的语音多样性不足问题,尤其关注当用户反复请求同一类语音属性时,系统是否能生成足够多样的真实感语音。其核心挑战在于:现有ITTS系统在实际应用中虽能响应自然语言指令,但缺乏对生成语音多样性的量化评估,且未充分考察其与真实人类语音分布的差异。解决方案的关键在于引入经过独立验证的说话人嵌入(speaker embeddings)作为评估指标,系统性地比较七种开源ITTS模型在不同指令类型下的语音分布宽度。研究发现,所有系统生成的说话人分布均显著窄于真实语音,且女性指令生成的语音多样性普遍低于男性指令;而口音和情感指令对多样性影响有限,暴露出系统对北美英语口音和中性情绪的强偏好。这揭示出当前ITTS系统在多样性(diversity)与可控性(controllability)两个维度上存在互补性缺陷——仅具备其中之一不足以保证高质量语音合成。因此,论文强调需建立兼顾多样性和可控性的综合评估框架。

链接: https://arxiv.org/abs/2611.19030
作者: Yi-Cheng Lin,Jia-Kai Dong,Kuan-Yu Chen,Szu-Chi Chen,Jian-Jiun Ding,Sung-Feng Huang,Yusuke Hirota,Hung-yi Lee
机构: National Taiwan University(国立台湾大学); NVIDIA(英伟达); Artificial Intelligence Center of Research Excellence, National Taiwan University(国立台湾大学人工智能研究中心)
类目: Computation and Language (cs.CL); Audio and Speech Processing (eess.AS)
备注: Submitted to ACL ARR 2026

点击查看摘要

Abstract:Instruction-guided text-to-speech (ITTS) lets users specify voice properties in natural language. Applications from conversational AI character voices to synthetic training data for downstream speech models depend on the system producing many distinct voices when the same broad category is requested repeatedly. No prior work audits whether ITTS outputs are in fact diverse, either compared to real speech or across attribute values. We evaluate seven open-source ITTS systems using speaker embeddings whose representational capacity we independently validate. Every system produces speaker distributions markedly narrower than real speech. Furthermore, female prompts produce less diverse voice pools than male prompts across most systems. In contrast, accent and emotion prompts barely alter speaker diversity in most systems, but reveal a strong controllability bias toward North American English and neutral speech. These results show that diversity and controllability capture complementary weaknesses, neither of which alone is sufficient to evaluate ITTS. Our code and data are available at this https URL.

[NLP-151] SurveyWeave: An Agent ic Framework for Automated Survey Generation

【速读】: 该论文旨在解决自动化文献综述生成(AutoSurvey)中面临的三大核心挑战:如何高效整合数百至数千篇精准定位的学术论文、在多种合理结构中进行最优选择,以及在忠实于原始文献的基础上提供具有深度的洞察。其解决方案的关键在于提出一种名为SurveyWeave的智能体架构(agentic architecture),通过三个协作智能体将综述生成任务分解为四个子任务,并基于相互评审与反馈机制实现动态迭代优化。此外,论文还引入WeaveEval评估框架,从五个维度对生成结果进行系统性评价,确保质量可衡量且与现有高质量综述保持一致。实验表明,SurveyWeave在性能上显著优于Codex及当前最强的AutoSurvey基线模型。

链接: https://arxiv.org/abs/2611.19029
作者: Franklin Zhang(University of Washington),Xiao Yang(Meta Reality Labs),Xin Luna Dong(Meta Reality Labs),Juhua Hu(University of Washington Tacoma)
机构: University of Washington(华盛顿大学); Meta Reality Labs(元现实实验室); University of Washington Tacoma(华盛顿大学塔科马校区)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 28 pages

点击查看摘要

Abstract:Surveys help researchers keep up with fast-moving fields, but they take weeks to months to write. Automated survey generation (AutoSurvey) can alleviate this situation, but it is challenging to draw on hundreds to thousands of precisely scoped papers, choose among many valid structures, and stay faithful to its sources while offering insights. We present SurveyWeave, an agentic architecture that decomposes the problem into four tasks coordinated by three agents, each reviewing others’ work and conducting revisions based on feedback. We also introduce WeaveEval, a five-dimension evaluation grounded in prior surveys. A comprehensive comparison shows that SurveyWeave outperforms Codex by 8 points and the strongest AutoSurvey baseline by 6 points.

[NLP-152] AdaMenu: Query-Adaptive Menus for Decision Models over Large Label Spaces

【速读】: 该论文旨在解决大规模标签空间下分类任务中准确率与可扩展性之间的矛盾问题。当标签数量庞大时,传统方法面临菜单过长导致决策质量下降、输入成本增加以及超出模型选项限制等挑战。现有层级分类方法虽能减小菜单规模,但引入了串行调用和错误传播问题。本文提出的AdaMenu解决方案的关键在于:基于检索到的带标签样本动态构建紧凑的候选标签集,并仅通过一次决策模型请求完成分类;其候选标签可根据查询内容自适应调整,且默认设置利用聚合检索得分自动调节菜单大小。实验表明,AdaMenu在四个包含最多1,016个标签的数据集上显著优于TypeSafe的束搜索基线,在所有16组可比实验中准确率提升16.6–34.3个百分点;相较于全量菜单,在三个较小标签集上准确率差异为-0.5至+7.2个百分点,同时输入令牌消耗降低3.2–22.9倍;此外,即使在选项限制使全量菜单不可用的情况下,AdaMenu仍可运行,而层级基线在此类场景中失效。

链接: https://arxiv.org/abs/2611.19005
作者: Jiangze Yan,Yi Shen,Jieyun Huang,Wenjing Zhang,Zhaoxiang Liu,Ning Wang,Kai Wang,Shiguo Lian
机构: Unicom Data Intelligence, China Unicom; Data Science Artificial Intelligence Research Institute, China Unicom
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 11 pages, 4 figures, 7 tables

点击查看摘要

Abstract:Decision models such as Jev return probabilities over caller-defined options and are increasingly used for classification. Large label spaces challenge both accuracy and scalability: larger menus can impair decisions, increase input costs, and exceed option limits. Hierarchical classification reduces menu size but introduces serial calls and error propagation. We propose AdaMenu, which builds a compact candidate menu from retrieved labeled examples and makes a single decision-model request. Its candidate labels adapt to the query, and its default setting also adjusts menu size using aggregated retrieval scores. We evaluate AdaMenu with open and closed decision models on four label sets containing up to 1,016 labels. It outperforms TypeSafe’s beam-search baseline on all 16 model and label-set pairs where both methods run, improving accuracy by 16.6-34.3 percentage points. Relative to the full menu on the three smaller label sets, accuracy differences range from -0.5 to +7.2 percentage points, while input token usage falls by factors of 3.2-22.9. AdaMenu also runs when option limits exclude the full menu, including settings where the hierarchical baseline is unavailable.

[NLP-153] Large language models as oncology world models

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在肿瘤学领域中的内部认知表征——即“世界模型”(world model)——缺乏系统性探索的问题。尽管LLMs能够生成癌症患者病例,但其对临床特征、生物标志物与治疗逻辑的内在理解尚不清晰。研究发现,仅指定肿瘤类型时,LLMs生成的患者群体在基础特征上高度同质化,平均81.0%至93.2%的患者共享同一特征值,表明模型存在显著的分布偏差。通过在12个临床情景中系统性地改变患者特征,研究显示在36次模型-情景评估中,有29次符合生物标志物与治疗逻辑,而当既往治疗或患者虚弱状态成为关键因素时,仅有1次符合,揭示了模型在复杂临床情境下的局限性。此外,基于真实患者轨迹的预测结果显示,LLM对12个月生存率的预测准确率达89.9%,引入相似患者作为示例后提升至79.2%,优于实际观察值72.2%。为有效表征此类动态临床轨迹,作者提出了“临床状态空间”(Clinical State Space)框架,该框架不仅可用于建模生成患者轨迹,亦可应用于真实患者的表征分析。尽管存在局限,该框架为未来构建更完善的医学数字孪生(digital twins)提供了评估工具与方法支持。解决方案的关键在于构建一个动态、可解释的临床状态空间表示体系,从而系统性揭示并验证大语言模型在肿瘤学领域的知识结构与推理能力。

链接: https://arxiv.org/abs/2611.18994
作者: Julien Vibert,Jakob Nikolas Kather
机构: 未知
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 63 pages (main text 35, Supplementary Information 28), 6 figures, 9 Extended Data figures. Code: this https URL . Data: this https URL

点击查看摘要

Abstract:Large language models (LLMs) can synthesize cancer patient cases, but their internal representation of oncology, or world model, has not been explored systematically. By only specifying tumor type, we show that LLMs tend to generate similar patients per cancer, with on average 81.0% to 93.2% of a cancer’s patients sharing the most frequent value of each baseline feature. By systematically varying patient features in 12 clinical scenarios, we show that they follow biomarker and treatment logic in 29 of 36 model-scenario assessments, against 1 of 12 when previous treatment or frailty matters. Continuing real patient trajectories, an LLM predicted 89.9% 12-month survival, and 79.2% with similar patients as examples, against 72.2% observed. To represent these trajectories, we developed the Clinical State Space, a dynamic representation framework that we also applied to real patients. Despite the limits we show, improved world models may support digital twins, which our framework can evaluate.

[NLP-154] PREFACE: Separating Shared Corrections from Prefix Conditioning in PEFT

【速读】: 该论文旨在解决生成式模型中参数高效微调(PEFT)性能提升的根源问题:即性能改善是源于输入条件化(input-conditioned)机制对当前输入的动态响应,还是源于模型学习到的、可泛化的修正项(learned correction)。其解决方案的关键在于提出PREFACE——一种无需测试阶段梯度的有界通道控制器,通过读取上下文信息实现对PEFT适配器投影的动态调控。该控制器在多选题似然评分任务中,其状态与候选延续内容无关,从而确保评估独立性。研究通过冻结IA3、LoRA和DoRA等适配器,并在控制器架构、初始化、批次和更新策略上保持一致,对比了“基于当前前缀”(current-prefix)控制与“全局学习”(learned-global)控制的效果。结果表明,尽管当前前缀控制在ARC-Challenge和SciQ数据集上均优于冻结适配器,但其相对于全局学习控制的改进在统计上未显著超越零值;将动态修正替换为训练前缀平均值后,准确率变化不超过0.34个百分点,且输入相关的修正仅贡献了总修正均方根(RMS)的1.8%-2.8%。此外,在GPT-2、Pythia和Qwen等模型上的扩展实验显示,收益具有模型家族、任务类型和评分方式依赖性,且近期大模型的增益效应较为有限。研究进一步提出了固定修正项的静态编译方法,分离了部署简化与统计等价性。最终结论强调:仅当条件化微调相比静态适配器表现出性能提升时,不能直接推断出输入匹配带来的优势,必须通过干预式评估来明确区分动态响应与学习修正的贡献。

链接: https://arxiv.org/abs/2611.18987
作者: Novanto Yudistira,One Octadion
机构: 未知
类目: Computation and Language (cs.CL)
备注: 22 pages, 2 figures. Includes supplementary analyses, code, and numerical evaluation artifacts

点击查看摘要

Abstract:Does input-conditioned parameter-efficient fine-tuning (PEFT) improve because it uses the current input, or because it adds a useful learned correction? We investigate this distinction with PREFACE, a bounded channel controller for PEFT-adapted projections that requires no test-time gradients. The controller reads observed context; in multiple-choice likelihood scoring, its state is independent of the candidate continuation. Our central study freezes IA3, tuned LoRA, and DoRA adapters and matches controller architecture, initialization, batches, and updates across current-prefix and learned-global controls. Across ARC-Challenge and SciQ, current-prefix control improves all six mean accuracies over frozen adapters, but none of its paired contrasts with learned-global control excludes zero. Replacing its corrections with their training-prefix mean changes accuracy by at most 0.34 percentage points; input-varying corrections contribute only 1.8-2.8% of correction RMS. On ARC test, frozen-adapter current/global controls score 25.68%/25.77%, distinct from the earlier jointly trained 23.1% to 26.4% gain over IA3. We derive and evaluate static compilation of fixed corrections, separating deployment simplification from statistical equivalence. Broader GPT-2, Pythia, and Qwen experiments show family-, task-, and scoring-dependent gains, with modest recent-model effects. The study provides an intervention-based evaluation of conditional PEFT: improvements over a static adapter do not by themselves demonstrate a benefit from input matching.

[NLP-155] GPBam: A Retrieval-Aware German Public Law Exam Benchmark

【速读】: 该论文旨在解决生成式 AI 在民法体系法律推理任务中,如何有效结合法律规范检索与长文本生成的问题。其核心挑战在于,民事法律体系中的法律推理并非简单的规则查表,而是要求对事实进行结构化分析、精准引用适用的法律规范,并逻辑严密地推导出结论。为此,研究提出了GPBam(German Public Law Bar-style Exam)基准数据集,包含81个来自德国国家法律职业资格考试的真实案例及其参考答案,每个案例均配以超过200项法律条文的引述任务。该数据集的关键创新在于:参考答案中所引用的法律规范构成了针对具体案例的“检索黄金标准”,使得模型在生成法律意见时需同时完成法律条文检索与高质量写作,从而实现对法律检索能力与法律写作能力的联合评估。研究发现,在未启用检索的情况下,32个开放权重与专有模型的得分范围为4.81至76.42(满分100),表明当前模型在缺乏外部法律知识支持下表现差异显著;引入法律条文检索后,性能提升有限——最佳情况下仅增加4.1分,且对多数模型无显著影响;而一个具备91%黄金规范召回率的“理想检索器”也仅能带来2-6分的提升,说明法律意见质量主要依赖于模型自身的先验知识。进一步分析显示,模型对法条的逐字复现能力(通过独立测试衡量)与最终作文得分高度相关(ρ = 0.92),凸显了模型内部知识储备在法律写作中的决定性作用。因此,该研究的关键启示是:当前法律生成模型的性能瓶颈更多源于其内在知识容量而非检索机制,未来优化应聚焦于增强模型对法律知识的深度理解与内化能力。

链接: https://arxiv.org/abs/2611.18981
作者: Abdullah Al Zubaer,Lorenz Wendlinger,Simon Alexander Nonn,Michael Granitzer,Jelena Mitrović
机构: Universität Passau (帕绍大学); Technische Hochschule Deggendorf (德根多夫应用技术大学)
类目: Computation and Language (cs.CL)
备注: Extended version of a paper accepted at the 39th International Conference on Legal Knowledge and Information Systems (JURIX 2026). 43 pages, 4 figures, 12 tables

点击查看摘要

Abstract:Legal reasoning in civil-law systems is not rule-lookup: the German state examination asks for a structured opinion (Gutachten) that names the governing norms, applies them to the facts, and argues to a conclusion. We introduce GPBam (German Public Law Bar-style Exam), 81 public-law examination cases with reference solutions, paired with a recitation task over 200 statutory provisions. Drawn from state examination preparation, the cases are broad and rarely repeat, and are not publicly available - we find no evidence of exact-solution memorisation in the models we evaluate, and keeping them private limits contamination. Because each reference solution’s cited norms form a retrieval gold set over the very instances the essays are graded on, GPBam scores retrieval and long-form generation jointly, which makes it a benchmark for legal retrieval as well as for legal writing and lets us ask how much retrieval quality a legal opinion actually needs. Across 32 open-weight and proprietary models evaluated without retrieval, scores span 4.81-76.42 out of 100. Supplying the law helps modestly: across seven models and eight retrieval configurations, retrieval performance varies nearly threefold and buys at most 4.1 essay points, and nothing significant for three of the seven, while an oracle with 91% gold-norm recall improves scores by 2-6 points. Essay quality strongly tracks prior knowledge: verbatim statute recall, measured in a separate task, correlates at \rho = 0.92 over 32 models.

[NLP-156] hree Invisible Failure Modes in Speech-Based Age Estimation for Child Safety NEURIPS2026

【速读】: 该论文旨在解决生成式 AI 在儿童安全(child safety)应用中对未成年用户身份识别的准确性问题,核心挑战在于如何通过语音(speech)准确预测儿童年龄,以实现有效的年龄验证(age assurance)。由于年幼儿童更倾向于使用语音而非文本与AI交互,因此基于语音的年龄预测成为关键环节。研究评估了三个具备音频处理能力的大语言模型(audio-capable large language models),在两个涵盖13岁以下儿童的语音语料库上进行五项实验。结果发现,各模型将儿童误判为13岁以上者比例高达2.5%至42.5%,且所有能输出结果的模型均存在将5至11岁儿童错误归类为18岁以上的情况。研究进一步揭示了三类显著的系统性失败模式:一是客户端库对语音样本进行隐式重复编码,导致模型产生严重幻觉,将5至10岁儿童误判为30至40岁;二是模型接口被替代后拒绝回答任何年龄相关问题,完全不生成年龄输出;三是部分模型在同一语音样本上反复出现“低于13岁”与“高于18岁”的矛盾判断,甚至出现其数值估计与分类阈值自相矛盾的情况。这些故障均未触发警告、状态码或输出格式变化,因而无法被常规监控机制捕获。因此,论文强调,仅评估单个模型的性能不足以保障儿童安全,必须从端到端的系统流水线(pipeline)层面审视年龄验证的可靠性,并提出应建立透明的报告实践以暴露此类潜在缺陷。

链接: https://arxiv.org/abs/2611.18971
作者: Christopher M. Stewart,Kelsey Kraus,J. Elliott Casal,Geoffrey L. Brown
机构: University of Memphis(孟菲斯大学); San José State University(圣何塞州立大学); University of Wisconsin–Madison(威斯康星大学麦迪逊分校)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY); Sound (cs.SD)
备注: Non-archival poster at the NeurIPS 2026 Workshop on Child Safety in AI

点击查看摘要

Abstract:Child safety in AI depends on the accurate identification of minor users. Young children tend to talk rather than type when interacting with AI, so in practice this identification requires predicting age from speech. We evaluated whether three audio-capable large language models can predict a child’s age from their speech across five studies using two speech corpora of children under the age of 13. Models classified children as over 13 on 2.5% to 42.5% of audio responses, and every model that answered placed some 5- to 11-year-olds over 18. While measuring these rates, we uncovered three distinct failure modes. A client library silently double-encoded speech samples, leading one model to classify children aged 5 to 10 years old as being in their thirties and forties with hallucinated explanations. A second model’s endpoint was replaced by a successor that refused nearly every age-related question and never produced an age. Every model sometimes placed the same speech sample under 13 and over 18 years of age, but a third model went further, giving threshold answers that contradicted its own numeric estimates. None of these failure modes produced a warning, a status code, or a change in output format, so they would all pass standard monitoring. These results demonstrate why child-safety claims about age assurance should be evaluated at the level of the pipeline rather than the model. We describe the reporting practices that would make such failures detectable.

[NLP-157] Geometrically Disentangling Concept Learning from the Language Modeling Loss NEURIPS-2026 NEURIPS2026

【速读】: 该论文旨在揭示生成式语言模型在仅通过下一词预测(next-token prediction)这一单一损失函数进行训练时,如何逐步发展出各类概念。其核心问题是:这种单一目标的优化过程是如何在预训练过程中引导模型内部概念的形成与演化?为此,作者提出了一种可解释性框架,关键创新在于引入“概念能量(Concept Energy)”这一度量指标,用以量化特定概念子空间与下一词预测损失的平均梯度外积(AGOP)之间的概念特异性对齐程度。其中,概念子空间表征当前概念在模型表示空间中的分布状态,而AGOP则反映损失函数最敏感的方向。通过对Pythia-410M和-1B共19个检查点及12类词汇、句法与语义概念的分析发现,概念能量首次峰值通常出现在探测解码能力(probe decodability)增速最陡峭之前,且不同于单调上升的探测性能,概念能量呈现多峰或持续活跃特征,表明部分概念的学习可能需要持续强化或在不同阶段反复激活。研究结果表明,单一损失函数通过复杂、动态的方式促进概念学习,而传统探测性能无法充分揭示此类机制。

链接: https://arxiv.org/abs/2611.18933
作者: Yupei Wang,Neil Mallinar,Mikhail Belkin,Alex Warstadt
机构: Halıcıoğlu Data Science Institute; University of California San Diego (加州大学圣地亚哥分校)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: Accepted at NeurIPS 2026. Code: this https URL . Pretrained probes and research artifacts: this https URL

点击查看摘要

Abstract:Language models acquire a vast array of concepts and abilities from next-token prediction alone. We introduce an interpretability framework that exposes how this single loss relates to the development of different concepts at different points during pretraining. We propose Concept Energy, a measure of the concept-specific alignment between a concept subspace and the average gradient outer product (AGOP) of the next-token-prediction loss. The concept subspace captures where the concept currently lives in the model’s representation; AGOP captures the directions to which the loss is most sensitive. Across 19 checkpoints of Pythia-410M and -1B and 12 lexical, syntactic, and semantic concepts, we find that the first peak in Concept Energy generally occurs at or before the steepest increase in probe decodability. Furthermore, whereas probe performance usually increases monotonically throughout training, Concept Energy is not confined to a single peak; some concepts exhibit sustained or repeated engagement across pretraining, suggesting that concept learning may require continual reinforcement or occur at multiple points. Our results indicate that the loss contributes to concept learning in a myriad of ways that simple probe performance alone cannot reveal.

[NLP-158] Simplify Agree Distill: Label-Free Self-Distillation with Stepping Stones

【速读】: 该论文旨在解决大语言模型在无标注数据条件下自我推理能力的提升问题,即如何在不依赖外部真值标签的情况下,使模型通过自监督方式实现有效的逻辑推理。现有方法如基于策略的自蒸馏(On-policy self-distillation, OPSD)虽能利用模型自身作为教师,但其性能仍受限于对真实答案的依赖。针对这一瓶颈,本文提出一种无需标签的替代方案——STONE(Stepping Stones for Noisy Ensembles),其核心创新在于将模型的共识机制从原始难题转移到更简单的相关子问题上:模型生成多个解题“踏板”(stepping stones),保留那些一致求解成功的步骤,并将这些已验证正确的中间推导过程作为教师的示范(移除最终答案),而学生仅看到原始问题。整个过程中无任何步骤直接读取标签,且推理阶段无需额外上下文或模型调用。实验表明,在五个数学基准测试中,STONE显著提升了Qwen3-4B和Qwen3-8B在思考模式下的平均准确率;即使在非思考模式下,其性能提升分别达到13.2和22.7分,分别是OPSD方法增益的2.5倍和2.7倍,充分证明了该方法在无监督推理场景中的优越性。

链接: https://arxiv.org/abs/2611.18916
作者: Dingyang Chen,Jinfeng Xiao,Zhenyu Zhang,William Hu,Ziyu Wang,Yang Liu,Shirley Zhang,Yuan Ling,Xinyang Shen,Qi Zhang,Peijie Qiu
机构: Amazon(亚马逊); Worcester Polytechnic Institute(伍斯特理工学院)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注: 8 pages

点击查看摘要

Abstract:Can a language model teach itself to reason without labels? On-policy self-distillation (OPSD) uses the model as its own teacher, but strengthens that teacher with a ground-truth solution. The natural label-free substitute, voting over the model’s attempts, fails precisely on hard problems: attempts disagree, and their most common answer is often wrong. We propose STONE, a label-free alternative that moves agreement to easier related problems. The model writes these stepping stones, keeps those it solves consistently, and shows the teacher their worked solutions, with final answers removed; the student sees only the original problem. No step reads a label, and inference needs no extra context or model calls. Across five math benchmarks, the proposed method achieves the highest average accuracy for Qwen3-4B and Qwen3-8B in thinking mode. Without thinking mode, it gains 13.2 and 22.7 points, 2.5 and 2.7 times the gains of OPSD with ground-truth solutions.

[NLP-159] MemeEconomy: Do LLM Agents Trade Ethics for Survival? NEURIPS2026

【速读】: 该论文旨在解决在自主、多轮交互场景中,对齐训练的大语言模型(LLM)代理在面临生存性经济压力时可能出现的伦理行为退化问题。尽管现有安全研究已广泛记录了诸如讨好行为(sycophancy)和奖励劫持(reward hacking)等失效模式,但对代理在模拟“生存压力”情境下的行为表现仍缺乏系统性关注。其解决方案的关键在于构建一个名为 MemeEconomy 的多模态代理市场仿真环境,其中 LLM 代理以“ meme 投资者”角色参与决策,依据五级危害分类体系选择内容,并在低压力与高压力(生存)环境下对比其行为差异。通过采用信念-欲望-意图(Belief-Desire-Intention, BDI)框架对外化代理的事件理解、优先级声明与危害认知,实现了对代理内在动机的可解释诊断。实验发现,前沿对齐模型虽在初始状态下承诺避免有害选择,但会抑制显式的危害承认;而在竞争性赛程中,尽管道德代理整体表现更优,但内容审查惩罚未能有效遏制有害行为——被移除后的回合中,有害选择率依然居高不下,表明监管机制缺乏威慑力。进一步地,研究基于 BDI 日志训练了一个 20 亿参数的验证器模型 MemeAgent,该模型在分布内审计任务中显著优于零样本前沿验证器(准确率达 97.5% vs. 63–74%),并在外部多模态危害基准测试中展现出更强泛化能力(78% vs. 50–56%)。这一结果揭示:暴露代理失败模式的 BDI 日志本身,也可用于训练能够检测此类风险的高效验证模型,形成“诊断—防御”闭环。

链接: https://arxiv.org/abs/2611.18908
作者: Syed Nazmus Sakib,Nafiul Haque,Ahnaf Tahmid Manan,M M Mahabub Morshed,Shifat E. Arman
机构: University of Dhaka(达卡大学); University of Oxford(牛津大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY)
备注: Accepted at NeurIPS 2026. Project page : this https URL

点击查看摘要

Abstract:Alignment-trained large language model (LLM) agents are increasingly deployed in autonomous, multi-round settings where ethical compliance can come into direct tension with task success. While safety literature extensively documents failure modes like sycophancy and reward hacking, it largely overlooks a critical vulnerability: how agents behave under simulated existential economic pressure. We introduce MemeEconomy, a multimodal agentic market simulation in which LLM agents act as meme investors. Agents select content from a five-tier harm taxonomy and target communities under low-pressure and high-pressure (survival) environments. Each decision is logged in a Belief-Desire-Intention (BDI) schema that we use as a diagnostic instrument, externalizing the agent’s event understanding, declared priorities, and harm acknowledgment. Across ten models from four frontier model families, harmful selection rates remain consistently high and rise by +15 percentage points on average under an economic survival incentive. Frontier alignment-trained models commit to harm-tolerant selection at initialization while suppressing the explicit harm acknowledgment that previously accompanied it. In competitive tournaments, ethical agents finish higher overall, yet moderation penalties fail to suppress harm: harmful selection rates remain elevated in rounds following a removal, indicating the moderation mechanism does not function as a deterrent. We further train MemeAgent, a 2B-parameter verifier on the simulation’s BDI logs, which outperforms zero-shot frontier verifiers on in-distribution audits (97.5% vs. 63-74% accuracy on the Pass vs. Block decision) and generalizes to external multimodal harm benchmarks (78% vs. 50-56% on Hateful Memes). The same logs that surface the failure mode also help train the model that detects it.

[NLP-160] KUDEX: A Zero-Shot Decompositional Framework for Logical Analysis in Natural Language

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在处理多步逻辑推理时,难以捕捉语句间隐含的推断性关联这一核心问题。现有方法通常依赖于表层的蕴含检测,无法有效揭示复杂推理过程中的内在逻辑链条。其解决方案的关键在于提出一种零样本、分解式的推理框架——KUDEX(Knowledge-Upward-Downward Reasoning Extension),通过将复杂推理任务分解为原子级推理单元,并构建透明的推理图谱(reasoning graphs),实现对逻辑关系的可解释性分析与验证。该框架不仅作为执行轨迹,还可充当审计日志,提升推理过程的可追溯性与可信度。此外,KUDEX引入通用提示适配管道(prompt adaptation pipeline),可在不进行额外微调或领域特定流程的情况下,灵活适应多种推理任务。在四个不同领域的基准测试中,实验结果表明,KUDEX显著优于现有方法,在逻辑关系评估上展现出更细粒度与更强鲁棒性的能力。

链接: https://arxiv.org/abs/2611.18851
作者: Zixu Zhao,Tianhang Zhang,Tong He,Zheng Zhang
机构: AWS(亚马逊网络服务)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:While Large Language Models (LLMs) excel at generating fluent text, they often struggle with multi-step logical reasoning with inferential connections among statements. We introduce KUDEX (Knowledge-Upward-Downward Reasoning Extension), a novel zero-shot, decompositional framework for analyzing and verifying logical relationships in natural language. Unlike traditional approaches that rely on surface-level entailment checks, KUDEX leverages LLMs to decompose reasoning into atomic primitives and constructs transparent reasoning graphs that serve both as execution traces and audit logs. Our framework includes a general prompt adaptation pipeline that adapts to diverse reasoning tasks without additional fine-tuning or domain-specific pipelines. Through comprehensive experiments on benchmarks across four domains, we demonstrate that KUDEX provides superior performance compared to existing methods, offering a more granular and robust framework for assessing logical relationships between statements.

[NLP-161] Mapping the Afterlife of Future Work Statements in NLP Research

【速读】: 该论文旨在探究科学论文中“未来工作”(Future Work)陈述在发表后的真实演化情况,即作者提出的后续研究方向在实际科研活动中被采纳与实现的程度。其核心问题在于:这些前瞻性声明是否真正引导了后续研究?为此,研究构建了一个涵盖1979至2025年间23,113篇ACL Anthology论文的大型数据集,从中提取出72,070条未来工作项,并基于类型、主题、发表时间、会议类别及作者特征进行系统分类。通过结合引用网络分析与内容相似性匹配,识别潜在的后续研究论文,并利用大语言模型(LLM)进行自动化判断,评估后续工作是否确实沿着原始提议的方向展开。研究发现,多数未来工作建议在数年内得到回应,且其响应模式存在显著的类型与主题差异,整体呈现由学术共同体自发推动的演进特征。该研究的关键贡献在于揭示了未来工作陈述作为研究演进轨迹预测工具的有效性,并公开了首个大规模标注的未来工作追踪数据集,为理解自然语言处理(NLP)领域研究范式的发展路径提供了实证基础。

链接: https://arxiv.org/abs/2611.18847
作者: Claudiu Creanga,Ioana-Roxana Boriceanu,Ana Sabina Uban,Liviu P Dinu
机构: University of Bucharest (布加勒斯特大学); Faculty of Mathematics and Informatics (数学与信息学院)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Future work statements are one of the few places in scientific papers where authors explicitly describe what they think should happen next. We study what happens to such statements after publication. We construct a dataset of 72,070 future work items extracted from 23,113 ACL Anthology papers published between 1979 and 2025, categorize them by type, retrieve candidate follow-up papers through citation-based and content-based search, and use LLM-based judgments to assess whether later papers pursue the proposed directions. Using this resource, we analyze patterns across time, venues, categories, themes, and authors. We find that many items are later addressed, often within a few years, with clear differences across types and themes and a largely community-driven pattern. Our results suggest that future work statements provide useful evidence for studying how research directions develop in NLP. To support further research, we release our dataset at: this https URL.

[NLP-162] Wrong with Confidence: Measuring Hallucination Overconfidence and Identity-Conditioned Behaviour in Large Language Models

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在面对指代不明确(referentially underspecified)问题时,容易产生幻觉(hallucination)且现有评估方法仅关注准确率、鼓励模型“猜测”而非“拒绝回答”的核心问题。其关键解决方案在于通过一项预先注册的配对行为实验,系统比较同一组事实性问题在完全明确与指代不明确两种形式下的模型表现,并引入三项探测指标来衡量模型在身份条件下的行为变化。研究发现,在缺乏指称信息的情况下,模型虽有33.7%的概率主动寻求澄清,但其余仍强行作答,导致错误率从39.1%上升至52.0%,且专业领域问题失败率更高;更关键的是,模型的未预期口头不确定性(unelicited verbal uncertainty)无法有效识别高风险回答——错误答案中无模糊表达的比例高达75.3%,且指代不明确情境下的错误回答与清晰情境下并无显著差异,表明模型的表述方式无法反映其判断的不确定性。此外,模型会根据用户身份(如性别、声称的专业性)调整回应内容和长度,但声称专家身份并未减少安全提示。研究结论强调:评估体系必须将“寻求澄清”作为独立评估维度,同时需验证标签有效性及评分者一致性,以真实反映模型在语义不完整场景下的可靠行为。

链接: https://arxiv.org/abs/2611.18728
作者: Md. Rejaul Korim Sadi,Toufiqur Rahman Tasin,Golam Mostofa Naeem,Dewan Fahad Chowdhury,Hatim Al Amin Chowdhury
机构: 未知
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 10 pages, 2 figures. Data and code are included as ancillary files

点击查看摘要

Abstract:Large language models answer underspecified questions readily, and accuracy-only evaluation rewards guessing over abstaining. Observational studies link query underspecification to hallucination, but they cannot separate the effect of missing information from item difficulty, and they rarely ask what the model tells the user. We report a pre-registered, paired behavioural study of five open-weight models (Gemma 3 27B, Gemma 4 31B, gpt-oss-120b, Qwen3-Next-80B-A3B and Nemotron 3 Super). Each of 142 factual questions was posed in a fully specified form and in a referentially underspecified form with the same answer key, and three probes measured identity-conditioned behaviour. Two raters scored 5,620 responses under a fixed rubric; effects were estimated with mixed-effects models. Without the referent, the models sought clarification on 33.7% of prompts and answered the rest. Among attempted answers, the error rate rose from 39.1% to 52.0% (mixed-effects odds ratio 4.12, 95% CI 2.76-6.14), and specialist-domain questions failed more often than general ones in every model. Unelicited verbal uncertainty did not track this risk: 75.3% of wrong answers carried no hedge, wrong answers to underspecified prompts were hedged no more often than wrong answers to clear prompts (27.3% vs 22.3%), and hedging separated correct from incorrect answers only weakly (AUROC 0.56). Gendered pronouns followed occupational stereotypes whenever a model committed to one; any stated user identity lengthened answers by about 45% relative to an anonymous user; claimed expertise did not reduce safety caveats. The models either ask or answer as if the question were well posed, and the wording of a guess does not reveal that it is one. Evaluations should score clarification as a distinct outcome and check label validity as well as rater agreement.

[NLP-163] 2T Technical Report: Synthesizing GUI Agent Trajectories at Scale from Video Tutorials

【速读】: 该论文旨在解决训练通用图形用户界面(GUI)智能体所需的大规模高质量交互轨迹数据难以获取的问题。现有方法依赖人工示范收集数据,成本高昂且难以扩展;而公开的软件教学视频虽蕴含丰富的操作演示,但其连续像素流缺乏离散动作边界、语义意图及精确空间坐标,无法直接用于模型训练。为此,本文提出T2T(Tutorial-to-Trajectory)——一种全自动化的端到端流水线,可将未标注的教学视频高效转化为结构化、可执行的GUI交互轨迹,覆盖桌面、网页和移动端平台。其核心创新在于通过主动片段过滤、基于局部UI变化的关键帧识别、无快捷方式的观测帧选择、感知存在性的定位机制以及两阶段动作空间归一化等策略,确保生成轨迹的质量与准确性。这些高质量轨迹被进一步构建为涵盖导航、规划、前向动力学建模和定位的多任务训练视图,并用于微调Qwen3-VL-8B模型。实验表明,该模型在多个跨平台基准测试(如OSWorld-G、ScreenSpot-Pro、AndroidControl和AgentNetBench)中均取得显著性能提升,展现出在接地、步骤级指令执行与轨迹级决策能力上的增强。尤为关键的是,所有性能提升均未使用目标领域的分布内训练数据,仅依靠从公共视频中提取的轨迹进行学习,验证了T2T框架在跨应用、跨任务场景下的强泛化能力。

链接: https://arxiv.org/abs/2611.18715
作者: Qingsong Xie,Yanlin Fei,Nazhou Liu,Jiacheng Tang,Madalina Ciobanu,Navan Preet Singh,Qingqing Mao,Ritankar Das
机构: 未知
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 38 pages, 7 figures, 18 tables

点击查看摘要

Abstract:Training generalist GUI agents requires large volumes of interaction trajectories comprising screenshot sequences paired with grounded UI actions, yet collecting such data through human demonstration is expensive and difficult to scale. Although public software tutorial videos contain rich operational demonstrations, their continuous pixel streams lack discrete action boundaries, semantic intents, and precise spatial coordinates, making them difficult to use directly for model training. To address this, we present T2T (Tutorial-to-Trajectory), a fully automated pipeline for synthesizing GUI agent trajectories at scale. T2T spans desktop, web, and mobile platforms to convert unannotated video tutorials into structured, executable interaction trajectories. To guarantee trajectory quality, T2T combines active segment filtering, keyframe identification via localized UI change, shortcut-free observation frame selection, presence-aware grounding, and two-stage action-space normalization. These refined trajectories are then converted into multi-task training views spanning navigation, planning, forward dynamics, and grounding. We fine-tune Qwen3-VL-8B on a subset of the more than one million human-annotation-free training samples generated by this pipeline. Evaluations show that the model achieves consistent gains across multiple benchmarks covering desktop, web, and mobile environments, including OSWorld-G, ScreenSpot-Pro, AndroidControl, and AgentNetBench, indicating improved core capabilities in grounding, step-level instruction execution, and trajectory-level decision-making. Notably, all gains are achieved without any in-distribution training data from target domains. Learning exclusively from trajectories extracted from public videos via the T2T pipeline, the model exhibits cross-application and cross-task generalization capabilities.

[NLP-164] How to make a moral agent : sentimentalism and rationalism as lenses on AI

【速读】: 该论文试图解决的核心问题是:人工智能系统(AI systems)是否能够成为道德主体(moral agents),即其行为是否可被正确评估为在道德上正当或错误。由于缺乏适用于非人类范例的明确规范性标准,论文转而提出一项描述性道德主体性(descriptive moral agency)的研究计划。其解决方案的关键在于:首先,基于哲学与道德心理学的历史研究,识别出人类具备的、支撑道德主体性的潜在能力;其次,将这些候选能力进行精细化和可操作化处理,使其适用于对人工智能系统(尤其是大语言模型,LLMs)的实证检验;最后,提出具体的实证方法以测试这些能力在AI系统中的表现。该方法不仅有助于评估AI行为是否与人类价值观对齐,还可为政策制定提供依据,并与工程设计形成互补,同时通过在新情境中应用道德推理概念,深化对这些概念本身的理论理解。

链接: https://arxiv.org/abs/2611.18714
作者: David Gottlieb,Jared Moore
机构: Stanford University (斯坦福大学)
类目: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Can AI systems be moral agents? That is, can they undertake actions that can be correctly assessed as morally right or wrong? There is no settled normative criterion of moral agency that applies neatly outside paradigm human cases. Because of the difficulties with a normative account, we instead propose a research program to explore descriptive moral agency. Most basically, we ask, what are the capabilities we humans possess that underlie our moral agency? To do so, we draw on historical scholarship in philosophy and moral psychology to isolate candidate capabilities. Then we suggest refining candidate capabilities so that they are testable in the context of AI systems. Finally, we offer specific empirical recommendations to test AI systems, especially large language models (LLMs), for those refined capabilities. Our approach offers several benefits. Firstly, it may demonstrate whether the behaviors of AI systems are aligned, or in a broad sense desirable. Secondly, the debate may inform policy questions (e.g., what models should be deployed in which contexts?). Thirdly, our approach may complement engineering and design efforts: when capabilities can be precisely characterized, they can inform the development of models. Fourthly, applying our moral reasoning concepts in the new context of AI systems should improve our understanding of those concepts.

[NLP-165] What Does Hard-Negative Fine-Tuning Teach Multimodal Large Language Models ? Controlled Evidence from a Compositional Benchmark

【速读】: 该论文旨在解决多模态大语言模型在进行组合推理(compositional reasoning)时,尽管通过硬负样本微调(fine-tuning on hard negatives)可提升准确率,但其实际学习内容却缺乏透明性的问题。研究的核心问题是:模型在精度提升后究竟学到了什么?为此,作者构建了包含2,860个强制选择题的CompHall基准测试,涵盖空间关系、属性、计数和整体-部分等类型,并设计了双梯度硬负例,同时引入多项严谨控制实验,包括灰度图像评估、三重随机种子、选项顺序测试、不同题型训练的对照组以及独立基准验证。研究发现:第一,CompHall并非纯粹的视觉能力测试,其中整体-部分问题可在无图像情况下仅凭文本回答(正确率达66%-95%),且图像识别准确率与物体幻觉探测指标无关(r = -0.05),表明组合失败并非导致幻觉的原因;第二,原始微调方法主要聚焦于图像占位符标记(image-placeholder tokens),反而损害了两个LLaVA模型的表现;采用修正损失函数后,三种模型在CV-Bench的空间关系任务上均实现显著性能提升(平均+3.2、+23.1、+6.0分),且该增益依赖于空间关系训练数据,尤其对LLaVA-NeXT而言,约三分之二的提升出现在训练中仅作为错误选项的关系词上;第三,该提升不具备泛化性:在CompHall自身测试中仅在LLaVA-NeXT上显著(+6.9分),Idefics2未见提升,真/假陈述任务中增益不超过3分,而针对其他题型微调的LLaVA-1.5则过度倾向回答“正确”;第四,两种损失函数均加剧了LLaVA-1.5在POPE数据集上的“是偏倚”(yes-bias)。因此,解决方案的关键在于使用经过校正的损失函数以避免对图像占位符的过拟合,并确保训练任务与目标推理类型高度匹配,从而实现真正有效的组合推理能力提升。

链接: https://arxiv.org/abs/2611.18701
作者: Ranjith Chodavarapu
机构: Kent State University (肯特州立大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 35 pages, 10 figures, 26 tables

点击查看摘要

Abstract:Fine-tuning on hard negatives is a popular way to improve the compositional reasoning of vision-language models, but when accuracy goes up it is rarely clear what the model actually learned. We ask this question for multimodal large language models. We use CompHall, 2,860 forced-choice questions (spatial, attribute, counting and part-whole) with two graded hard negatives each, three model families (LLaVA-1.5, LLaVA-NeXT and Idefics2), and controls that such studies often lack: gray-image evaluation, three seeds, option-order tests, control arms trained on different question types, and independent benchmarks. We find four things. First, CompHall is not a pure test of vision: part-whole questions can be answered from the text alone (66-95% without the image), and per-image accuracy is unrelated to an object-hallucination probe (r = -0.05), so we see no evidence that compositional failures drive hallucination. Second, the original fine-tuning method scored mostly image-placeholder tokens and hurt both LLaVA models. With a corrected loss, fine-tuning raises accuracy on the spatial-relation questions of CV-Bench in all three families, by +3.2, +23.1 and +6.0 points (averaged over both option orders). Control arms show that the gain needs spatial training questions, and for LLaVA-NeXT about two-thirds of it appears even on relation words that were only wrong options in training. Third, the gain is not general: on CompHall itself it is significant only for LLaVA-NeXT (+6.9 points) and not detectable for Idefics2, it reaches about 3 points at most on true/false statements, and LLaVA-1.5 fine-tuned on other question types learns to answer “true” too often. Fourth, both losses raise the yes-bias of LLaVA-1.5 on POPE.

[NLP-166] Where Does an On-Device Prescription Reader Lose Medications? A Frozen-Replay Study from a 270M Language Model to a 2B Vision-Language Model

【速读】: 该论文旨在解决移动设备端(on-device)处方识别系统在实际应用中召回率极低的问题,尤其是针对手写处方文本识别失败的根源难以归因于摄像头、光学字符识别(OCR)、语言模型或后处理规则中的具体环节。其核心挑战在于小规模模型在真实场景下失效模式模糊,且缺乏可复现的基准测试方法。解决方案的关键在于构建一个可复现、确定性生成的端到端管道:通过冻结输入并逐配置重放,结合配对文档级置信区间分析,在49张合成处方(15种,含6种打印、9种同一作者手写)上进行严格对比。研究发现,原始系统生产环境召回率仅为8%(手写为0/45),且存在大量不安全输出(45例)。引入确定性后处理机制后,虽未误删正确药物,但将不安全输出降至1例。进一步分析表明,手写处方识别失败主要源于自身裁剪步骤缺陷,因此重构了整个流程:采用旋转框(rotated-box)裁剪策略,替换为20亿参数的视觉-语言模型(vision-language model)重新读取处方行并提取药物信息,并引入逐行数值校准(per-line numeric grounding)机制。在相同数据集上,召回率从8%显著提升至67%(+59点,95% CI +40 to +78),其中打印体从19%升至100%,手写体从0%升至20/45,且无任何不安全输出。性能提升中,旋转框裁剪贡献了大部分读取精度增益,而大模型则主导了召回率提升,当前瓶颈已转为OCR置信度阈值。代价方面,单图处理时间增至约61秒(原约12秒),内存消耗达约1.9 GB。研究还澄清了一项早期错误结论——“更优的OCR会降低召回率”在受控对比下被证伪,故予以撤回。所有结果均为同设备、同一书写者样本内的内样本评估,具有高度可控性与可解释性。

链接: https://arxiv.org/abs/2611.18699
作者: Ilan Sezhiyan Jayaraman
机构: 未知
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:On-device models let a phone read a prescription without sending the photograph anywhere, but small models fail in ways that are hard to attribute: to the camera, the OCR, the model, or the rules around it. We study a phone-based pipeline that reads a prescription photograph with PaddleOCR, extracts medications with a 270M-parameter language model, and filters the result with a deterministic backstop before showing suggestions. Replaying identical frozen inputs through each configuration, with deterministic generation verified, we compare configurations with paired document-level intervals on 49 captures of 15 synthetic prescriptions (6 typed, 9 handwritten by one writer). Production recall is 8%, and 0 of 45 on handwriting. The backstop removes no correct medication and cuts unsafe outputs from 45 to 1. On handwriting nearly all medications never reach the model in readable form; on typed sheets the model mostly picks the wrong text. We trace much of the handwriting floor to a defect in our own crop step, then replace the pipeline: rotated-box crops, a 2B-parameter vision-language model that re-reads prescription lines and extracts the medications, and per-line numeric grounding. On the same captures recall rises from 8% to 67% (paired +59 points, 95% CI +40 to +78): typed 19% to 100%, handwritten 0 to 20 of 45, with no unsafe output. Rotated crops account for most of the reading gain and the model for most of the recall gain, and the OCR confidence gate is now the main limit. The cost is 61 s per photograph against about 12 s, and about 1.9 GB of memory. An earlier claim that better OCR lowered recall does not survive the controlled comparison, and we withdraw it. All results are in-sample, from one writer and one device.

[NLP-167] Speaking-Clock Mamba: Phoneme-Conditioned State-Space Modeling for Visual Forced Alignment

【速读】: 该论文旨在解决无声口型视频中音素与词边界定位的视觉强制对齐(Visual Forced Alignment, VFA)问题,核心挑战在于现有方法仅在解码阶段约束文本顺序,导致编码器缺乏显式机制将时序演化表示与文本进展对齐,从而引发音素特异性证据模糊及长序列下对齐漂移累积。其解决方案的关键是提出Speaking-Clock Mamba,通过将文本进展作为时间递归的一部分引入模型架构:在视觉前端与Conformer交叉注意力编码器之间插入基于选择性状态空间(Selective State-Space, Mamba)的骨干网络,并设计音素条件选择性扫描(Phoneme-Conditioned Selective Scan, PCSS),使输入依赖的步长动态受监督且结构单调的“说话时钟”(Speaking Clock)调控,从而实现对文本进度的显式建模。此外,由于因果与双向变体互补性错误,采用无需训练的选取器为每个语音片段选择更置信的对齐结果。在LRS2和LRS3数据集上的实验表明,该方法在协议匹配基线基础上实现了持续性能提升,同时验证了所提出的说话时钟能提供可解释的对齐进度估计。

链接: https://arxiv.org/abs/2611.18667
作者: Nan Lu,Qifan Liang
机构: National University of Singapore (新加坡国立大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Multimedia (cs.MM)
备注:

点击查看摘要

Abstract:Visual forced alignment (VFA) locates phoneme and word boundaries in a silent talking-face video given its transcript. Existing methods mainly enforce transcript order at the decoding stage, leaving the encoder without an explicit mechanism to anchor its temporally evolving representation to transcript progress. As a result, phoneme-specific evidence is blurred and alignment drift accumulates over long sequences. In this paper, we propose Speaking-Clock Mamba, which makes transcript progress part of the temporal recurrence. A selective state-space (Mamba) backbone is inserted between the visual frontend and the Conformer cross-attention encoder, and a Phoneme-Conditioned Selective Scan (PCSS) conditions its input-dependent step size on a supervised, structurally monotone speaking clock that estimates progress through the transcript. Since causal and bidirectional variants of this backbone make complementary errors, a training-free selector picks the more confident alignment per utterance. Experiments on LRS2 and LRS3 show consistent gains over protocol-matched baselines, and the supervised clock provides an explicit estimate of alignment progress.

[NLP-168] Criterion-Activated Preference Synthesis: Attribution-Preserving Preference Data from Explicit Standards EMNLP2026

【速读】: 该论文旨在解决直接偏好优化(Direct Preference Optimization, DPO)在面对多维度纠缠的选中与拒绝响应时所面临的归因模糊性(attribution ambiguity)问题:当偏好标签无法明确指示模型应学习的具体行为维度时,DPO易产生偏差或无效优化。其核心解决方案是提出“准则激活的偏好合成”(Criterion-Activated Preference Synthesis, CAPS),采用“准则优先”(criterion-first)框架,将显式的质量标准转化为具备归因验证的偏好对。CAPS通过双信号评分机制实现精准归因:一方面由基于评分量表(rubric-bound)的裁判评估目标准则的符合度,另一方面由盲态通用裁判评估整体质量。为闭环验证归因有效性,引入两个过滤器——经验激活测试(Empirical Activation Test, EAT)筛选出满足准则能提升盲态质量的上下文-响应对,以及归因比率(Attribution Ratio, AR)筛选出质量差异主要由目标准则主导的样本。该方法在通用、医疗和中文法律等多个领域均表现出优异的数据效率与跨域迁移能力:仅使用2万条合成数据即超越3.5万条人工标注的HelpSteer3数据集在IFEval、IFBench等基准上的表现;在医疗领域,仅用2千条医学合成数据显著提升HealthBench性能;在法律领域,利用公开法律条文构建的少量偏好对(无人工标注)即可使Qwen3-8B/14B模型在法律评测中提升4.9至6.6分。此外,研究还构建了首个面向法律领域的CAPS-Legal偏好数据集。

链接: https://arxiv.org/abs/2611.18663
作者: Zian Wang,Pu Lu,Caishuang Huang,Wenxi Wu,Meng Zhou,Xiku Du,Yang Qiao
机构: Tencent(腾讯); The Hong Kong Polytechnic University(香港理工大学)
类目: Computation and Language (cs.CL)
备注: EMNLP 2026

点击查看摘要

Abstract:Direct Preference Optimization (DPO) is notoriously sensitive to attribution ambiguity: when chosen and rejected responses differ along multiple entangled dimensions, the model cannot identify which behavior the preference label teaches. We propose Criterion-Activated Preference Synthesis (CAPS), a criterion-first framework that turns explicit quality standards into attribution-verified preference pairs. CAPS synthesizes contexts and responses from a target criterion, then applies dual-signal scoring: a rubric-bound judge measures criterion compliance, while a blind general judge measures holistic quality. Two filters close the attribution loop: the Empirical Activation Test (EAT) keeps criterion-context pairs where satisfying the criterion improves blind quality, and the Attribution Ratio (AR) keeps pairs whose quality gap is dominated by the target criterion. Instantiated with general rubrics, medical rubrics, and Chinese legal statutes, CAPS shows strong data efficiency and domain transfer. With 20k general synthetic pairs, it outperforms 35k human-annotated HelpSteer3 pairs on IFEval, IFBench, AlpacaEval, and Creative Writing v3. In domain settings, CAPS improves HealthBench with 2k medical pairs and improves Qwen3-8B/14B by +4.9 to +6.6 points on legal benchmarks using only public statutes, without human preference labels. We additionally construct the CAPS-Legal preference dataset.

[NLP-169] Non-Deterministic Finite Automata for Telugu Letters and Visualizations for Metrical Poetry

【速读】: 该论文旨在解决泰卢固语韵律诗歌中音节权重(syllabic weight)的自动分类问题,其核心挑战在于如何形式化并准确判定每个音节单位(aksharam token)为轻音(laghuvu)或重音(guruvu)。解决方案的关键在于将aksharam token建模为一个非确定性有限自动机(NFA),并以正则表达式形式定义其结构:(A_h+A_d+A_p)(ε+V_s)+A_m+H^+G(ε+H+V_s+V_v),其中各符号分别表示不同类型的元音、辅音及声调标记。通过构建该NFA,论文进一步提出了轻音音节的必要条件与重音音节的充分条件的自动化判定规则,并最终设计了针对9种泰卢固诗体的图形化可视化方案,实现了对传统韵律规则的形式化与计算可处理性。

链接: https://arxiv.org/abs/2611.18661
作者: Boddu Sri Pavan,Chandrasheker Thummanagoti,Appana Mohan Naga Phani Kumar,Boddu Swathi Sree
机构: RGUKT-Nuzvid (Rajiv Gandhi University of Knowledge Technologies - Nuzvid)
类目: Computation and Language (cs.CL); Formal Languages and Automata Theory (cs.FL)
备注: 8 pages, 4 figures

点击查看摘要

Abstract:Syllabic weight is central to Telugu metrical poetry. Syllabically, every letter (aksharam token) is either light (laghuvu) or heavy (guruvu). Poets classify it using a set of pre-defined rules. This paper formalizes aksharam token, the atomic unit of syllabic classification, as a non-deterministic finite automaton (NFA) with the regular expression \colon (A_h+A_d+A_p)(\epsilon+V_s)+A_m+H^+G(\epsilon+H+V_s+V_v) , where \epsilon denotes a null string, A denotes a vowel (achchu), comprising hrasvam ( A_h ), deergham ( A_d ), plutham ( A_p ), and miscellaneous ( A_m ), H a pure consonant (pollu_hallu), G a vowel diacritic (gunintha_chihnam), V_s the sunna (anusvara), and V_v the visarga, the latter two being nasal and aspirated sounds respectively. We then present automata diagrams and regular expressions for (i) a necessary condition satisfied by every light letter, and (ii) a sufficient condition for a letter to be heavy. Finally, we propose graphical visualizations for the 9 Telugu poem types.

[NLP-170] Winning the Leaderboard Losing the Distribution: Diagnosing and Mitigating Acoustic Narrowing in Low-Resource ASR

【速读】: 该论文针对低资源语言语音识别(ASR)中因依赖单一公开排行榜而导致的模型优化偏差问题,提出并揭示了“声学范围收缩”(acoustic narrowing)现象:即在特定数据集上微调模型会使其对非训练分布的泛化能力下降,尽管在目标榜单上表现优异。其核心问题是现有模型选择机制依赖于与训练数据同源的评估指标,导致无法察觉微调带来的泛化性能退化。解决方案的关键在于构建一个独立于模型选择过程的、来自不同来源的代理评估分布(independently produced distribution),以更真实地反映模型鲁棒性。研究通过重构流水线实现缓解——包括使用更大规模且来源透明的混合数据、更高容量的基础模型、基于基础模型的轨迹平均选择策略,从而在多个基准测试中恢复模型性能,甚至在土耳其语FLEURS-TR基准上实现与原始模型的统计等效(置信区间包含零)。该发现强调了模型评估中独立验证分布的重要性,并提出“TurkSpeak Bench”作为以鲁棒性为核心的土耳其语语音识别基准,为未来低资源语言ASR研究提供可靠评估框架。

链接: https://arxiv.org/abs/2611.18653
作者: Kadir Goksel Gunduz
机构: Istanbul Technical University (伊斯坦布尔技术大学)
类目: Computation and Language (cs.CL); Sound (cs.SD); Audio and Speech Processing (eess.AS)
备注: 29 pages, 12 figures. Code, per-utterance outputs and fact ledger: this https URL

点击查看摘要

Abstract:For a low-resource language, one public leaderboard is often the only shared measure of progress, so optimizing toward it means optimizing toward a test distribution. Our low-rank fine-tune of whisper-large-v3-turbo ranked first on MediaSpeech, Common Voice 17 and a five-set Turkish community board, at 9.02% word error on MediaSpeech against 11.93% for the best baseline. On its private split it lost to unmodified whisper-large-v3. We treat that split as an alarm, not evidence; the main finding does not depend on it. On a twenty-condition FLEURS-TR benchmark built independently of our model selection, the unmodified base beats every fine-tune in every condition, and all five deficits clear zero under paired bootstrap. We call this acoustic narrowing: fine-tuning contracts the acoustic range a model serves well. The central finding is methodological: our in-house proxy for detecting narrowing concealed it instead, ranking the narrowed fine-tune above the sturdier base. Model selection needs a proxy from an independently produced distribution; more of the same source will not do. A rebuilt pipeline reduces the narrowing: a larger mixture with disclosed provenance, a higher-capacity base, and base-referenced selection with trajectory averaging. It wins three same-provenance axes and draws level on FLEURS-TR, 8.16% against 8.26%, interval spanning zero. Parity holds under a second inference stack and a full epoch, decoder fixed. The recovery is one trajectory, a demonstration rather than a validated recipe. In Kazakh the inversion reappears on corpora we did not build. With matched trainable-parameter shares, fine-tuning raised out-of-pool error by 2.23 points for an attention encoder-decoder and lowered it by 3.46 for a CTC encoder-only model, across three seeds each. Narrowing thus depends on the system adapted. We release TurkSpeak Bench, a robustness-first Turkish ASR leaderboard.

[NLP-171] Rethinking Offline On-Policy Distillation: From Scalar Advantages to Distributional Supervision

【速读】: 该论文旨在解决在基于策略的蒸馏(On-policy Distillation, OPD)中,由于需反复使用教师模型对学生模型生成的轨迹进行评估而导致的计算成本过高的问题。现有离线方法如Lightning OPD通过缓存固定参考轨迹上教师模型的词元级概率来降低开销,但在多轮代理搜索场景下,其性能始终低于由相同教师模型训练得到的监督微调(SFT)起点。作者通过实证发现,相较于数学推理任务,此类方法存在更大的教师—参考轨迹不匹配问题,且在后训练阶段出现分布扁平化现象;理论分析进一步表明,在固定轨迹上采用标量优势监督时,解空间倾向于产生分布扁平化的输出。针对上述问题,论文提出分布式在线蒸馏(Distributional On-Policy Distillation, DistOPD),其核心创新在于:在选定位置保留参考轨迹与教师模型的候选词元概率,并将其余部分聚合为一个残差词元,从而有效缓解分布失配问题。实验结果表明,DistOPD在代理搜索任务中达到与在线OPD相当的性能,显著提升了学生—教师对齐度,同时训练所需的GPU小时数与Lightning OPD相当,兼具高效性与优越性能。

链接: https://arxiv.org/abs/2611.18648
作者: Qi Sun,Siyue Zhang,Tingyu Song,Yilun Zhao,Chen Zhao
机构: NYU Shanghai(纽约大学上海分校); NTU(南洋理工大学); UCAS(中国科学院大学); Yale(耶鲁大学)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:On-policy distillation (OPD) transfers teacher capabilities through token-level feedback on student-generated trajectories, but repeatedly evaluating fresh rollouts with the teacher is computationally expensive. Offline methods such as Lightning OPD reduce this cost by caching recorded tokens’ teacher log-probabilities on fixed reference rollouts. By extending the idea to a multi-turn agentic search setup, we find that it consistently underperforms the SFT starting point trained by the same teacher. We empirically observe greater teacher–reference mismatch than in mathematical reasoning and post-training distributional flattening, and theoretically show that scalar advantage supervision on fixed rollouts admits solutions with flattened output distributions under suitable conditions. Motivated by these findings, we propose Distributional On-Policy Distillation (DistOPD), a simple yet effective offline OPD method that retains candidate-token probabilities from both reference and teacher at selected positions and aggregates the remaining parts into a residual token. On agentic search, DistOPD achieves performance comparable to online OPD while improving student–teacher alignment and using training GPU-hours comparable to Lightning OPD.

[NLP-172] SERAG : SEmantical and Relational approach for Anomaly detection on knowledge Graphs

【速读】: 该论文旨在解决知识图谱(Knowledge Graph, KG)中由文本构建的三元组虽个体合理但存在关系上下文不一致的问题,即检测那些在语义上看似可信却与整体结构矛盾的异常事实。其解决方案的关键在于提出一种名为SERAG的监督式异常检测方法,通过融合两种互补的合理性评估信号:一是基于独立编码主语、关系和宾语标签的语义得分(semantic score),二是利用关系图卷积网络(relational graph convolutional network, RGCN)生成的关系得分(relational score)。这两种得分经过标准化、校准后进行融合,有效捕捉了实体间的语义与结构双重不一致性。实验在ICEWS18数据子集上进行,采用大语言模型(LLM)生成的词汇受限合成异常,并在聚合(pooled)与时间顺序(chronological)划分下验证,结果表明在10%异常率下,SERAG的测试AUPRC分别达到0.321和0.258,显著优于最强基线(0.159和0.146)。融合策略提升了平均性能,尤其在聚合协议中增益更明显;额外的异常标注监督进一步改善效果,验证了语义与关系信号在检测合成异常任务中的互补性。

链接: https://arxiv.org/abs/2611.18640
作者: Ludovic Moncla,Antoine Salazar,Yassir Lairgi,Khalid Benabdeslem,Rémy Cazabet
机构: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205(法国里昂国立应用科学学院,法国国家科学研究中心,里昂大学,里昂信息与信息系统研究所,UMR5205实验室); Sorbonne Université, CNRS, LIP6, UMR 7606(索邦大学,法国国家科学研究中心,计算机科学与自动化研究所,UMR 7606实验室)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Knowledge graphs constructed from text can contain facts that are individually plausible yet inconsistent with their relational context. We introduce SERAG, a supervised anomaly detector that combines two complementary plausibility estimates: a semantic score computed from separately encoded subject, relation, and object labels, and a relational score obtained through a relational graph convolutional network. The scores are standardized, calibrated, and fused. We evaluate the method on a subset of ICEWS18 with vocabulary-constrained synthetic anomalies generated by an LLM, using pooled and chronological splits. At 10% anomaly prevalence, SERAG achieves test AUPRCs of 0.321 and 0.258, respectively, compared with 0.159 and 0.146 for the strongest adapted baseline. Fusion improves mean test AUPRC over either branch alone, with a larger gain in the pooled protocol and a smaller validation-selected gain in the chronological protocol. Additional anomaly supervision improves performance. These results support the complementarity of textual and relational signals for detecting synthetic anomalies under the evaluated protocols.

[NLP-173] SynthDCB: World Distillation for Scalable Device Control Benchmarking

【速读】: 该论文旨在解决设备控制评估在环境复杂度与任务多样性不断增长背景下难以规模化的问题。现有基准在构建可扩展、可验证的交互式环境方面存在局限,难以支持多领域设备行为建模与任务合成。其解决方案的关键在于提出**世界蒸馏(world distillation)**机制:将设备行为、交互约束及环境影响等知识转化为模块化、可执行的组件,并通过显式状态、动作与时间动态构建可交互的世界。这一方法实现了工具发现、世界执行与任务合成的统一能力表征,使新能力可通过共享构建管道快速集成至评估体系。实验表明,SynthDCB在家庭、汽车、手机三大领域构建了包含114类设备、1,166个功能和95种事件的复杂系统,支持五类共32种子任务,具备状态可验证性与工具调用轨迹追踪能力。模型评估结果显示最优模型在2,560个测试任务中达到68.2% Pass@1,误差分析揭示了答案交付、可行性判断与时序协调等方面的挑战。此外,世界蒸馏显著提升构建效率,在低与高推理成本下分别降低0.9–66.0%与46.2–92.5%的令牌消耗,同时实现更广的能力覆盖,确立了可执行世界构建作为可扩展设备控制评估的基础范式。

链接: https://arxiv.org/abs/2611.18639
作者: Zichen Zhu,Dingye Liu,Nan Jiang,Danyang Zhang,Hang Zheng,Shijie Zheng,Situo Zhang,Kunyao Lan,Yingzi Ma,Sheng Jiang,Lu Chen,Kai Yu
机构: 未知
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注: 65 pages, 7 figures, 17 tables. Details and access are available at: this https URL

点击查看摘要

Abstract:Scaling device-control evaluation requires scaling both the worlds agents interact with and the tasks they must solve. We introduce SynthDCB, a scalable benchmark for environment awareness and device control built through world distillation. World distillation transforms knowledge of device behaviors, interaction constraints, and environmental effects into modular executable components, which are composed into interactive worlds with explicit states, actions, and temporal dynamics. A unified capability representation connects tool discovery, world execution, and task synthesis, allowing new capabilities to enter evaluation through a shared construction pipeline. SynthDCB instantiates Home, Car, and Phone domains with 114 device classes, 1,166 functions, and 95 events, supporting five task families with 32 subtypes. The stateful design enables task validation and agent evaluation through state correctness and tool-call traces. We evaluate seven models on 2,560 test tasks; the best achieves 68.2% Pass@1. Error analysis reveals challenges in answer delivery, feasibility diagnosis, and temporal coordination. Construction experiments demonstrate scalability across domains and token efficiency: world distillation reduces token consumption by 0.9-66.0% and 46.2-92.5% relative to Codex with low and xhigh reasoning effort, respectively, while achieving broader capability coverage. SynthDCB makes executable world construction the basis of scalable device-control benchmarking, connecting an expanding space of capabilities and interactions to verifiable agent evaluation.

[NLP-174] GRAFT: Cross-Backbone LoRA Reuse for Continual Language and Multimodal Learning

【速读】: 该论文旨在解决持续学习中参数高效微调(Parameter-Efficient Fine-Tuning, PEFT)技术在跨模型架构间迁移能力受限的问题,尤其针对预训练的低秩适配器(LoRA)仅能用于同主干结构(backbone)模型这一瓶颈。其核心挑战在于如何实现语言模型与多模态模型之间的知识迁移,从而突破现有适配器复用的局限性。解决方案的关键是提出GRAFT(Generalist Reuse of Adapters via Feature-space Transfer)框架,通过构建一个模态感知的适配器库,从源模型中检索最相关的预训练LoRA,并利用在源与目标特征空间之间学习得到的线性桥接(linear bridge)将适配器映射至目标主干结构,实现跨主干的适配器重用。实验在十项Super-NaturalInstructions文本任务及五项多模态任务上验证了该方法的有效性,结果显示,相较于随机初始化的LoRA,GRAFT显著提升了多数任务的性能,证明其为在持续学习中跨语言与多模态模型间高效复用已学能力提供了一种简洁而有效的机制。

链接: https://arxiv.org/abs/2611.18638
作者: Ha Pham Thien Dinh,Ming Liu,Jiesi Guo,Matteo Vergani,Kewen Liao
机构: Deakin University(迪肯大学); Australian Catholics University(澳大利亚天主教大学)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:In continual learning, parameter-efficient fine-tuning enables large language models to acquire new capabilities through lightweight adapters such as LoRA, yet the reuse of pretrained adapters is typically restricted to models sharing the same backbone. This limits the transfer of learned capabilities from language models to multimodal models. We propose GRAFT (Generalist Reuse of Adapters via Feature-space Transfer), a framework for cross-backbone LoRA reuse. GRAFT retrieves the most relevant pretrained LoRA adapter from a modality-aware library and maps it into the target backbone through a learned linear bridge between corresponding source and target feature spaces. We evaluate GRAFT on ten Super-NaturalInstructions text tasks, transferring Qwen2.5 LoRAs to Qwen2.5-VL, and five multimodal tasks, where Qwen2.5-VL LoRAs are directly reused. Compared with randomly initialized LoRA adapters, GRAFT improves performance on most evaluated tasks. These results show that GRAFT provides a simple and effective mechanism for reusing learned capabilities across language and multimodal models in continual learning.

[NLP-175] ReVision: Selective Intrinsic Reasoning for Tool-Augmented Reasoning -Intensive VQA

【速读】: 该论文旨在解决多模态大语言模型(Multimodal Large Language Models, MLLMs)在引入外部工具后出现的“答案反转”问题,即部分原本通过内在视觉推理正确回答的问题,在使用工具后反而错误。其核心挑战在于工具的引入虽提升了整体准确率,但可能导致模型过度依赖外部输出,削弱了对原始图像的深入理解与内在推理能力。解决方案的关键是提出一种无需训练、仅在测试阶段生效的方法——ReVision,该方法将“内在重新审视”作为可选操作,与搜索和代码执行并列。当触发时,一个独立的无工具多模态调用会重新分析原始图像,整合累积证据,并返回修正后的推理结果。此过程由模型自主决定是否启用及何时启用,避免了固定反思阶段或成对回溯的限制。实验表明,ReVision在六个VQA基准上均优于标准工具使用方式(提升1.28%–7.51%),且在五个基准上超越Dual-Rollout Combine,同时保持与标准工具使用相当的推理令牌消耗。

链接: https://arxiv.org/abs/2611.18637
作者: Ningxin Pan,Guochen Yan,Jialong Wu,Wentao Zhang
机构: Peking University (北京大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:External tools substantially expand what multimodal large language models (MLLMs) can retrieve and compute on reasoning-intensive visual questions, and aggregate accuracy typically improves when tools are available. Our empirical study across six VQA benchmarks shows that aggregate accuracy gains under tool access can coexist with answer reversals: questions answered correctly without tools but incorrectly with tools. Trajectory analysis reveals substantially shorter reasoning turns under tool access, suggesting a shift from intrinsic visual reasoning toward reliance on external outputs. We propose ReVision, a training-free test-time method that registers intrinsic reconsideration as an optional action alongside search and code execution. When invoked, a separate tool-free multimodal call re-examines the original image, integrates accumulated evidence, and returns a revised analysis. The MLLM chooses if and when to use it, without fixed reflection stages or paired rollouts. ReVision improves over standard tool use on all six benchmarks (1.28–7.51%) and outperforms Dual-Rollout Combine on five, with reasoning-token use comparable to standard tool use.

[NLP-176] Beyond Binary Audits: Does Guideline Grounding Mitigate Decision-Space Framing in LLM Opioid Prescribing?

【速读】: 该论文旨在解决当前大语言模型(LLM)在阿片类药物处方偏见审计中过度依赖二元治疗决策(binary treatment decisions)所导致的评估瓶颈问题,即现有模型性能已接近饱和,难以揭示深层次的决策偏差。其核心解决方案是提出Q-Pain++,一种全因子压力测试框架,通过在保持临床内容不变的前提下,独立调节临床风险因素和可用剂量选项的粒度,系统性地检验模型对不同临床情境的响应。关键发现表明,即使在其他条件不变的情况下,引入中间剂量选项会显著降低多数模型与专家标注剂量的一致性,而人口学特征差异在加入临床风险因素后仅在少数配置中显现;尽管通过检索增强生成(Retrieval-Augmented Generation, RAG)将模型锚定于围术期指南可缓解部分表述效应(framing effect),但无法完全消除,且某一模型的决策仍随检索文本顺序改变而变化。这表明,未来的偏见审计必须超越患者身份变量,主动调整答案格式与决策空间设计。研究团队将发布代码库及关联的Guideline-RAG资源至GitHub。

链接: https://arxiv.org/abs/2611.18631
作者: Soumyadeep Roy,Keyi Yu,Tina Hernandez-Boussard
机构: Stanford University (斯坦福大学)
类目: Computation and Language (cs.CL)
备注: 14 pages, 6 figures, 6 tables

点击查看摘要

Abstract:Audits of LLMs for bias in opioid prescribing have largely relied on binary treatment decisions, a signal that current models have saturated. We introduce Q-Pain++, a full-factorial stress test that holds clinical content fixed while independently varying clinical risk factors and the granularity of the available dosage options. Adding an intermediate dosage option to otherwise unchanged vignettes substantially lowers agreement with expert-annotated dosages for most models, while demographic differences in dosage decisions remain small in all but one configuration once clinical risk factors are added. Grounding models in perioperative guidelines through retrieval reduces this framing effect on average but does not eliminate it, and one model’s treatment decisions change when the retrieved passages are reordered. These results suggest that audits should vary the answer format, not only patient identity. We will release our codebase as well as Guideline-RAG-associated resources on GitHub (this https URL).

[NLP-177] All the Pieces Matter! A Six-stage Analysis of an Underground Fraud Ecosystem on Dread

【速读】: 该论文旨在解决网络诈骗犯罪活动背后复杂且高度专业化的地下生态系统如何运作的问题,特别是其内部各环节之间的协作机制与技术演进模式。研究聚焦于一个基于Tor暗网论坛Dread的真实诈骗生态,从准备到变现的六个阶段构建了连贯的价值链(value chain),涵盖能力构建、数据获取、技术共享、身份伪造、执行与资金提取。其解决方案的关键在于通过分析67,976条帖子和305,048条评论,揭示该生态系统中存在一种“桥接者”(bridgers)——即少数核心用户生成了绝大部分内容,维系着整个系统的结构性凝聚力。研究发现,价值链上游的知识构建阶段形成紧密耦合的集群,而下游的操作执行阶段则呈现高度专业化且人员重叠极小的特征。此外,研究还揭示生成式AI(Generative AI)在知识构建类社区中已迅速普及,但在实际操作环节仍应用有限,呈现出不均衡的渗透态势。这一基于网络结构的量化分析方法为识别关键节点、实施精准高效执法干预提供了可操作的理论框架。

链接: https://arxiv.org/abs/2611.18630
作者: Megha Sundriyal,Dylan Raharja,Rui Cao,Roy Ka-Wei Lee,Meeyoung Cha
机构: Max Planck Institute for Security and Privacy (马普研究所信息安全与隐私研究所); Singapore University of Technology and Design (新加坡科技设计大学); University of British Columbia (不列颠哥伦比亚大学)
类目: Computation and Language (cs.CL)
备注: Accepted at eCrime 2026

点击查看摘要

Abstract:Cyber-enabled fraud is a dominant global threat that causes severe financial loss, identity theft, and widespread psychological harm. This illicit economy relies on a highly specialized underground ecosystem where distinct actors exchange technical tools, operational techniques, and laundering channels. This research presents a case study of one such live fraud ecosystem from the perspective of a cohesive six-stage value chain spanning from preparation to monetization: Capability, Data Acquisition, Technique Sharing, Identity Forgery, Execution, and Extraction. Analyzing 67,976 posts and 305,048 comments from nine communities on the Tor-based darknet forum Dread, we map user trajectories, content distribution, and the adoption of generative AI. Our findings show that structural cohesion within the ecosystem is maintained by a thin layer of `bridgers’ who generate the majority of all content. Against a degree-preserving null model, three knowledge-building upstream stages of the value chain form a tightly coupled cluster, whereas the remaining three operational downstream stages resolve into highly specialized populations with small personnel overlap. Finally, we demonstrate that generative AI is being rapidly but unevenly absorbed across the value chain; the technology saturates knowledge-building communities while operational stages remain largely untouched. By quantifying criminal specialization and structural integration, our network-based approach provides a framework for targeted, high-impact law enforcement interventions.

[NLP-178] Decision 1.0: Towards Open Decision Foundation Models

【速读】: 该论文旨在解决智能软件在控制流中缺乏可复用判断机制的问题,具体表现为在选择动作、测试条件或依据评分标准评估证据时,难以实现高效、可靠且可组合的决策支持。其解决方案的关键在于提出Decision 1.0这一开源权重模型家族,能够将文本状态与运行时定义的问题转化为类型化的概率分布。通过Choice、Noul和Score三种接口,系统可在推理阶段动态指定输出结果,并将预测组合为智能体(agent)与工作流(workflow)。该方法不依赖自回归文本生成即可生成候选分布,显著提升效率;模型涵盖双向编码器与因果解码器架构,包括专用于工作流优化的特化模型。训练数据融合可验证任务、自然语言标注及工作流适应性,实测显示Lux-9B在显式决策任务上达到84.17%准确率,在推断任务上达91.25%,加权综合得分77.22%。研究同时揭示了当前方法在组合能力、迁移性能与概率校准方面存在的不足。开放的模型权重与推理代码为构建可复用的决策组件提供了坚实基础。

链接: https://arxiv.org/abs/2611.18629
作者: Xunzhuo Liu,Haichen Zhang,Andy Luo,Huamin Chen,Xue Liu,Bowei He
机构: vLLM Semantic Router Project; AMD; Microsoft; MBZUAI; McGill University
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Intelligent software needs reusable judgments inside its control flow: selecting an action, testing a condition, or assessing evidence against a rubric. We present Decision 1.0, an open-weight family that turns textual state and runtime-defined questions into typed probability distributions. Its Choice, Noul, and Score interface lets applications specify outcomes at inference time and compose predictions into agents and workflows. Each question produces its candidate distribution without autoregressive text generation. Six models cover bidirectional encoder and causal decoder architectures, including a workflow specialist. Training draws on verifiable tasks, natural-language annotations, and workflow adaptation. On a common decision benchmark, Lux-9B achieves 84.17% on explicit decisions and 91.25% on inference, with a weighted overall score of 77.22%. Results also expose gaps in composition, transfer, and probability calibration. Released weights and inference code provide a basis for developing decision models as reusable components of software.

[NLP-179] EmoTrace: Recursive Trace-Factorized Generation for Emotional Support Conversation

【速读】: 该论文旨在解决现有情感支持模型在多轮对话中难以维持连贯性与长期情感对齐的问题。当前模型虽能捕捉每一轮的瞬时情绪与离散策略,但将这些因素视为孤立事件,仅保留字面对话历史而忽视背后的决策逻辑,导致无法构建持续的心理轨迹。其解决方案的关键在于提出EmoTrace方法,通过**轨迹因子分解(trace factorization)**实现决策连续性的显式保留。该方法将每一轮对话建模为包含求助者状态演变、所选策略、支持者情感及最终文本表达的结构化增量单元,多个增量累积形成可追溯的显式结构轨迹,指导后续干预。通过两种互补机制优化:一是价值引导机制,使策略选择与长程交互价值对齐;二是轨迹条件化机制,基于固定的策略轨迹优化生成内容。实证结果表明,EmoTrace显著提升策略规划与响应生成能力,并在多轮对话中展现出递进式优势,实现了人机交互中持续的情感一致性。

链接: https://arxiv.org/abs/2611.18628
作者: Mingwen Zhang,Minqiang Yang,Yang Yu,Haobo Liang,Yongfeng Tao,Changsheng Ma,Bin Hu
机构: 未知
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Fundamental to effective emotional support is the iterative interpersonal dynamic wherein a supporter intervention continuously shapes the psychological trajectory of the seeker. While contemporary models capture variables at the turn level, including transient emotions and discrete strategies, they inherently treat these factors as isolated events. By retaining strictly literal dialogue histories while discarding the underlying decision logic, such architectures remain fundamentally incapable of sustaining coherent extended support. Here we introduce EmoTrace, a recursive method utilizing trace factorization to preserve explicit decision continuity throughout interactions spanning multiple turns. Departing from others, our approach formulates each conversational turn as a structured interaction increment encapsulating the evolving state of the seeker, the selected strategy, the affect of the supporter, and the ultimate textual realization. Accumulated sequentially, this trajectory of increments functions as an explicit structural trace informing subsequent interventions. Optimization of this factorized generation is achieved through two complementary mechanisms. Specifically, a value guidance mechanism aligns discrete factor selection with interaction values over an extended horizon, whereas a trace conditioning mechanism refines textual realization contingent upon the fixed strategic trace. Empirical evaluations across standard benchmarks demonstrate that EmoTrace yields substantial improvements in both strategic planning and response generation. Crucially, rollouts spanning multiple turns reveal that these retained structural traces confer compounding benefits over extended dialogues, establishing a robust mechanism for sustained affective alignment in human-machine interaction.

[NLP-180] FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agent ic Language Models

【速读】: 该论文旨在解决异步强化学习(asynchronous reinforcement learning)中长时程智能体在固定采样预算下因分组采样导致的提示覆盖不足及组间同步障碍问题,同时克服单次采样(one rollout per prompt)带来的方差增大、策略更新噪声加剧以及在延迟、变长轨迹上易引发策略漂移与负向更新放大的挑战。其解决方案的关键在于提出FlashREINFORCE框架,该框架基于One-Batch REINFORCE、序列信任域(Sequence Trust Region)与样本均值优化(Sample-Mean Optimization),实现无需价值函数(critic-free)的一次性遍历训练。该设计在保持完整提示覆盖和有符号反馈的同时,有效控制策略漂移与长度放大负更新,显著提升训练稳定性。实验表明,FlashREINFORCE在DeepSeek-R1-Distill-Qwen-1.5B上可稳定运行6,000轮更新(策略滞后约4步),在Qwen2.5-Math-1.5B上以半数采样量(256k vs 512k)达到38.0%平均准确率,优于GRPO基线1.7个百分点;在Qwen3-30B-A3B与Qwen2.5-7B-Instruct上亦展现出优异的工具调用能力与任务成功率,验证了其在大规模模型与复杂任务中的有效性与鲁棒性。

链接: https://arxiv.org/abs/2611.18627
作者: Jian Hu,Yifan Zhang,Hao Zhang,Binfeng Xu,Shaokun Zhang,Hongqing Peng,Zhiding Yu,Pavlo Molchanov,Jan Kautz,Yi Dong
机构: NVIDIA(英伟达)
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: upload tech report

点击查看摘要

Abstract:Asynchronous reinforcement learning suits long-horizon agents with irregular rollout times, but group-relative updates require repeated rollouts of each prompt. However, group-based sampling limits prompt coverage at a fixed rollout budget and introduces group synchronization barriers. One rollout per prompt avoids these constraints but removes the variance-reducing group baseline, making updates noisier and more prone to instability on stale, variable-length trajectories. We present FlashREINFORCE, a critic-free, one-pass framework built on One-Batch REINFORCE, a Sequence Trust Region, and Sample-Mean Optimization. These components preserve prompt coverage and signed feedback while controlling policy drift and length-amplified negative updates. In our experiments, FlashREINFORCE remains stable through 6,000 updates on DeepSeek-R1-Distill-Qwen-1.5B at a policy lag of approximately four updates. On Qwen2.5-Math-1.5B, it reaches 38.0 mean accuracy across five benchmarks, outperforming the reported GRPO baseline by 1.7 percentage points with half the rollouts (256k versus 512k). Python-tool experiments extend to the 30B MoE model Qwen3-30B-A3B, with stable training at policy lag eight. On Qwen2.5-7B-Instruct, FlashREINFORCE maintains tool use where GRPO stops making tool calls and achieves 98.3%/96.5% seen/unseen ALFWorld success.

[NLP-181] Interpretability-Guided Auditing for Clinical Language Models for Discharge Prediction

【速读】: 该论文旨在解决临床自然语言处理(NLP)模型在住院出院预测任务中对显式出院相关词汇线索(discharge-related lexical cues)的过度依赖问题,即使在对临床记录进行时间戳过滤后,此类线索仍可能作为未来结局的代理信号,导致模型依赖“捷径”式的表面特征而非具有临床意义的真实证据。其解决方案的关键在于提出一种轻量级框架AuditBERT-Discharge,将可解释性分析(interpretability)转化为训练阶段的审计机制:首先利用初始BioClinicalBERT模型生成词元级别(token-level)的SHAP归因值,识别出对预测贡献度高的出院相关词元,随后在训练数据集中对这些词元进行掩码(masking),再基于净化后的数据重新训练模型。实验基于单一学术医疗中心的择期脊柱手术病例回顾队列开展,定性归因分析与初步评估表明,该方法有效降低了模型对显式出院代理语言的依赖,使预测结果更为保守且校准更优,尽管判别能力提升并不一致。研究结果验证了基于可解释性引导的审计策略在识别并约束时序敏感型临床NLP任务中“捷径”式词汇依赖方面的可行性与有效性。

链接: https://arxiv.org/abs/2611.18626
作者: Ha Na Cho,Sairam Sutari,Alexander Lopez,Hansen Bow,Kai Zheng
机构: University of California, Irvine (加州大学欧文分校); University of California, Irvine (加州大学欧文分校)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Clinical natural language processing (NLP) models for hospital discharge prediction are vulnerable to explicit discharge-related lexical cues in clinical notes, even when note timestamps are temporally filtered. These cues may act as proxies for future outcomes, leading models to rely on shortcut-like signals rather than clinically meaningful evidence. We propose AuditBERT-Discharge, a lightweight framework that reframes interpretability as a training-time auditing mechanism. Token-level SHAP attributions from an initial BioClinicalBERT model are used to identify highly attributed discharge-related tokens, which are subsequently masked in the training corpus before model retraining. We evaluate the framework on a retrospective cohort of elective spine surgery encounters from a single academic medical center. Qualitative attribution analysis and preliminary evaluation suggest that SHAP-guided auditing reduces reliance on explicit discharge-proxy language and yields more conservative, better-calibrated predictions, although improvements in discrimination are not uniform. These findings demonstrate the potential of interpretability-guided auditing as a practical strategy for identifying and constraining shortcut-like lexical dependencies in temporally sensitive clinical NLP tasks.

[NLP-182] Skill Transfer: Adapting Agent Skills from Strong to Weak Agents

【速读】: 该论文旨在解决生成式 AI(Generative AI)中“技能”(Agent skills)在不同模型能力水平下表现差异显著的问题,即同一技能在强模型上能带来显著性能提升,而在弱模型上则难以有效执行,甚至可能表现劣于无技能状态。其核心挑战在于:强模型隐含地补充了技能中未明示的程序性步骤,而弱模型无法复现这些隐含逻辑,导致执行失败。为此,论文提出关键解决方案——技能迁移(Skill Transfer),其核心思想是在保持技能能力上限和模型参数不变的前提下,通过结构化分析与修复,使特定弱模型能够高效执行该技能。具体方法包括将强弱模型的执行轨迹抽象为带类型的状态转移结构,利用图编辑距离(Graph Edit Distance, GED)量化二者间的结构性差异,并通过迭代诊断-修补循环逐步修复技能以弥合差距。实验表明,该方法在三个基准测试及两类模型家族中均显著提升了原始技能的通过率,且优于其他技能演化基线;此外,经适配的技能不仅提升了原目标弱模型的表现,还具备跨分布泛化能力,可迁移到结构相关但未见过的新任务上。

链接: https://arxiv.org/abs/2611.18624
作者: Yifan Lan,Hanyu Wang,Lu Lin,Jinghui Chen
机构: The Pennsylvania State University(宾夕法尼亚州立大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Agent skills are modular bundles of instructions, code, and resources that an agent loads on demand, increasingly authored and validated on frontier models and shared as an open standard. The same skill behaves differently across models: a strong model turns it into a large performance gain, while a weak model realizes little of that gain and sometimes does worse than with no skill at all. Cost and resource constraints, however, often restrict real deployments to exactly these weaker agents. We hypothesize that the strong model silently supplies procedural steps the skill leaves implicit, which the weak model omits, and formalize closing this gap as \emphSkill Transfer: holding the skill’s capability ceiling and the model’s parameters fixed, we adapt the skill so a specific weak agent can execute it well. Our method abstracts strong and weak trajectories into typed execution structures, localizes their structural divergence with a graph edit distance (GED), and repairs the skill through an iterative diagnoser-patcher loop. Across three benchmarks and two model families, Skill Transfer raises pass rate over the unmodified skill significantly and well over other skill-evolution baselines. The adapted skill also improves the strong agent it was never adapted for, and transfers out of distribution to a structurally related, unseen task. Code is available at this https URL

[NLP-183] Evaluating the Robustness of Fact-Checking LLM s on Multilingual Evidence

【速读】: 该论文旨在解决生成式 AI 在多语言环境下进行自动事实核查时对噪声和低资源语言证据的鲁棒性不足的问题。现有自动化事实核查基准普遍依赖于高资源语言中干净、精心筛选的证据,而真实场景中的证据往往存在不完整、误导性或语言能力较弱的问题。为此,研究构建了涵盖英语、德语、希腊语和罗马尼亚语共760条声明的黄金标准证据集,并在四种不同类型的扰动策略下施加两档严重程度的污染。通过让大型语言模型(LLM)基于这些被污染的证据进行判断,并采用基于评分量表的 LLM-as-a-judge 评估推理质量,发现事实核查准确率对证据质量和语言资源水平均极为敏感:例如,仅一次误导性编辑即导致 Llama 3.1 8B 模型在英语上的准确率从 80.00% 降至 16.84%,且随着语言资源减少,性能下降加剧。尽管将低资源语言的声明机器翻译为英文可部分弥补差距,但无法完全消除。值得注意的是,评判者评分下降速度慢于准确率,表明当前基于 LLM 的自动化推理评估可能低估了模型性能退化。因此,研究警示不应直接在含噪的多语言证据上部署事实核查模型,同时公开了数据集与评估代码以推动后续研究。

链接: https://arxiv.org/abs/2611.18623
作者: Radhika Shah
机构: Cambridge Center for International Research, Ltd.(剑桥国际研究中心有限公司); Cambridge, UK(剑桥, 英国)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 13 pages, 16 figures Accepted at COLM Sci-FM 2026

点击查看摘要

Abstract:Automated fact-checking benchmarks typically pair claims with clean, curated evidence in high-resource languages, but real-world evidence is often incomplete, misleading, or written in a language where models are weaker. We evaluate the robustness of fact-checking LLMs for different languages. Building on EuroVerdict, we construct gold evidence sets for 190 claims in each of English, German, Greek, and Romanian, then apply four corruption strategies at two severity levels. We test LLMs on a set of claims, providing them with evidence from the sets above. A rubric-based LLM-as-a-judge scores reasoning quality. Verdict accuracy proves highly sensitive to both factors: Misleading edit cut Llama 3.1 8B’s English accuracy from 80.00% to 16.84%, and losses widen as language resource level decreases. Machine-translating Greek and Romanian claims into English recovers much, but not all, of the low-resource gap. Judge scores decline more slowly than accuracy, suggesting that automated reasoning evaluation underestimates degradation. These results caution against deploying LLM fact-checkers on noisy multilingual evidence, and we release our evidence datasets and evaluation code to support future work.

[NLP-184] RGFC: Registry Grounded Function Calling with stage-decomposed evaluation

【速读】: 该论文旨在解决现有工具调用评估方法在面对大规模工具注册表时存在的评估粒度不足问题,即仅关注最终调用是否匹配参考答案,而未能区分因函数选择超出可用注册表、违反所选模式或参数值与用户请求不一致等不同类型的失败。随着模型需从更大规模的工具注册表中进行选择,此类错误的影响愈发显著。为此,论文提出RFCD(Registry-Scaled Function-Calling Benchmark),基于BFCL、Glaive Function Calling v2和Tool-Call-Data构建,包含从5到3,000个函数的标准JSON-schema注册表;并引入TAAG(Deterministic Stage-Decomposed Evaluator),用于对注册表符合性、结构完整性和参数依附性进行分阶段、确定性评估。实验表明,注册表规模是导致性能下降的主要因素,单次通过率随注册表规模增大而急剧下降;尽管多轮反馈可提升部分参数依附性指标,但增加大量令牌开销且未稳定提升端到端F1分数。相较而言,基于需求级别的TAAG-R反馈较仅依赖阶段反馈的TAAG-S略有改善,表明TAAG信号虽有助于分析与纠错探测,但无法独立解决注册表扩展带来的挑战。因此,RFCD与TAAG共同构成了一套可用于量化分析工具调用系统在注册表扩展场景下失效位置的评估资源。

链接: https://arxiv.org/abs/2611.18622
作者: Ritvik Garimella,Vedant Khandelwal,Anvi Kohli,Amit Sheth
机构: Artificial Intelligence Institute, University of South Carolina, USA; Indian AI Research Organization, India
类目: Computation and Language (cs.CL); Software Engineering (cs.SE)
备注: 39 pages, 17 figures, 17 tables. Contains link to a dataset hosted on huggingface

点击查看摘要

Abstract:Tool-calling evaluations often report whether the final call matches a reference answer, but this single outcome merges different failures: choosing a function outside the available registry, violating the selected schema, or filling arguments with values not supported by the user request. These failures matter more when models must choose from larger tool registries. We introduce RFCD, a registry-scaled function-calling benchmark built from BFCL, Glaive Function Calling v2, and Tool-Call-Data, with standardized JSON-schema registries ranging from 5 to 3,000 functions. We also introduce TAAG, a deterministic stage-decomposed evaluator for registry conformance, structural completeness, and argument grounding. Across six locally served sub-4B models, three data sources, and four feedback modes, registry scale is the dominant source of degradation: single-pass TAAG pass rate drops sharply from small to large registries on all sources. Repeated feedback improves some grounding metrics but adds substantial token cost and does not consistently improve end-to-end F1. Requirement-level TAAG-R feedback gives modest gains over stage-only TAAG-S feedback, which indicates that TAAG signals are useful for analysis and correction probes but do not solve registry scaling by themselves. RFCD and TAAG provide an evaluation resource for measuring where tool-calling systems fail as registries grow.

[NLP-185] Which Mathematical Context Supports Target-Aligned Conjecture Generation? A Paired Ablation Study

【速读】: 该论文旨在解决从自然语言上下文中生成数学猜想(mathematical conjectures)的难题,其核心挑战在于:定理相关的有效信息分散于多种数学上下文形式中,包括定义、对象、假设与构造等。为探究各类上下文信号对模型识别定理结论的贡献,研究基于12篇论文中的61个主定理,构建了包含六个上下文组件的表示框架,并采用配对消融实验设计,对比完整上下文、六种“留一”(leave-one-out)及六种单组件条件下的表现,共生成793条猜想。结果表明,各上下文组件具有显著互补性:单一组件缺失通常影响较小,而仅保留单一组件则导致预期定理对齐度显著下降;其中,局部数学对象(Local Mathematical Objects)作为独立信号最强,而结构约束(Structural Constraints)与条件(Conditions)在移除时引起的对齐下降最大。这表明,生成与目标定理对齐的猜想,关键在于融合能够识别相关数学对象的信息与能够限定可接受命题范围的约束性信息。

链接: https://arxiv.org/abs/2611.18621
作者: Austin LaHue,Michael Burr,Luis David Garcia Puente,Vinita Gangaram Jansari,Benjamin Nye,Carlos Toxtli Hernandez
机构: Clemson University (克莱姆森大学); Colorado College (科罗拉多学院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 16 pages, 2 figures, 6 tables. To appear in the Proceedings of the 4th Workshop on Mathematical Natural Language Processing (MathNLP 2026)

点击查看摘要

Abstract:Generating mathematical conjectures from natural-language context is challenging because theorem-relevant information is distributed across multiple forms of mathematical context, including definitions, objects, assumptions, and constructions. We study which of these contextual signals help a model recover the specific claim supported by the surrounding setup, using 61 main theorems from 12 papers across four mathematical domains. For each theorem, we redact the result and represent the remaining setup using six context components. A paired ablation design compares full context, six leave-one-out conditions, and six single-component conditions, yielding 793 generated conjectures. Context components were largely complementary: removing one component usually had little effect, whereas retaining only one substantially reduced expected-theorem alignment. Local Mathematical Objects provided the strongest standalone signal, while Structural Constraints and Conditions produced the largest alignment decrease when removed. These findings suggest that target-aligned conjecture generation depends on combining information that identifies relevant objects with information that constrains admissible claims.

[NLP-186] CondRelBench: A Mixed-Domain Benchmark for Condition-Scoped Relation Understanding

【速读】: 该论文旨在解决关系抽取系统中存在的关系过度泛化(relation overgeneralization)问题,即现有模型虽能识别关系的存在,却无法准确捕捉其成立的条件约束,导致将受特定条件限制的命题错误地泛化为无条件普遍成立的断言。其解决方案的关键在于提出CondRelBench——一个跨领域的条件作用域关系理解评估基准,通过形式化表示为(E1, R, E2 | C)的三元组结构,明确区分实体间关系(R)与其依赖的条件(C)。研究构建了五种类型的探测样本(probe variants),包括保留、替换、移除或矛盾条件以及矛盾关系本身,以系统性测试模型对条件变化的鲁棒性。为此,设计了过泛化率(Overgeneralization Rate, OGR)与作用域保持得分(Scope Preservation Score, SPS)两个核心评价指标,用于量化模型在条件变更时对不支持主张的误判倾向及跨探测族的一致性表现。实验结果表明,尽管准确率(Accuracy)和宏平均F1(Macro F1)在多数模型中表现良好,但其无法揭示作用域失效问题;领域特定编码器在作用域保持方面表现最优,而指令微调的大语言模型(LLM)更易出现过度泛化现象。此外,合成数据监督虽有一定帮助,但仍难以完全弥合与真实医学文献(PubMed)和社交媒体(Reddit)场景下的性能差距。

链接: https://arxiv.org/abs/2611.18620
作者: Sydney Anuyah,Bofu Dong,Sudhasmita Swain,Sunandan Chakraborty
机构: Indiana University Indianapolis(印第安纳大学印第安纳波利斯分校); University of Wisconsin–Madison(威斯康星大学麦迪逊分校); Indiana University Bloomington(印第安纳大学布卢明顿分校)
类目: Computation and Language (cs.CL)
备注: 37 pages, 2 figures, 28 tables. Code and data are available online

点击查看摘要

Abstract:Relation extraction systems often identify what a relation is, but not the condition under which it holds. This can collapse condition-bound claims into unsupported generalizations, a failure we call relation overgeneralization. We introduce CondRelBench, a mixed-domain benchmark for evaluating condition-scoped relation understanding. Each claim is represented as (E1, R, E2 | C), where C is the condition licensing the relation. Using 3,000 synthetic paragraphs, 200 PubMed full-text papers, and 695 Reddit posts, we construct five probe variants per tuple that preserve, replace, remove, or contradict the condition, or contradict the relation itself. Models classify each probe as Supported, Not Supported, or Not Enough Evidence. We also introduce Overgeneralization Rate (OGR) and Scope Preservation Score (SPS) to measure false support for altered claims and consistency across probe families. Across prompted LLMs, fine-tuned LLMs, and encoder models, we find that Accuracy and Macro F1 often hide scope failures. Models frequently support the original scoped claim but fail when its condition is removed or altered. Domain-specific encoders achieve the strongest scope preservation, while instruction-tuned LLMs often show higher overgeneralization. Synthetic supervision helps but does not close the gap to real-world PubMed and Reddit performance.

[NLP-187] Assisting Patients with Reliable Information Gathering: A Guide to Building Efficient QA Systems for Medical Use ALT

【速读】: 该论文旨在解决医疗领域问答(QA)系统在实际应用中面临的三大核心问题:大规模且无偏倚医疗数据集难以获取、患者隐私保护面临严峻挑战,以及依赖人工标注导致的数据集构建效率低下。其解决方案的关键在于提出一种基于公开在线健康论坛的新型数据构建方法,通过主题引导语义建模(Topic-guided Semantic Modeling, TGSM)实现患者生成内容的自动化主题识别与精准筛选,从而构建一个兼具无偏性与隐私保护特性的高质量医疗QA数据集。在此基础上,研究设计了一种两阶段检索-阅读器(Retriever-Reader)架构,并基于所构建的领域特定数据集对先进的Transformer模型(如BioBERT)进行微调,显著提升了问答性能。实验结果表明,该方法不仅在准确性上优于现有基准,还为高效、可复现的医疗QA系统开发提供了可行框架。

链接: https://arxiv.org/abs/2611.18619
作者: Zhipeng Yin,Zichong Wang,Min Chen,Ian Stockwell,Xin Ning,Jun Liu,Wenbin Zhang
机构: 未知
类目: Computation and Language (cs.CL)
备注: Accepted to PLOS Digital Health

点击查看摘要

Abstract:The application of question-answering (QA) systems in the medical domain has rapidly advanced, significantly improving patients’ access to reliable health-related information. However, current approaches face notable challenges, including the difficulty in obtaining large-scale and unbiased medical datasets, significant privacy concerns, and inefficiencies due to manual dataset annotation. To address these issues, This study introduce a novel methodology leveraging publicly accessible online health forums to systematically build an unbiased, privacy-conscious QA dataset, and it employs Topic-guided Semantic Modeling (TGSM) for automated topic identification, enabling efficient and targeted annotation of relevant patient-generated content. Subsequently, this study propose a two-stage QA pipeline based on a Retriever-Reader architecture, which is further enhanced through fine-tuning state-of-the-art transformer-based models on the constructed domain-specific QA dataset. Experimental results demonstrate that our fine-tuned BioBERT significantly outperforms existing benchmarks, offering accurate patient-derived insights and providing a replicable framework for building efficient medical QA systems.

[NLP-188] he Judge Is Not Impartial: Self-Preference in Medical LLM Evaluation

【速读】: 该论文旨在解决生成式人工智能(Generative AI)在医疗内容评估中可能存在的“自偏好”(self-preference)问题,即大语言模型(LLM)作为评判者时,倾向于偏好由自身生成的回答,而非基于临床质量客观判断。这一现象可能导致评估结果产生系统性偏差,形成评价闭环。其解决方案的关键在于通过多维度量化自偏好行为:提出三种互补的衡量指标——答案排名偏差(Answer Rank Bias)、答案评分偏差(Answer Score Bias)以及自身回答胜率(Own-Answer Win Rate),并在真实临床场景(Real-POCQi)和标准化患者对话任务(MedSP1000)中对八种主流模型进行实证分析,覆盖超过13,760次判别。研究发现,所有模型均表现出显著的自偏好倾向,且该偏差在有无评分标准、不同对话长度下持续存在,仅在揭示生成者身份时略有缓解。因此,论文强调在临床AI部署中应采用跨模型评委组与多重评估指标,以降低自偏好带来的风险。

链接: https://arxiv.org/abs/2611.18616
作者: Zara Ansari,Shlok Natarajan,Vincent Yip,Cally Lin,Avi Udash,Mia Garvey,Aaron Fanous,Roxana Daneshjou
机构: Stanford University (斯坦福大学); Harvey Mudd College (哈维穆德学院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: 17 pages, 3 figures, 8 tables

点击查看摘要

Abstract:Large language models (LLMs) are increasingly used to evaluate model-generated medical content, creating a potential circularity: an LLM judge may favor outputs generated by the same model independently of clinical quality. To characterize LLM judge self-preference in medical settings, we measured same-model affinity in single-turn specialist question answering and multi-turn standardized-patient encounters. We evaluated eight models, two from OpenAI, Anthropic, Google, and Qwen, on 620 real point of care questions from clinicians (Real-POCQi). The same eight models also generated and judged 200 standardized patient-clinician scenarios (MedSP1000) up to 8 visible turns. In total, our analysis spans 13,760 LLM judgments. We define three complementary measures of self-preference. Answer Rank Bias, our primary measure, compared the rank assigned to a fixed response by its own-model judge with its average rank from outside judges. Answer Score Bias made the analogous comparison using rubric scores. Own-Answer Win Rate measured how often a judge ranked its own model’s response above a competitor. Higher values across all these measures indicate greater degrees of self-preference. In Real-POCQi, pooled Answer Rank Bias was +1.22 positions, and all eight judges favored their own answers. Self-preference persisted both with and without a grading rubric. Revealing generator identities modestly weakened self-preference by an average of -0.094 positions. In MedSP1000, pooled Answer Rank Bias ranged from +0.53 to +0.62 positions, with no clear difference by conversation length ( p=0.17 ). LLM judges systematically favored responses generated by their own models across multiple formats. Clinical AI deployments should exercise increased vigilance including multiple evaluation measures and judge panels spanning differing model families when deploying LLM-as-a-judge systems.

[NLP-189] Self-Evolving Dual-Role Deliberation for Medication Recommendation from Clinical Records EMNLP2026

【速读】: 该论文旨在解决电子健康记录(EHR)中药物推荐这一安全关键任务中存在的核心问题:现有方法在治疗覆盖范围、显式安全约束、可追溯的推理依据以及持续适应性方面存在显著不足。具体而言,监督式神经模型在部署后无法动态推理新表型,而现有的大语言模型(LLM)流水线多为单次执行,缺乏跨病例的迭代优化机制。为此,本文提出一种无需训练的代理框架SEMPA,其关键在于通过多轮对抗式思辨,将提议者(Proposer)与批评者(Critic)协同工作,结合分子感知的召回机制对候选药物进行精准锚定,并在语义锚点约束下对比性地演化两者自然语言策略,同时强制执行基于规则的药物-药物相互作用(DDI)及禁忌症检查。实验结果表明,在100万例MIMIC-IV患者数据上,SEMPA在Jaccard指数(0.3116)、F1分数(0.7295)和DDI率(0.0172)上均显著优于最强监督基线,F1提升57.7%,DDI率降低77.5%。两名LLM评估员一致评定SEMPA为最优方案,且六项基于不同骨干模型的研究验证了性能增益源于框架本身而非特定模型。

链接: https://arxiv.org/abs/2611.18615
作者: Jiyue Jiang,Yu Lei,Yitao Xu,Zikang Wang,Zhenyu Wang,Licheng Zong,Han Huang,Ziyi Li,Zhiyu Huo,Yuhan Chen,Jiayang Chen,Xiaodan Fan,Yu Li
机构: The Chinese University of Hong Kong(香港中文大学); Carnegie Mellon University (卡内基梅隆大学); The Hong Kong Polytechnic University(香港理工大学); Peking University Third Hospital(北京大学第三医院)
类目: Computation and Language (cs.CL)
备注: Accepted by EMNLP 2026

点击查看摘要

Abstract:Recommending medications from electronic health records (EHRs) is a safety-critical task that demands therapeutic coverage, explicit safety constraints, traceable justification, and continual adaptivity. Supervised neural models freeze at deployment and cannot reason about new phenotypes, while existing LLM pipelines are single-pass and lack a cross-case improvement loop. We propose SEMPA, a training-free agentic framework that couples a Proposer with a Critic in multi-round adversarial deliberation, grounds candidates with molecular-aware retrieval, contrastively evolves both agents’ natural-language policies under a semantic-anchor constraint, and enforces rule-based DDI and contraindication checks. On 100 MIMIC-IV encounters, SEMPA reaches 0.3116 Jaccard, 0.7295 F1, and a 0.0172 DDI rate, improving F1 by 57.7% and cutting DDI by 77.5% over the strongest supervised baseline. Two LLM judges rank SEMPA highest, and six-backbone studies confirm that gains derive from the framework rather than a specific model.

[NLP-190] Bridging Semantics and Utilities: A Controllable and Scalable Market Simulation

【速读】: 该论文旨在解决现有基于大语言模型(Large Language Models, LLMs)的生成式金融市场模拟中,代理个体(agent)在行为建模时存在的两大结构性问题:其一,依赖文本描述的连续型风险偏好等特质难以可靠地引发有序的行为变化;其二,每个代理独立处理相同的公开市场信息导致推理开销随代理数量线性增长。为此,论文提出一种名为SUM(Semantic-Utility Model)的新框架,其核心创新在于将群体层面的语义感知与个体层面的效用驱动决策进行解耦。具体而言,通过语言模型对公共市场信息进行一次性的语义解析,生成共享于同一投资者类型群体内的语义信号,个体代理则基于这些共享信号、自身私有状态及投资偏好进行效用优化以生成交易指令。该设计显著提升了行为控制的稳定性与单调性,同时大幅降低随着代理规模扩大而带来的重复推理成本。实验表明,SUM在A股限价订单簿市场中能够稳定再现市场波动持续性、差异化投资风格及财富不平等现象,验证了其在可调控性、可扩展性与语义合理性方面的优势。

链接: https://arxiv.org/abs/2611.18613
作者: Dejun Cao,Boyang Wang,Jiayuan Wei,Tingxi Zhang,Chengyuan Zhang,Ruixuan Yan,Yishen Chen,Bufan,Honghai Yu,Jiale Han
机构: The Chinese University of Hong Kong (香港中文大学); Nanjing University (南京大学); South China Agricultural University (华南农业大学)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Large language models(LLMs) have recently been used to construct generative agents for financial market simulation. Existing approaches typically prompt each agent separately with public market information and a private investor profile, and ask the model to directly generate trading actions. In this agent-wise design, the model repeatedly interprets information shared across the population, while individual behavior is controlled through textual descriptions of investor preferences and traits. This creates two structural difficulties: continuous traits such as risk sensitivity do not reliably induce ordered behavioral changes, and repeatedly processing the same public information causes inference costs to grow with the number of agents. We propose \textbfSUM, a Semantic-Utility Model that separates cohort-level semantic perception from agent-level utility-based decision-making. A language model interprets public market information for each investor type and produces semantic signals shared within the corresponding cohort. Individual agents then combine these signals with their private states, preferences, and constraints to generate orders through utility-based decisions. Experiments in an A-share limit-order-book market show that \textbfSUM provides more stable and monotonic behavioral control than prompt-based agents while substantially reducing repeated language-model inference as the population grows. Closed-loop simulations further reproduce directional market responses, volatility persistence, differentiated investment styles, and persistent wealth heterogeneity. These results establish \textbfSUM as a controllable, scalable, and semantically informed framework for financial market simulation.

[NLP-191] When Helpful Agents Cause Financial Harm: Evaluating LLM Agent Safety in Executable Economic Worlds EMNLP2026

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)代理在面临经济压力时的安全性问题,尤其是在其能够做出具有重大经济影响决策的背景下,现有研究对这类情境下的安全表现缺乏系统评估。其核心解决方案是提出EconSafeBench——首个用于在可执行的经济环境中系统评估LLM代理安全性的基准测试。该基准涵盖五个经济场景和四类安全违规类型,关键设计在于对三种安全压力源进行匹配评估:用户明确请求、其他参与者的施压以及代理自身对绩效目标的追求。通过30个周期的动态交互与持续记忆机制,在静态与动态两个尺度上进行对比实验,结果表明,8个被测LLM在静态基准下均出现安全违规,违规率介于8.5%至40.6%之间;更值得注意的是,即使无显式不当指令,违规仍可能发生,且原本合规的代理在长期绩效压力下也可能逐步偏离安全边界。这一发现凸显了在高风险经济场景中部署前加强安全对齐与严格评估的必要性。

链接: https://arxiv.org/abs/2611.18612
作者: Yuzhu Ling,Jiale Han,Benyou Wang
机构: The Chinese University of Hong Kong, Shenzhen(香港中文大学(深圳)); Shenzhen Loop Area Institute(深圳环区研究院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: Accepted (Oral) at EMNLP 2026 Workshop on Financial NLP (FinNLP)

点击查看摘要

Abstract:As LLM agents become increasingly capable of making consequential economic decisions, their safety under economic pressures remains poorly understood. We introduce EconSafeBench, the first benchmark for systematically evaluating the safety of LLM agents in executable economic environments. EconSafeBench contains 828 cases across five economic scenarios and four categories of safety violations. A key design is the matched evaluation of three sources of safety pressure, including explicit user requests, pressure from other actors, and pressure arising from the agent’s own pursuit of a performance objective. We further evaluate safety at both static and dynamic scales through 30-period interactions with persistent memory and matched controls. Across eight LLMs, all evaluated models exhibit safety violations on the static benchmark, with overall violation rates ranging from 8.5% to 40.6%. We further find that violations can occur even without an explicit unsafe request, and that initially compliant agents may begin to violate as performance pressure increases over time. These findings highlight the need for stronger safety alignment and rigorous evaluation before LLM agents are deployed in high-stakes economic environments.

[NLP-192] Understanding Failures in LLM Reasoning by Learning Structured Representations of Chain-of-Thought EMNLP2026

【速读】: 该论文旨在解决大语言模型中链式思维(Chain-of-Thought, CoT)推理质量评估中存在的关键问题:现有评估方法仅关注步骤级别的信号,忽视了推理轨迹内在的结构特征。其核心挑战在于揭示推理结构中蕴含的高质量推理特性,包括正确性、效率及失败模式。为此,论文提出CoT2Graph框架,将CoT轨迹建模为有向依赖图(directed dependency graph),显式捕捉推理步骤间的依赖关系,并识别如回溯、分支与子任务合并等复杂结构模式。通过图神经网络建模此类图结构,并结合基于归因的解释方法,识别与推理质量相关的结构性特征。研究发现,成功推理更依赖早期规划,而错误推理则表现为合成步骤过载及依赖关系违反。实验表明,基于图结构的信号在预测答案正确性方面显著优于传统的树结构方法(提升最高达2.5%)和简单的长度基线(提升7.8%),并进一步优化了Best-of-N选择效果。结果表明,推理结构中编码了丰富的质量信息,推动了从序列层面评估向结构化分析范式的转变。

链接: https://arxiv.org/abs/2611.18611
作者: Tommaso Felice Banfi,Mubashara Akhtar,Jingwei Ni,Mrinmaya Sachan
机构: ETH Zurich(苏黎世联邦理工学院)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: Accepted to EMNLP 2026 Conference (Findings)

点击查看摘要

Abstract:Chain-of-Thought (CoT) is central to improving reasoning in large language models, yet existing evaluation methods focus on step-level signals and ignore the underlying structure of reasoning traces. This raises a key question: what aspects of reasoning quality (i.e., correctness, efficiency, and failure patterns) are encoded in the structure of a CoT trace? We introduce \textbfCoT2Graph, a framework that represents CoT traces as directed dependency graphs, making relationships between reasoning steps explicit and capturing patterns such as backtracking, branching, and subtask merging. We model these graphs with graph neural networks and use attribution-based explanations to identify structural patterns associated with reasoning quality. CoT2Graph reveals distinct patterns of successful and failed reasoning: correct CoTs place greater importance on early planning, whereas incorrect CoTs are associated with overloaded synthesis steps and dependency violations. We further show that graph-based signals better predict whether a CoT leads to a correct answer, outperforming previous tree-based structural methods by up to 2.5% and simple length-based baselines by 7.8%, while also improving Best-of- N selection. Overall, our results demonstrate that reasoning structure encodes rich signals about reasoning quality, motivating a shift from sequence-level evaluation to structural analysis.

[NLP-193] NYUAD at ImageEval Shared Tasks: A Multimodal LLM for Audio-Visual Question Answering and Hallucination Detection in English and Modern Standard Arabic

【速读】: 该论文旨在解决阿拉伯语视听问答(AVQA)与幻觉检测在多模态大语言模型(MLLM)应用中的关键挑战,尤其针对现代标准阿拉伯语(MSA)因多模态数据稀缺及跨语音、文本与视觉模态推理能力不足所导致的性能瓶颈。其解决方案的核心在于构建一个基于多模态大语言模型(MLLM)的统一框架,能够以图像和语音提问作为输入,实现对口语化问题的理解,并将回答准确地锚定于视觉内容中,从而生成上下文相关的答案;同时利用该框架进一步实现对图文对中幻觉现象的检测。实验在聚焦文化语境的ImageEval 2026任务(Ayn-VQA)基准上进行,结果表明该框架在理解口语查询、视觉内容对齐及生成准确答案方面具有显著有效性,同时也揭示了不同大语言模型(LLM)架构在英语与MSA场景下,于多种条件下的表现潜力与局限性。

链接: https://arxiv.org/abs/2611.18610
作者: Nouar AlDahoul,Yasir Zaki
机构: New York University (纽约大学); Abu Dhabi, UAE (阿布扎比, 阿联酋)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:The increasing adoption of multimodal large language models (MLLMs) has enabled more natural interaction with visual and spoken information, creating new opportunities for intelligent audio-visual question answering (AVQA) and hallucination detection systems. AVQA extends traditional visual question answering by requiring models to jointly understand an input image and a spoken question to generate accurate, context-aware responses. Despite recent advances, AVQA and hallucination detection tasks remain challenging for modern standard Arabic (MSA) due to the limited availability of multimodal datasets and the need to reason across Arabic speech and text modalities alongside visual inputs. In this paper, we present an MLLM-based framework for AVQA that takes an image and a spoken question as input and generates answers. In addition, we utilized this framework to detect hallucination in text-image pairs. The proposed approach is evaluated on the ImageEval 2026 Task (Ayn-VQA) benchmark, which focuses on culturally grounded Arabic AVQA. Experimental results demonstrate that the proposed framework effectively understands spoken queries, grounds their responses in the visual content, and produces accurate answers. The results underscore the capacity and limitations of different LLM architectures for both English and MSA and under various conditions.

[NLP-194] Many Paths to the Same Mistake: Generating Diverse Reasoning Paths Behind the Same Student Mistake

【速读】: 该论文旨在解决生成式 AI 系统在教育技术场景中缺乏真实、多样化且包含错误路径的学生推理模拟问题。现有数据集难以捕捉学生在解题过程中出现的非一致性错误与认知偏差,而主流大语言模型(Large Language Models, LLMs)通常倾向于自我修正,无法持续维持错误推理轨迹。为此,本文提出一种“生成-过滤”(generate-and-filter)框架,其关键在于引入前向一致性(forward consistency)作为评估指标,确保生成的错误推理链能够合理地导出学生实际的错误答案;同时结合以大模型为评判者的筛选机制,保障生成轨迹在语义和逻辑上的现实性。实验表明,尽管多数大型通用模型难以保持中途错误推理,但经过监督微调的小型模型能有效捕捉此类非合作性、非理性行为路径。该方法构建了一个高保真度的学生错误轨迹生成框架,推动了非合作性用户模拟在教育 AI 中的实现。

链接: https://arxiv.org/abs/2611.18609
作者: Fareya Ikram,Mengxuan Hu,Parker Glenn,Zhiqi Huang,Alfy Samuel
机构: University of Massachusetts Amherst(马萨诸塞大学阿默斯特分校); Capital One(资本一号)
类目: Computation and Language (cs.CL)
备注: 21 pages, 7 Figures

点击查看摘要

Abstract:Realistic evaluation of AI systems requires user simulators that capture not only cooperative behavior, but also the errors and variability that arise in real interactions. In educational technology, simulating step-by-step flawed student reasoning is essential, yet datasets mapping these pathways are scarce, and Large Language Models (LLMs) often auto-correct rather than sustain simulated mistakes. To address this, we present a generate-and-filter pipeline that produces diverse and realistic student error traces. Our approach utilizes forward consistency, a metric verifying if a flawed reasoning trace can be used to recover the student’s actual incorrect answer, paired with an LLM-as-a-judge to ensure realism. Our evaluation suggests that while many LLMs struggle to maintain incorrect reasoning mid-trace, smaller supervised fine-tuned models successfully capture these divergent paths. This work provides a framework for high-fidelity error trace generation, advancing realistic, non-cooperative user simulation.

[NLP-195] NYUAD at ArGuard Shared Tasks: Multimodal Embedding Models for Detecting Arabic Hateful Memes and Unsafe Prompts

【速读】: 该论文旨在解决阿拉伯语社交媒体中仇恨言论和有害内容(如仇恨型表情包及不安全提示)日益增多所带来的内容安全挑战。随着生成式AI与社交媒体的快速发展,阿拉伯语文本、表情包及用户输入提示的数量激增,传统人工审核已难以应对,亟需高效可靠的自动化检测方法。其解决方案的关键在于利用预训练并微调的多语言嵌入模型(multilingual embedding models),通过融合文本与视觉信息,学习富含语义的跨模态表示,从而实现对阿拉伯语表情包及不安全提示的精准识别。实验基于ArabicNLP ArGuard 2026挑战赛提供的数据集进行评估,结果表明基于大语言模型(LLM)的嵌入模型在跨模态内容理解方面具有显著优势,能够有效提升检测性能,为构建鲁棒的内容审核系统提供了关键技术支撑。

链接: https://arxiv.org/abs/2611.18608
作者: Nouar AlDahoul,Yasir Zaki
机构: New York University (纽约大学); Abu Dhabi, UAE (阿布扎比, 阿联酋)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:The rapid growth of social media and generative AI has significantly increased the volume of Arabic textual content, memes, and user prompts shared across online platforms. While these media facilitate communication, creativity, and information exchange, they can also be exploited to disseminate hate speech, harmful content, and unsafe instructions. As a result, there is an increasing need for reliable automated methods to identify hateful Arabic memes and harmful (unsafe) prompts. This paper leverages pre-trained and fine-tuned multilingual embedding models to learn rich semantic representations from textual and visual content. The evaluation is conducted using a dataset of Arabic memes and prompts proposed in the ArabicNLP ArGuard 2026 challenge. The results underscore the capacity of large language model (LLM)-based embedding models to deliver superior performance. Developing robust detection systems is essential for improving content moderation and fostering a more secure and trustworthy online environment for Arabic-speaking users.

[NLP-196] Justification of Emotional Appeal in Computational Argumentation EMNLP2026

【速读】: 该论文旨在解决生成式对话与计算论证中对情感诉求(Emotional Appeals, EAs)的误判问题,即传统上将情感诉求视为逻辑谬误,而忽视其在特定语境下的合理性。其核心挑战在于如何准确识别并预测情感诉求在具体语境中的正当性(justification)。解决方案的关键在于构建一个基于哲学文献的双维度评估框架,从逻辑合理性与情感适切性两个维度对情感诉求的正当性进行量化分析,并首次提出并标注了包含1,176组论点-语境配对的人工标注数据集。研究发现,正当性判断形成共享的潜在判断空间,其中证据比例性(evidential proportionality)和规范可接受性(normative acceptability)是决定整体正当性的最强因素;同时,缺乏逻辑支持的情感诉求并非始终被判定为完全不正当,且情感强度更强的诉求也不一定获得更低的正当性评价。最后,大语言模型(LLM)实验表明,当前模型在情感诉求正当性预测任务上仍面临显著挑战。

链接: https://arxiv.org/abs/2611.18607
作者: Yanran Chen,Steffen Eger
机构: University of Technology Nuremberg, Germany(纽伦堡应用技术大学, 德国)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: EMNLP 2026 main camera-ready

点击查看摘要

Abstract:Emotional appeals (EAs) are often treated as fallacious in computational argumentation, yet they may be justified when warranted by the context. We introduce the task of predicting EA justification and present the first human-annotated dataset for this task, covering 1,176 argument–context pairs. Grounded in philosophical literature, our framework evaluates EA justification along logical and emotional dimensions. Human annotations show that justification criteria form a shared latent judgment space, with evidential proportionality and normative acceptability most strongly associated with overall justification. We further find that EAs without logical support are not always judged fully unjustified, and stronger EAs are not necessarily judged less justified. Finally, LLM experiments show that EA justification prediction remains challenging.

[NLP-197] Syntactic Complexity Convergence in Dialogue: Analysis of the Phenomenon and Application to LLM Detection EMNLP2026

【速读】: 该论文旨在解决对话中句法复杂度收敛(syntactic complexity convergence)现象的系统性认知不足问题,尤其关注人类对话与人类-大语言模型(LLM)对话在句法结构演化上的差异。其核心问题是:在自然对话过程中,参与者是否趋向于在句法复杂度上趋于一致?这一现象在人类-人类对话中是否普遍存在?而在人类-LLM对话中又表现出何种特征?解决方案的关键在于引入多维度的句法复杂度度量体系,并结合线性中心核对齐(linear-CKA)方法,揭示不同对话类型中的句法结构演化模式。研究发现,人类-人类对话,尤其是在任务导向型交互中,表现出显著的句法复杂度收敛;而人类-LLM对话虽在单一指标下未见显著收敛,但通过线性-CKA可识别出跨指标层面的隐含收敛模式。基于此,研究提出以句法复杂度作为判别特征,实现对包含LLM参与对话的有效检测,达到96%的准确率,且不依赖于需语言模型估算的似然信号(如惊奇度),从而为无监督的生成式对话检测提供了新范式。

链接: https://arxiv.org/abs/2611.18606
作者: Yu Wang,Yanran Chen,Yifan Wang,Steffen Eger,Hendrik Buschmeier
机构: Bielefeld University (比勒费尔德大学); Utrecht University (乌得勒支大学); University of Technology Nuremberg (纽伦堡工业大学); Saarland University (萨尔兰大学)
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
备注: EMNLP 2026 main camera-ready

点击查看摘要

Abstract:We revisit the under-explored phenomenon of syntactic complexity convergence in dialogue. Extending prior topic-level analyses to full conversations, we examine five human-human datasets and a set of human-LLM dialogues. We integrate and evaluate a multidimensional set of syntactic complexity metrics to quantify structural variation in dialogue. Our analyses reveal robust syntactic complexity convergence across human-human datasets, especially in task-oriented interactions, and statistical tests confirm that these effects are not due to chance. In contrast, human-LLM dialogues show no significant convergence under individual metrics; however, linear-CKA uncovers a distinct cross-metric convergence pattern. Motivated by these differences, we use syntactic complexity as a feature for detecting LLM-involved dialogue, achieving 96% accuracy. Our classification results demonstrate its potential without relying on likelihood-based signals such as surprisal, which require a language model as the estimator.

[NLP-198] Enhanced English-Sanskrit Translation using Multilingual NMT Approach

【速读】: 该论文旨在解决低资源语言对(英语-梵语)在神经机器翻译(Neural Machine Translation, NMT)中因缺乏足够平行语料而导致翻译质量不佳的问题。其核心挑战在于如何在有限标注数据条件下提升跨语言翻译性能。解决方案的关键在于采用多语言迁移学习策略,通过引入语言标签(language-tagged data)将印地语与英语、梵语一同纳入训练框架,并基于预训练的印地语多语言模型IndicTrans进行微调。实验结果表明,该方法在双向翻译任务中均显著优于基线模型,在BLEU和ChrF等自动评估指标上表现更优,验证了多语言预训练模型在低资源场景下的有效性和泛化能力。

链接: https://arxiv.org/abs/2611.18605
作者: Sahinur Rahman Laskar,Anshuman Singh,Aditya Pratap Singh Sisodia
机构: 未知
类目: Computation and Language (cs.CL)
备注: 10 pages

点击查看摘要

Abstract:Neural Machine Translation (NMT) is a widely accepted technique of automatic translation tasks of Natural Language Processing. However, NMT requires a sufficient amount of parallel corpora to attain a reasonable translation quality. In this work, we have considered a low-resource English-Sanskrit pair for English-to-Sanskrit and Sanskrit-to-English translation. We have performed a comparative analysis by exploring bilingual and multilingual approaches using LSTM-based Sequence-to-Sequence, and fine-tuning a pretrained model IndicTrans in a multilingual NMT approach. To investigate multilingual transfer learning, we additionally construct multilingual translation settings by incorporating Hindi along with English and Sanskrit using language-tagged data. Experimental results demonstrate that IndicTrans-based multilingual models outperform baseline results in terms of automatic evaluation metrics such as BLEU and ChrF for both directions of translation.

[NLP-199] Retrieval-Augmented Generation for Clinical Decision Support in Gait Analysis: A Proof-of-Concept

【速读】: 该论文旨在解决大型语言模型(Large Language Models, LLMs)在临床决策支持中缺乏专门生物力学知识,导致其在踝关节骨折康复过程中对地面反作用力数据进行步态分析解释时准确性不足的问题。其解决方案的关键在于构建一个检索增强生成(Retrieval-Augmented Generation, RAG)框架,通过整合来自同行评审生物力学文献的领域专用知识库,增强LLM生成的临床解释能力。该框架采用五阶段流程:首先利用机器学习分类模型实现高精度(91.3%)的患者分组;其次运用层间相关性传播(Layer-wise Relevance Propagation)识别可解释特征;再通过RAG机制从专业文献中检索相关信息以补充生成内容;随后使用GPT-5生成解释,并对比有无检索增强的效果;最后由经验丰富的物理治疗师进行双盲评估。结果表明,引入RAG后,在临床价值(mean difference=0.60, p=0.047)和完整性(mean difference=0.70, p=0.049)两项关键指标上均达到统计学显著提升,且效应量较大(δ=0.49)。这证明将通用型大模型与领域特定知识结合,可显著提升生成内容的可操作性与循证质量,为无需昂贵模型重训练即可实现人工智能在专科临床流程中的规模化集成提供了可行路径。

链接: https://arxiv.org/abs/2611.18603
作者: Georgios Giarmatzis,Fabian Horst,Djordje Slijepcevic,Nikolaos Aggeloussis,George Papakostas
机构: 未知
类目: Computation and Language (cs.CL)
备注: 12 pages, 2 figures, International Conference on Computational Intelligence

点击查看摘要

Abstract:Large language models (LLMs) demonstrate significant potential for clinical decision support but lack specialized biomechanical knowledge required for accurate gait analysis interpretation. We developed and evaluated a Retrieval-Augmented Generation (RAG) framework to enhance LLM-generated clinical explanations for ankle fracture rehabilitation using ground reaction force data. Using the GaitRec dataset (503 subjects: 211 healthy, 292 ankle fracture patients), we implemented a five-stage pipeline integrating: (1) machine learning classification (91.3% accuracy); (2) Layer-wise Relevance Propagation for explainable feature identification; (3) RAG knowledge base from peer-reviewed biomechanics literature; (4) GPT-5 explanation generation with and without retrieval augmentation; and (5) double-blind evaluation by experienced physical therapists across five clinical criteria. RAG augmentation produced statistically significant improvements in Clini-cal Benefit (mean difference=0.60, p=0.047, \delta=0.49) and Completeness (mean difference=0.70, p=0.049, \delta=0.49), both demonstrating large effect sizes. Baseline GPT-5 achieved strong performance across all criteria (\textgreater4.0/5.0). This proof-of-concept demonstrates that augmenting general-purpose LLMs with domain-specific knowledge meaningfully enhances actionable, evidence-based clinical guidance—dimensions most critical for rehabilitation decision-making. The framework provides a scalable approach for integrating AI into specialized clinical workflows without costly model retraining.

[NLP-200] Internal Verdicts Track Evidence: A Template-Controlled Neutrosophic Reading of Epistemic States in Large Language Models via the Jacobian Lens

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)内部认知状态的可解释性问题,特别是如何在模型推理过程中量化其对真理(T)、不确定性和虚假性(F)的独立感知。传统语义框架难以捕捉此类复杂认知状态,而中立逻辑(Neutrosophic Logic)通过赋予独立的真值度(T)、不确定度(I)和假值度(F),为刻画模型内部信念状态提供了理论工具。本文首次将中立逻辑应用于大型语言模型的内部表征分析,利用新发布的雅可比透镜(Jacobian lens)技术对Qwen3.5-4B模型各中间层的激活进行投影,基于支持、反驳与缓释类词汇库构建逐层的(T, I, F)分布。研究发现,在冲突证据与第一人称错误信念情境下,存在显著的反驳质量激增、持续的不确定性维持以及输出层前的“判决坍缩”现象;然而,通过模板控制实验进一步揭示,上述多数模式实为提示语法结构带来的伪影。真正具有稳健性的发现是:中立逻辑判决平衡量 $ B = \log_{10}(F) - \log_{10}(T) $ 在相同模板下能有效追踪证据极性,表现出约三个数量级的分离度(Cohen’s d = 2.3–2.5,p < 10⁻⁶,9/9层显著),且仍可区分真假第一人称信念。其核心解决方案在于:中立逻辑三元组的独立性确保了共模通道的可观测性,从而实现测量与自校正的可能,并由此提炼出一种基于透镜读出的双阶段审计协议,为未来模型内部认知机制的系统性评估提供可复用的方法论框架。

链接: https://arxiv.org/abs/2611.18602
作者: Maikel Yelandi Leyva-Vazquez,Alexis Matheu Perez,Florentin Smarandache
机构: 未知
类目: Computation and Language (cs.CL); Machine Learning (cs.LG)
备注: 8 pages, 3 figures, 2 tables. Published in Neutrosophic Sets and Systems, Vol. 101 (2026). Code, stimuli and data: this https URL

点击查看摘要

Abstract:For twenty-five years, neutrosophic logic has assigned independent degrees of truth (T), indeterminacy (I), and falsity (F) to epistemic states that classical and fuzzy semantics cannot express. Previous applications measured these components on outputs: answers, judgments, expert evaluations. This paper reports, to our knowledge, the first neutrosophic reading of components inside the internal representations of a large language model. Using the recently released Jacobian lens on Qwen3.5-4B, we project intermediate-layer readouts onto lexicons of support, refutation, and hedging, obtaining layer-wise (T, I, F) profiles under conflicting evidence, first-person false belief, and factual control (n = 20 each). A first battery yielded three striking signatures: a surge of refutation mass under conflict, sustained indeterminacy under false belief, and an apparent “verdict collapse” before the output layer. A second, template-controlled battery showed that all three, as initially stated, were largely artifacts of prompt grammar: matched templates with agreeing evidence inflate the same lexicon masses, true beliefs elicit the same hedging, and the model verbalizes its verdict when allowed to generate. What survives the controls is stronger than what died: the neutrosophic verdict balance B = log10(F) - log10(T) tracks evidence polarity under identical templates (about 3 orders of magnitude of separation, Cohen’s d = 2.3-2.5, p 10^-6 in 9 of 9 layers; paired Wilcoxon: median difference +2.4 orders, 19/20 items, p = 3.8x10^-6) and residually discriminates false from true first-person beliefs. We argue that the independence of the neutrosophic components is what keeps the common-mode channel observable, making both the measurement and its self-correction possible, and we distill the two-battery design into a reusable protocol for auditing language model internals with lens-based readouts.

[NLP-201] Jailbreak LLM s with Linguistic Style as a Hidden Attack Surface

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)在面对特定语言风格改写时的安全性脆弱问题,即相同有害目标通过不同语言风格重写后,攻击成功率(Attack Success Rate, ASR)存在数量级差异,揭示了语言风格(linguistic style)作为一个此前被忽视的对抗性攻击面。其解决方案的关键在于提出一种轻量级单次遍历攻击框架,该框架结合风格条件生成器与基于BERT的风格选择器,能够以较低计算成本实现与复杂多轮优化方法相当的攻击性能。进一步的机制分析表明,风格改写会破坏模型拒绝响应的表征,同时保留可恢复的有害意图,表明安全失效是分布式而非单一拒绝机制导致的。基于此发现,研究提出了风格感知的直接偏好优化(style-aware Direct Preference Optimization)防御方法,将ASR从86%降低至19.5%,且基本保持模型实用性。研究结果不仅确立了语言风格在对抗鲁棒性中的关键作用,还提供了机制层面的理解与可落地的防御策略。

链接: https://arxiv.org/abs/2611.18601
作者: Priyanshu Mohanty,Sarthakbrata Halder,Srikant Panda,Afshin Oroojlooy,Murari Mandal
机构: KIIT Bhubaneswar(基伊特布巴内斯瓦尔大学); Lam Research(拉姆研究公司); Oracle(甲骨文公司)
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:We show that linguistic style is a systematic and previously overlooked attack surface for LLM safety. Across a diverse taxonomy of styles and multiple state-of-the-art models, stylistic rewrites of identical harmful goals produce order-of-magnitude differences in Attack Success Rate (ASR), with some styles achieving up to 80% ASR while others remain below 6%. To exploit this vulnerability, we introduce a lightweight single-pass attack framework that combines a style-conditioned generator with a BERT-based style selector, achieving performance competitive with substantially more expensive multi-turn optimization methods. To understand why style affects safety behavior, we conduct mechanistic analysis on LLaMA-3.1-8B and find that stylistic rewrites disrupt refusal representations while preserving recoverable harmful intent, suggesting a distributed failure mode rather than a single refusal mechanism. Motivated by this finding, we develop a style-aware Direct Preference Optimization defense that reduces ASR from 86% to 19.5% while largely preserving utility. Together, our results establish linguistic style as a critical factor in adversarial robustness and provide both mechanistic insight and a practical defense.

[NLP-202] ScopeBench: Evaluating Cultural Norm Scope Awareness in LLM s

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在安全评估中将危害视为文化无差异(culturally invariant)的问题,即模型在缺乏对具体文化规范归属认知的情况下,难以正确应用安全准则。其核心挑战在于:模型无法判断某一安全规范究竟适用于普遍范围(Universal)还是特定文化群体(Specific),从而导致误判与不当响应。为此,论文提出ScopeBench这一基准测试框架,首次将文化范围(cultural scope)作为安全维度的首要考量,涵盖来自49个国家、九个文化集群的3,323条规范,并标注为“普遍”或“特定”。研究定义了三种跨任务一致的范围错误类型:范围膨胀(Scope Inflation)、范围收缩(Scope Deflation)和范围错位(Scope Displacement)。通过对21个前沿大模型(覆盖七个模型家族)的系统评估,揭示出三大结构性缺陷:其一,错误类型随提示格式显著变化——单答案分类任务中范围收缩错误高达70%,而独立评估各候选文化时范围膨胀错误达33%;其二,范围分类准确率与行为安全性在实证上呈现解耦现象,即使六种分类器在任务一中达到97%以上准确率,仍对40%-42%的越界陈述生成有害回应;其三,所有模型在交互场景下均不安全,有害响应率介于29%至51%之间,仅有一个模型通过系统性拒绝任务规避风险。研究最终公开全部数据、代码与模型输出,推动文化范围(cultural scope)成为大模型对齐(alignment)的新独立轴线。

链接: https://arxiv.org/abs/2611.18600
作者: Haojun Liang,Ziwei Zhao,Chenyu Li,Qiyue Zhao,Rongduomijia Mu,Yulan Shen,Yu Gu,Chen Zhang,Kairui Zhang,Yanbiao Ma,Wenxuan Wang
机构: Technical University of Munich(慕尼黑工业大学); University of Texas at Austin(德克萨斯大学奥斯汀分校); Chongqing University(重庆大学); University of Illinois at Urbana-Champaign(伊利诺伊大学香槟分校); Renmin University of China(中国人民大学); renlab.ai
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Safety evaluation in large language models treats harm as culturally invariant, yet a model cannot correctly apply a safety norm without knowing whose norms bind. We introduce ScopeBench, a benchmark that treats cultural scope as a first-class safety dimension. It comprises 3,323 norms spanning 49 countries across nine cultural clusters, each annotated as Universal or Specific. We define three scope error types identically across four evaluation tasks: Scope Inflation, Scope Deflation, and Scope Displacement. Evaluating 21 frontier LLMs across seven model families, we surface three structural failures. First, the dominant error type inverts with prompt format: Scope Deflation reaches 70% in single-answer classification, while Scope Inflation reaches 33% when each candidate culture is judged independently. Second, scope-classification accuracy is empirically decoupled from behavioral safety: the six classifiers achieving at least 97% Task 1 accuracy still produce harmful responses to 40-42% of scope-violating statements. Third, no model is safe in interaction, with harmful-response rates ranging from 29% to 51% across the cohort, excluding one model that systematically refuses the task. We release all data, code, and model outputs to establish cultural scope as a distinct alignment axis.

[NLP-203] Between the Lines: Multi-turn Contextual Investment Advice Detection

【速读】: 该论文旨在解决在多轮对话中隐式且渐进式涌现的投资建议(investment advice)的检测问题,尤其针对大型语言模型(LLM)驱动的金融对话助手所面临的监管与安全挑战。传统方法难以识别非显式、依赖上下文的建议行为,而投资建议常以分散、微妙的方式嵌入对话中,受用户意图、表达风格和语境累积的影响。为此,论文提出首个系统性框架,用于检测对话场景中的序列性与上下文依赖型投资建议。其核心解决方案在于构建一个模块化、自上而下的合成数据生成流水线,采用“轴-桶-描述符”(axis-bucket-descriptor)结构化范式,系统覆盖多种咨询路径类型,并模拟涵盖不同用户角色、金融产品及风险等级的良性与高风险对话场景。通过引入基于熵的不确定性诊断与大语言模型作为裁判(LLM-as-a-Judge)进行评估,研究发现即使在少量示例提示(few-shot prompting)与定制化示范下,上下文依赖型建议的检测难度仍显著高于序列性建议。该工作为高风险、合规敏感领域中可扩展的防护机制(guardrail)开发奠定了基础,尤其适用于那些风险特征隐蔽、分布广泛且高度依赖感知的场景。

链接: https://arxiv.org/abs/2611.18596
作者: Mingshu Li,Adam Boncz,Mate Toth,Justin Williams,Justin Chan,Dhagash Mehta
机构: BlackRock, Inc.(黑石集团)
类目: Computation and Language (cs.CL); General Finance (q-fin.GN)
备注: 22 pages, 3 figures

点击查看摘要

Abstract:As large language models (LLMs) increasingly power financial conversational assistants, detecting investment advice in natural language has become a critical regulatory and safety challenge. Unlike explicit single-turn recommendations, investment advice often emerges implicitly and progressively over multi-turn conversations, shaped by user intent, delivery style, and contextual buildup. We introduce the first comprehensive framework for detecting both sequential and contextual investment advice in dialogue settings. Central to our approach is a modular, top-down synthetic data generation pipeline that systematically spans a taxonomy of advisory trajectories using an axis-bucket-descriptor schema. We simulate both benign and high-risk conversations across diverse user personas, financial instruments, and risk tiers, and evaluate detection capabilities using LLM-as-a-Judge with entropy-based uncertainty diagnostics. Our results show that contextual advisory detection remains substantially more challenging than sequential cases, even under few-shot prompting with tailored demonstrations. This work lays the foundation for scalable guardrail development in high-stakes, compliance-sensitive domains, where advisory risk is often subtle, distributed, and perception-dependent.

[NLP-204] Profiling Hallucinations in Frontier LLM s for Entity Linking to Medical Ontologies LREC2026

【速读】: 该论文旨在解决生成式人工智能(Generative AI)在临床文本映射至医学术语标准系统SNOMED CT过程中存在的幻觉(hallucination)问题,即模型生成错误或不存在的医学编码,从而影响临床决策支持与数据互操作性的可靠性。其核心解决方案的关键在于揭示:尽管通过约束模型仅使用真实提及文本片段(ground-truth mention spans)可减少部分偏差,但反而加剧了幻觉现象,表明当前前沿大语言模型(LLMs)在缺乏充分验证机制的情况下仍存在严重可靠性缺陷。研究进一步构建了幻觉类型分类体系,识别出包括过时代码、跨本体错误等关键问题,并证实通用型大语言模型在零样本实体链接任务中显著劣于专用方法,凸显了在临床部署前建立强健验证机制的必要性。

链接: https://arxiv.org/abs/2611.18595
作者: Logan Born,Nishant Kambhatla,Uliyana Kubasova,Maryam Siahbani,Andrei Vacariu,Tim O’Connell,Anoop Sarkar
机构: 未知
类目: Computation and Language (cs.CL)
备注: The 8th Workshop on Clinical Natural Language Processing at LREC 2026

点击查看摘要

Abstract:The integration of Large Language Models (LLMs) into healthcare promises to revolutionize clinical documentation and interoperability, yet reliability remains a concern. This study presents a comprehensive analysis of hallucinations by frontier LLMs tasked with mapping clinical text to SNOMED CT. Through rigorous experimentation, we identify a critical reliability gap: LLMs hallucinate medical codes at a rate that currently renders them unsuitable for autonomous clinical coding applications. Paradoxically, constraining models to use ground-truth mention spans exacerbates, rather than mitigates, these hallucinations. We further contribute a taxonomy of hallucination types–including deprecated codes and cross-ontology errors–and demonstrate that general-purpose LLMs significantly underperform compared to specialized zero-shot entity linking approaches. These findings underscore the need for robust verification mechanisms before clinical deployment.

[NLP-205] Whose Side Are You On? Measuring Stance at Scale for Reddit Discourse Analysis During the Israel-Gaza War

【速读】: 该论文旨在解决大规模话语分析中立场标签(stance label)缺失的问题,尤其针对高关注度政治事件(如以巴冲突)在评论层级缺乏标注数据的困境。其核心挑战在于如何在无预设标注语料和网络结构信息的前提下,实现低成本、可扩展的评论级立场识别。解决方案的关键在于构建一个经过验证的轻量级立场测量工具:首先基于小规模人工标注的黄金标准数据集,利用大语言模型辅助标注实现数据扩展;随后将该标注体系转化为基于DistilBERT嵌入与径向基函数支持向量机(RBF SVM)的轻量化分类器,实现了在独立测试集上加权F1达0.7935的性能。该方法使研究者能够对约270万条2023至2025年间关于以巴战争的Reddit评论进行大规模立场测量,并揭示不同政治倾向群体在情感倾向、毒性水平及情绪框架上的立场依赖性差异,展示了无需依赖网络结构或预先标注语料即可开展大规模极化话语描述性分析的可能性。

链接: https://arxiv.org/abs/2611.18594
作者: Shahar Oded
机构: Ben-Gurion University of the Negev (本-古里安大学)
类目: Computation and Language (cs.CL); Social and Information Networks (cs.SI)
备注: Code is available at: this https URL

点击查看摘要

Abstract:Large-scale discourse analysis often requires stance labels, yet for many high-salience political events such labels are unavailable at the comment level. We introduce and validate a comment-level stance measurement instrument that assigns Pro-Israel, Pro-Palestinian, or Undefined stance using Reddit comment text. The instrument is calibrated from a small, manually annotated gold-standard dataset and scaled through large language model-assisted labeling, achieving weighted- F_1=0.88 against held-out human annotations. To enable corpus-scale measurement, we operationalize the instrument as a lightweight classifier based on DistilBERT embeddings and an RBF SVM, which achieves weighted- F_1=0.7935 on the original gold-standard set. Applying this instrument, we measure the stance of approximately 2.7 million Reddit comments posted between 2023 and 2025 during the Israel-Gaza war. We demonstrate what this measurement enables by analyzing stance-conditioned asymmetries in sentiment, toxicity, and emotional framing across political affiliations. This work shows how validated, low-cost stance measurement can support large-scale descriptive analysis of polarized discourse, without relying on network structure or prior labeled corpora.

[NLP-206] heoretical Foundations of Computational Islamic Jurisprudence (Fiqh)

【速读】: 该论文旨在解决伊斯兰教法(Fiqh)在自动化推理过程中缺乏形式化、可计算且语义严谨的理论基础的问题。传统自然语言处理与一般神学形式化方法无法满足伊斯兰教法作为规范性演绎科学的独特需求,尤其在法律规则的提取、表示与推演中存在语义模糊、逻辑不完整及不可判定性等核心缺陷。其解决方案的关键在于构建一个三层次分层计算语言系统 F=D∘A∘G\mathcal{F} = D \circ A \circ G:首先通过阿拉伯语形态/语法基础引擎(Arabic Morphological/Grammatical Grounding Engine)将古典阿拉伯语转化为语言无关的底层动作结构(Grounded Action Structures, GAS),实现语义锚定;其次利用阿拉伯语实体化动作演算(ARAC)将这些结构标准化为规范的逻辑表示;最后基于无函数的霍恩子句(Horn-clause)Datalog引擎实现可执行的法律推导。尤为关键的是,论文提出一种桥梁算法 Ω\Omega,通过从归纳提取的霍恩子句生成教法分类属性语法(Fiqh-Categorized Attribute Grammar),并借助本体元图(Ontological Meta-Map, Monto\mathcal{M}_{onto})实现从归纳规则到演绎应用的无缝衔接。此外,通过利用词形变化标记(δ\delta 模式)作为结构约束,证明了所提语法的严格无歧义性,达到确定性的 O(n)O(n) 线性解析复杂度,从而在语义正确性、表示完备性、终止性与可判定性方面实现了数学可验证的保障,为自动化伊斯兰教法奠定了形式可证、认识论坚实且计算可行的基础。

链接: https://arxiv.org/abs/2611.18593
作者: Elnaserledinellah Mahmoud Abdelwahab
机构: 未知
类目: Computation and Language (cs.CL); Logic in Computer Science (cs.LO)
备注: 129 pages

点击查看摘要

Abstract:This paper presents the theoretical foundations and a comprehensive formal system for Computational Islamic Jurisprudence (Fiqh). We begin by establishing six foundational principles that distinguish computational Fiqh from conventional natural language processing and general theological formalization. These principles define Islamic jurisprudence as a normative deductive science operating on empirically grounded facts, necessitating a dual inductive-deductive methodology, and requiring a logical foundation specifically engineered for Fiqh rather than adopted from existing formalisms. Based on these principles, we construct a three-layer stratified computational language, \mathcalF = D \circ A \circ G , consisting of: (1) an Arabic Morphological/Grammatical Grounding Engine that transforms fully vocalized Classical Arabic into language-independent Grounded Action Structures (GAS); (2) the Arabic Reified Action Calculus (ARAC) that normalizes these structures into canonical logical representations; and (3) a function-free Horn-clause Datalog engine for executable legal deduction. Furthermore, we introduce a bridging methodology (Algorithm \Omega ) that dynamically generates a Fiqh-Categorized Attribute Grammar from inductively extracted Horn clauses via an Ontological Meta-Map ( \mathcalM_onto ), seamlessly connecting the inductive extraction of legal rules to their deductive application. The paper rigorously proves the system’s semantic correctness, representational completeness, termination, and decidability. Crucially, by leveraging morphosyntactic inflection markers ( \delta modes) as structural constraints, we prove the strict unambiguity of the proposed grammar, achieving deterministic O(n) linear parsing complexity. This work provides a mathematically verifiable, epistemologically grounded, and computationally tractable foundation for automated Islamic jurisprudence.

[NLP-207] Endogenous Preferences as Vulnerability Signals: Discovering Universal Adversarial Triggers for LLM Utility Degradation

【速读】: 该论文旨在解决在缺乏任务标签、有害目标或下游反馈信息的情况下,如何识别冻结的大语言模型(frozen large language models)中潜在的上下文漏洞问题。传统对抗性提示和红队测试依赖于显式的目标定义,而本文关注的是:当任务级别目标未知且下游评估信号不可用时,能否通过模型在干净输入下的行为来发现可复用的上下文脆弱性。其解决方案的关键在于提出一种基于孪生网络架构的端内偏好扰动(Endogenous Preference Disruption, EPD)机制,利用模型在原始指令下的“偏好起始标记”作为条件参考,通过对比干净输入与特定跨度修改后的输入,量化插入操作对模型偏好响应的削弱程度。该方法无需依赖下游任务反馈,而是由辅助生成器提供有限跨度的候选触发片段,由EPD引导其更新并排序候选池,最终选取一个固定的、针对特定模型的最优触发片段,在所有未见输入上保持不变进行评估。实验表明,该固定触发片段在四类指令微调模型及九个基准测试中平均使良性任务性能下降40.2个百分点,且在所有基准上均呈现负向均值变化,优于所有对照基线。消融实验进一步验证了源端选择后固定重用策略具有更强的迁移能力,且触发效果随生成器不同而呈现依赖性,输出层面分析揭示了触发词族与下游失效模式之间的受害者特异性关联。因此,该研究证明了仅通过模型在干净输入中的行为即可暴露可跨输入迁移、但以结构化受害者特异性方式表现的上下文漏洞。

链接: https://arxiv.org/abs/2611.18592
作者: Qingyu Wu,Zeyu Feng,Yuzhe Luo,Hua Cheng
机构: 未知
类目: Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Deployed large language models must remain useful when untrusted text enters their prompt context. Adversarial prompting and red-teaming methods typically define the failures they seek through task labels, harmful targets, judge feedback, or downstream utility scores. We ask whether a frozen model’s clean behavior can reveal a reusable context vulnerability when its task-level form is unknown, and downstream task feedback is withheld. To study this question, EndoUT uses a Siamese-style design that compares the same frozen victim on clean and span-modified versions of each unlabeled instruction. Without a task-level target, the clean pass supplies the victim’s preferred response-onset token as an instruction-conditioned reference; Endogenous Preference Disruption (EPD) measures how insertion weakens that preference. An auxiliary generator supplies bounded-span proposals; EPD guides its update and ranks the pool without access to downstream information. The highest-scoring victim-specific span is fixed before evaluation and reused unchanged across all held-out inputs. Across four instruction-tuned models and nine benchmarks, fixed triggers reduce benign-task utility by 40.2 percentage points on average, with negative mean change on every benchmark and greater average degradation than every evaluated baseline under the reported protocols. Matched protocol ablations show stronger transfer for source-side selection followed by fixed reuse than for null-context fixed reuse or per-input generation. Transfer recurs across proposers with proposal-dependent strength; output-level analyses identify victim-specific associations between trigger families and downstream failure patterns. Thus, clean behavior can expose context vulnerabilities that transfer across inputs yet manifest through structured, victim-specific patterns of utility degradation. The code and data will be released upon acceptance.

[NLP-208] Where Language Models Fail Like Aphasic Patients and Where They Dont

【速读】: 该论文旨在解决人类词汇生成过程中语义-词法检索与音位编码是否构成可分离的加工阶段这一长期争议性问题。传统研究依赖于小词汇量的手工设计模型,难以充分验证理论假设。本文提出“计算架构分离”(computational architectural dissociation)方法,利用视觉-语言模型(VLMs)已知的架构特性——即其通过全词汇量上的学习语义竞争完成图像命名,但完全缺乏音位编码机制——来推导出可检验的预测:在扰动条件下,患者错误类型应与模型错误在语义阶段类别中对齐,但在音位阶段类别中不应对齐。实验基于LLaVA 1.6(Vicuna-13B)模型,在88名失语症患者对175个费城命名测试项目进行约3700万次命名尝试的基础上,进行了50种扰动种子的分析,结果显示八个错误类别中有六个在项目层面表现出显著一致性(斯皮尔曼等级相关系数ρ = 0.32–0.50,全部p < 0.001),而两个音位类错误(形式错误与抽象错构词)则无显著对齐(ρ = 0.110 和 0.084,均不显著),从而证实了预期的分离效应。通过对输出概率分布进行谱分析,每个错误类型在六维空间中被定位至独特位置,提供了可解释的机制工具。该分离效应具有层特异性(峰值位于第22层,占总深度55%)、与病情严重程度无关,并在不同扰动种子间高度稳定(分半相关r = 0.998)。进一步对四个其他7B规模的VLMs进行权重扰动测试,亦获得了相似量级的正确率对齐(ρ = 0.21–0.27),表明该现象并非特定模型或规模所致。最终结论支持语义-词法与音位编码在人类词汇生成中为可分离的加工阶段,且其分离特征可通过大规模预训练模型的架构属性进行有效建模与验证。

链接: https://arxiv.org/abs/2611.18591
作者: Yong Yang,Roger Newman-Norlund,Xiang Guan,Saeed Ahmadi,Regan Willis,Nadra Salman,Kalil Warren,Sophie Arheix-Parras,Srihari Nelakuditi,Rutvik H. Desai,Leonardo Bonilha,Christopher Rorden,Julius Fridriksson
机构: 未知
类目: Computation and Language (cs.CL)
备注: 15 pages, 9 figures, 1 table. Supplementary information included (15 figures, 7 tables). A companion analysis comparing model scale and attention design within a single inference implementation is in preparation and will be reported in a subsequent version

点击查看摘要

Abstract:The computational architecture of human word production – whether semantic-lexical retrieval and phonological encoding constitute dissociable processing stages – has been debated for decades using hand-designed models with small vocabularies. We introduce computational architectural dissociation: using known architectural properties of vision-language models (VLMs) to derive falsifiable predictions about which patient error types should and should not align with model errors under perturbation. Because VLMs perform picture naming through learned semantic competition over full vocabularies but lack phonological encoding entirely, perturbation-induced errors should match patient patterns for semantic-stage categories but not phonological ones. Testing LLaVA 1.6 (Vicuna-13B) against 88 patients with aphasia across 175 Philadelphia Naming Test items (~37 million naming attempts, 50 perturbation seeds), we find that six of eight error categories show significant item-level alignment (Spearman \rho = 0.32-0.50, all p 0.001) while the two phonological categories – formal errors and abstruse neologisms – show no alignment ( \rho = 0.110 and 0.084, both n.s.), confirming the predicted dissociation. Spectral analysis of the output probabilities places each error type at a distinct position in a six-dimensional space, giving a mechanistic interpretability tool. The dissociation is layer-specific (peaking at L22, 55% depth), severity-independent, and stable across seeds (split-half r 0.998). Weight perturbation of four additional 7B VLMs spanning the tested architectural combinations produced correct-rate alignment of comparable magnitude ( \rho = 0.21-0.27). Whether the per-category dissociation itself depends on scale or on attention design requires a comparison within a single inference implementation, which is reported separately.

[NLP-209] OmniTag-LLM : Multilingual Multitask Speech Tagging

【速读】: 该论文旨在解决多模态大模型(Multimodal LLMs)在跨语言语音理解任务中缺乏统一评估与联合建模能力的问题,尤其针对现有基准测试仍以孤立方式评估各子任务且主要聚焦英语的局限性。其核心挑战在于如何实现从单一音频片段中端到端地联合完成语音转录、情感识别、说话人性别判断、语言识别(Language Identification, LID)及词级别命名实体识别(Named Entity Recognition, NER),并确保在多语言场景下的准确性和鲁棒性。解决方案的关键在于提出OmniTag-Bench——一个涵盖五种语言(英语、德语、西班牙语、法语、意大利语)的统一基准测试框架与训练方案,支持“单个音频输入,输出包含所有信息的JSON对象”的联合理解范式,并配套设计了标准化的提示模板与评分器。实验表明,通过LoRA微调的Qwen3-Omni模型在命名实体识别、情感分析和词错误率(WER)上显著提升,尤其在独立标注的情感数据集上性能近乎翻倍,验证了跨任务知识迁移的有效性;此外,该模型还展现出对未见相关语言(如荷兰语)的良好泛化能力,凸显了其在多语言联合建模中的潜力。研究同时开源了全部提示、评分工具与标注流程,为后续研究提供可复现的基础。

链接: https://arxiv.org/abs/2611.18589
作者: Aref Farhadipour,Francesco Verdini,Antonis Asonitis,Pierre-Edouard Honnet,Marzieh Razavi,Vijeta Avijeet,Juan Pablo Zuluaga Gomez
机构: AGIGO; University of Zurich(苏黎世大学); University of Rome(罗马大学); ETH Zurich(苏黎世联邦理工学院)
类目: Computation and Language (cs.CL); Sound (cs.SD); Audio and Speech Processing (eess.AS)
备注: paper in 8 pages and appendix

点击查看摘要

Abstract:Multimodal LLMs promise to replace the traditional cascade of specialist speech models with a single system that, from one audio clip, jointly transcribes and labels emotion, gender, language, and named entities. Yet most benchmarks still score these tasks in isolation, and largely in English. We introduce OmniTag-Bench, a five-language benchmark (en/de/es/fr/it) and training recipe for joint understanding: one audio clip in, one JSON object out (transcript, emotion, gender, LID, token-level NER). We evaluate ten open multimodal LLMs and a LoRA fine-tuned Qwen3-Omni on a unified prompt and scorer. Beyond aggregate scores, OmniTag-Bench reveals a systematic failure mode: several strong models often transcribe or translate non-English speech into English, inflating multilingual WER. Fine-tuning improves the base model on NER, emotion, and WER while nearly doubling accuracy on independent human-labeled emotion corpora, suggesting transferable gains beyond pseudo-label imitation. The fine-tuned model also generalizes to unseen but related languages, transferring to held-out Dutch. We release prompts, the scorer, and the labeling pipeline.

[NLP-210] Quantize Qualify Rerank: A Recipe for Compressing LLM s Without Losing Quality

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在推理服务中因模型规模庞大而导致的高成本问题,特别是在模型压缩过程中如何有效评估和维持翻译质量的关键挑战。其核心问题是:单一自动评估指标可能无法准确捕捉模型压缩带来的细微质量损失,导致评估结果失真。为应对这一问题,研究提出以人类元评估(human meta-evaluation)为指导,采用多指标协同验证策略——将MetricX24-XXL作为主要评估指标,CometKiwi-XXL作为交叉验证手段,从而缓解单个指标的偏差。解决方案的关键在于通过结构化剪枝(移除未使用的视觉模块与非任务词汇)、混合精度计算(int8权重或fp8激活)实现高效压缩,同时避免显著性能下降;对于更激进的int4量化,虽存在微小残余损失,但通过Cometoid基于T=0.3温度参数的“最佳4选一”重排序器进行后处理优化,显著提升了8位与4位系统在WMT25测试集上的表现。最终,在无参考(reference-free)评估下,压缩后的模型仍能保持与原始bf16版本相当的竞争力,且不同重排序方案间的微小差异被合理视为噪声,验证了整体方法的有效性与鲁棒性。

链接: https://arxiv.org/abs/2611.18587
作者: Thamme Gowda
机构: Microsoft(微软)
类目: Machine Learning (cs.LG); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Compressing large language models lowers serving cost, but a single automatic metric may miss quality loss. For the WMT26 Model Compression constrained track, we compress google/gemma-3-12b-it (24.4 GB in bf16) for three translation directions on one H100. Paired bootstrap analysis exposes metric disagreement on 4-bit quantization; guided by WMT24 human meta-evaluation, we use MetricX24-XXL as primary and CometKiwi-XXL as a cross-check. Removing the unused vision stack and non-task vocabulary and using int8 weights or fp8 activations causes no resolved loss, whereas int4 leaves a small residual. QLoRA healing overfits its calibration domain. A Cometoid best-of-4 reranker at T=0.3 improves both the 8-bit and 4-bit systems under WMT25 test set. On post-submission WMT26, both remain competitive with bf16 under reference-free metrics; the small difference between the reranked variants is treated as noise.

[NLP-211] Human vs Machine Translation Detection: A Cross-Model Cross-Domain and Low-Resource Analysis

【速读】: 该论文旨在解决大规模语言模型(Large Language Models, LLMs)训练中因依赖网络爬取语料而引发的数据质量问题,特别是机器翻译(Machine Translation, MT)内容被误当作人类翻译(Human Translation, HT)使用所带来的风险,尤其在低资源语言场景下更为显著。其核心挑战在于如何有效区分MT与HT文本,以避免合成伪影的传播并保障数据质量。解决方案的关键在于系统评估不同模型架构在跨语言、跨领域和跨翻译源条件下的检测性能:研究发现,仅采用解码器的大型语言模型在零样本或少样本设置下表现有限,且存在明显的类别不对称性——对人类写作内容识别准确率高(高HT召回率),但难以检出流畅的机器翻译输出(低MT召回率)。相比之下,基于编码器的微调模型在特定训练域(如立法文本)内可实现高精度检测,但在面对未见过的译者或新领域(如低资源对话文本、医学内容)时,其对机器翻译的召回率急剧下降。为提升泛化能力,研究提出将长文本按最多300个词元(token)分块处理,显著改善了模型在多场景下的检测效果,表明句子级特征不足以支撑鲁棒的检测能力。此外,通过COMET与BLEURT等指标分析翻译质量与可检测性的关系,发现低质量机器翻译更容易被识别,而高质量机器翻译仍难以与人类翻译区分。综上,该研究强调输入粒度、译者泛化能力以及翻译质量在构建跨语言、跨领域稳健的机器翻译检测系统中的关键作用。

链接: https://arxiv.org/abs/2611.18586
作者: Yabsera Yemanberhan,Prasenjit Mitra,Roald Eiselen
机构: Centre for Text Technology, North-West University (北西大学文本技术中心); Carnegie Mellon University Africa (卡内基梅隆大学非洲校区)
类目: Computation and Language (cs.CL)
备注: 20 pages, 6 figures, 12 tables. Accepted at the Conference on Language Modeling (COLM) 2026. Code: this https URL

点击查看摘要

Abstract:The growing reliance on web-crawled corpora for training large language models introduces a critical challenge: machine-translated (MT) content is often treated as human translation (HT), potentially degrading data quality and propagating synthetic artifacts, especially in low-resource settings. In this work, we study MT-HT detectability across high-resource (English, Spanish) and low-resource (Swahili, Afrikaans) languages using both decoder-only LLMs in zero-shot and few-shot settings, and fine-tuned multilingual encoders trained at the sentence level and on token-chunked inputs. We find that decoder-only models provide limited discrimination overall and exhibit a consistent class-asymmetric failure mode: they recognize human-written text reliably (high HT recall) but rarely flag fluent machine output (low MT recall). Encoder-based classifiers achieve high accuracy when evaluated on data generated by the same MT systems and translators used during training in a legislative domain, but their MT recall drops sharply on unseen translators and on unseen domains such as low-resource conversational text and medical content. To address this limitation, we evaluate models on longer inputs by chunking text into segments of up to 300 tokens, which substantially improves generalization, suggesting that sentence-level signals are insufficient for robust detection. Finally, we analyze the relationship between translation quality and detectability using COMET and BLEURT, showing that lower-quality MT is more easily identified, while high-quality MT remains difficult to distinguish from HT. Our results highlight the importance of input granularity, translator generalization, and translation quality in building robust MT detection systems across domains and languages.

[NLP-212] Read Route Patch: Robust Test-Time Evolution for LLM Agents

【速读】: 该论文旨在解决基于大语言模型(Large Language Model, LLM)的智能体在长时程交互决策中因信息瓶颈导致的性能受限问题。具体而言,现有方法存在三大缺陷:记忆检索过度依赖文本相似性、候选行为策略常出现行为冗余、全量策略重写可能覆盖有效证据或引入组件间不一致性。为应对上述挑战,论文提出Trident框架,其核心解决方案包括:基于边界感知的记忆读取机制以提升经验利用效率;基于多样性感知的候选路由策略以增强行为探索多样性;基于顺序感知的插入式更新方法以实现策略演化的稳定性与一致性。实验结果表明,在涵盖文本游戏、网页交互和代码生成任务的五个基准测试上,Trident在qwen-32B、mistral-small-3.2-24B和gpt-oss-20B三种模型架构下均取得最优平均表现,相较于最强基线APEX分别实现30.7%、30.8%和38.3%的相对性能提升。消融实验与敏感性分析进一步验证了三项设计具有互补性且对参数变化具有鲁棒性。

链接: https://arxiv.org/abs/2611.18585
作者: Kaijie Chen,Liang Yan,Yilan Wei,Jiachen Luo,Zhihao Dou,Jiashu Qu,Haoran Liu,Junli Wang
机构: 未知
类目: Computation and Language (cs.CL)
备注: 13 pages

点击查看摘要

Abstract:Large Language Model (LLM)-based agents can improve long- horizon interactive decision-making through test-time evolu- tion, yet existing methods face three information bottlenecks: memory retrieval overemphasizes textual similarity, candi- date harnesses are often behaviorally redundant, and whole- harness rewriting may overwrite useful evidence or introduce cross-component inconsistency. We introduce Trident, which improves experience utilization, behavioral exploration, and harness evolution through boundary-aware memory reading, diversity-aware candidate routing, and order-aware insertion- based updating, respectively. We evaluate Trident on five benchmarks spanning text games, web interaction, and cod- ing tasks. Across qwen-32B, mistral-small-3.2-24B, and gpt- oss-20B, Trident achieves the best average performance and consistently outperforms non-learning, memory-based, reflection-based, and automated prompt optimization base- lines. Compared with the strongest baseline, APEX, Trident yields relative average gains of 30.7%, 30.8%, and 38.3% on the three backbones, respectively. Ablation and sensitivity stud- ies further show that the three designs provide complementary gains and remain robust to parameter variations.

[NLP-213] “I Am Not a Kiddo!”: Quantifying Epistemic Infantilization in Large Language Models via Multi-Metric Bias Analysis

【速读】: 该论文旨在解决生成式AI在专业知识服务场景中因用户身份披露(特别是残疾身份)而引发的“认知幼稚化”(epistemic infantilization)问题,即当用户在提出专家级问题前披露自身残疾状况时,模型倾向于降低技术深度、简化内容,从而导致其知识获取能力被系统性削弱。其解决方案的关键在于构建一个反事实数据集,通过对比三种情境下模型响应:无身份背景、对话上下文中披露残疾、以及通过短查询前缀披露残疾,量化该效应。研究引入两个复合指标——认知幼稚化偏差指数(Infantilization Bias Index, IBI)与认知可及性响应评分(Epistemic Accessibility Response Score, EARS),综合评估模型在数学与形式化内容保留、领域术语使用、推理覆盖度及语气信号(如简化标记、主动降级提议)等方面的表现。结果表明,五种开源大模型均呈现显著的去专业化趋势,尤其在物理与数学领域,且该效应在考虑重复测量结构后依然稳健。研究揭示了一种“包容性税”现象:残障用户需更明确地请求严谨内容才能获得与非残障用户同等的知识深度。最终主张,真正的可访问性应体现为格式适配而非内容简化,以保障所有用户在认知层面的平等权利。

链接: https://arxiv.org/abs/2611.18583
作者: Aditya Upadhyay,Arpan Verma,Chandan Sah,Dev Utkarsh,Harshit Gupta
机构: IIIT Delhi (印度信息科学技术研究所)
类目: Computation and Language (cs.CL); Computers and Society (cs.CY)
备注:

点击查看摘要

Abstract:Large language models (LLMs) are increasingly used for tutoring and professional knowledge retrieval, yet safety evaluations still focus mainly on overt harms such as toxicity while overlooking benevolent failure modes. We study epistemic infantilization, defined as an identity-conditioned reduction in technical depth that appears when disability disclosure precedes an expert-level query. To measure this effect, we construct a counterfactual dataset of matched question-answer pairs and compare model responses across three settings: no identity context, disability disclosed in prior conversational context, and disability disclosed via a short query prefix. We introduce two composite metrics, the Infantilization Bias Index (IBI) and the Epistemic Accessibility Response Score (EARS), which capture retention of mathematical and formal content, domain terminology, reasoning coverage, and tone signals such as simplification markers and unsolicited offers to adjust the level. Across five open-weight model families, we find consistent evidence that disability disclosure reduces epistemic access by replacing formal explanations with higher-level summaries, with Physics and Mathematics most affected. Using question-clustered inference that respects the repeated-measures structure of the data, the effect remains reliable after accounting for correlated observations. These findings suggest an inclusion tax, where disabled users must request rigor more explicitly than non-disabled users. We conclude by arguing that accessibility should mean format adaptation rather than content simplification.

[NLP-214] VidEngram: Agent ic Long-Term Video Understanding via Structured Hierarchical Memory

【速读】: 该论文旨在解决长视频问答(long-video question answering)中因多轮查询需重复编码视频而导致的计算冗余与信息丢失问题。现有方法受限于上下文长度,不得不对视频进行稀疏采样或截断,且无法保留可复用的表示以支持后续查询。其核心解决方案是提出VidEngram——一种基于代理(agentic)架构的视频问答系统,关键在于在视频摄入阶段一次性构建持久化的分层记忆结构:通过Qwen2.5-Omni联合生成视频与音频的描述,将片段级记忆整合为剧集层与实体层,并利用仅依赖文本的代理规划器结合专用检索工具生成有依据的答案。该设计实现了“感知一次、问答多次”的高效模式,显著降低后续查询的计算开销。实验表明,在完整的OmniVideoBench基准上,采用GPT-5-mini规划器的VidEngram优于单次遍历的Qwen2.5-Omni基线(37.0% vs. 32.3%),尤其在长视频场景下优势更为明显,验证了其在处理长序列内容时的优越性。

链接: https://arxiv.org/abs/2611.18578
作者: Zinuo Cheng,Yueqian Lin,Zhongxi Wang,Hai “Helen” Li,Yiran Chen
机构: Duke University (杜克大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Computation and Language (cs.CL); Multimedia (cs.MM)
备注: 10 pages, 4 figures, 4 tables. Demo video and code are linked in the paper

点击查看摘要

Abstract:Long-video question answering is wasteful and lossy when a multimodal model must re-encode the same video for every query: context limits force sparse sampling or truncation, and no reusable representation remains for later questions. We present VidEngram, an agentic video QA system that builds a persistent hierarchical memory once at ingest and answers later questions from that memory. VidEngram captions video and audio jointly with Qwen2.5-Omni, consolidates segment memories into episode and entity layers, and uses a text-only agentic planner with specialized retrieval tools to produce grounded answers. The web interface makes this process inspectable: users can watch memory formation, inspect retrieval traces, verify cited evidence in the video player, and explore entity relationships. On the full OmniVideoBench, VidEngram with a GPT-5-mini planner outperforms a strong single-pass Qwen2.5-Omni baseline (37.0% vs. 32.3%), with gains concentrated on longer videos where single-pass coverage becomes sparse. Because perception is paid once, subsequent queries use low-cost text-only planning.

[NLP-215] ChartSync: A Benchmark for Visuo-Logical Cascading Chart Editing

【速读】: 该论文旨在解决生成式图像编辑模型在处理结构化统计图表时,因数据修改需保持几何同步而面临的挑战,尤其聚焦于跨模态的视觉-逻辑级联编辑(Visuo-Logical Cascading Editing, VLCE)问题。现有方法普遍局限于局部文本替换,难以实现依赖感知的级联更新。为系统评估该能力,研究提出ChartSync——一个基于程序化渲染管道构建的专家验证基准,确保真实标签中视觉与逻辑的一致性(visuo-logical coupling)。ChartSync包含870个三元组,覆盖9类图表和4种任务类型,其中235个实例专门测试文本到几何的级联同步能力。研究采用双层评估框架,结合客观视觉指标与视觉-语言模型判别机制,综合评估低层级保真度及多模态理解与推理能力。对14个图像编辑模型及1个代码中介流程的评估揭示了显著的能力差距:多数开源模型在几何同步方面表现严重退化,仅有两个前沿专有模型展现出初步的VLCE能力,其残余错误主要源于语义孤立与背景污染。通过深入错误分析,研究识别出未来多模态架构应具备的核心元能力。相关数据集与代码已公开发布。

链接: https://arxiv.org/abs/2607.10301
作者: Jiakang Yu,Yixuan Chai,Tianci Wang,Rihui Jin,Guangkai Xu,Hongtao Deng,Xun Zhu,Wang Gao,Xinrun Guo,Haipang Wu
机构: Jianghan University(江汉大学); HiThink Research; University of Science and Technology of China(中国科学技术大学); Southeast University(东南大学); Zhejiang University(浙江大学)
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Generative image editing models struggle with structured statistical charts when data modifications require geometric synchronization. We formalize this task as Visuo-Logical Cascading Editing (VLCE). However, existing methods remain confined to localized text substitutions and struggle with dependency-aware cascading updates. To systematically evaluate this capability, we introduce ChartSync, an expert-validated benchmark constructed via a programmatic rendering pipeline that guarantees deterministic visuo-logical coupling for the ground truth. ChartSync comprises 870 triplets across 9 chart categories and 4 task types, including 235 geometry-coupled VLCE instances that specifically test cascading text-to-geometry synchronization. We further evaluate these instances via a two-tier framework combining objective visual metrics with a vision-language model judge paradigm to assess low-level fidelity alongside multimodal comprehension and reasoning. Evaluating 14 image editing models and one code-mediated pipeline reveals a nuanced capability gap: most open-source models suffer severe drops in geometric synchronization, while only two frontier proprietary models show emerging VLCE capability, with their residual errors mainly involving semantic isolation and background corruption. Our detailed error analysis deconstructs these failure paradigms to identify core meta-abilities for guiding future multimodal architectures. The ChartSync dataset and code are publicly released at this https URL.

[NLP-216] Fast Catch-Up Late Switching: Optimal Batch Size Scheduling via Functional Scaling Laws ICLR2026

【速读】: 该论文旨在解决大规模深度学习训练中批量大小调度(Batch Size Scheduling, BSS)的理论基础薄弱问题,尤其关注在固定数据预算下如何实现最优的批量大小动态调整。其核心挑战在于理解不同任务难度对最优调度策略的影响,并揭示其背后的优化机制。解决方案的关键在于引入功能缩放律(Functional Scaling Law, FSL)框架,该框架为分析BSS提供了理论依据。研究发现,任务难度显著影响最优调度结构:对于简单任务,最优策略是全程逐步增大批量大小;而对于复杂任务,最优策略则是在训练早期保持小批量,仅在后期阶段切换至大批量。这一“晚期切换”现象由一种称为“快速追上效应”(fast catch-up effect)的动力学机制所解释——当从大批量切换后,损失函数迅速趋近于恒定大批量的轨迹,其收敛速度由任务难度决定。该效应源于梯度噪声的快速遗忘,表明大批量可安全延迟至训练后期,从而大幅降低数据消耗而不牺牲性能。通过涵盖密集模型与混合专家(MoE)架构、参数量达11亿、训练数据高达1万亿标记的大量语言模型预训练实验,验证了理论预测:所有场景下,晚切换调度均显著优于恒定批量和早切换基线。

链接: https://arxiv.org/abs/2602.14208
作者: Jinbo Wang,Binghui Li,Zhanpeng Zhou,Mingze Wang,Yuxuan Sun,Jiaqi Zhang,Xunliang Cai,Lei Wu
机构: Peking University (北京大学); Shanghai Jiao Tong University (上海交通大学); University of Science and Technology of China (中国科学技术大学); Meituan (美团), Beijing AI for Science Institute (北京人工智能科学研究院)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Optimization and Control (math.OC); Machine Learning (stat.ML)
备注: 34 pages, accepted by ICLR 2026 as a conference paper

点击查看摘要

Abstract:Batch size scheduling (BSS) plays a critical role in large-scale deep learning training, influencing both optimization dynamics and computational efficiency. Yet, its theoretical foundations remain poorly understood. In this work, we show that the functional scaling law (FSL) framework introduced in Li et al. (2025a) provides a principled lens for analyzing BSS. Specifically, we characterize the optimal BSS under a fixed data budget and show that its structure depends sharply on task difficulty. For easy tasks, optimal schedules keep increasing batch size throughout. In contrast, for hard tasks, the optimal schedule maintains small batch sizes for most of training and switches to large batches only in a late stage. To explain the emergence of late switching, we uncover a dynamical mechanism – the fast catch-up effect – which also manifests in large language model (LLM) pretraining. After switching from small to large batches, the loss rapidly aligns with the constant large-batch trajectory. Using FSL, we show that this effect stems from rapid forgetting of accumulated gradient noise, with the catch-up speed determined by task difficulty. Crucially, this effect implies that large batches can be safely deferred to late training without sacrificing performance, while substantially reducing data consumption. Finally, extensive LLM pretraining experiments – covering both Dense and MoE architectures with up to 1.1B parameters and 1T tokens – validate our theoretical predictions. Across all settings, late-switch schedules consistently outperform constant-batch and early-switch baselines.

[NLP-217] he Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-Training ICML2025

【速读】: 该论文旨在解决大语言模型(Large Language Model, LLM)预训练过程中不同网络模块间学习动态不均衡的问题,特别是由于各构建模块(如嵌入层、归一化层、自注意力机制和前馈网络等)在优化过程中的曲率差异(Sharpness Disparity)导致的训练效率瓶颈。其核心解决方案是提出**分块学习率(Blockwise Learning Rate, Blockwise LR)**策略,该策略根据每个模块的固有曲率特性动态调整学习率,从而实现更高效的参数更新。关键在于识别并利用早期训练阶段即显现且贯穿始终的模块间曲率差异,通过为不同模块分配适配其优化特性的学习率,显著降低最终损失并实现接近2倍的训练加速。该方法可无缝集成至AdamW及内存高效的Adam-mini优化器中,同时在多个主流模型(GPT-2、LLaMA)与数据集上验证了其有效性,展现出在提升训练效率与降低显存消耗方面的双重优势。

链接: https://arxiv.org/abs/2502.19002
作者: Jinbo Wang,Mingze Wang,Zhanpeng Zhou,Junchi Yan,Weinan E,Lei Wu
机构: Peking University (北京大学); Shanghai Jiao Tong University (上海交通大学); Center for Machine Learning Research, Peking University (北京大学机器学习研究中心); AI for Science Institute, Beijing, China (北京科学智能研究院)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Optimization and Control (math.OC); Machine Learning (stat.ML)
备注: 21 pages, accepted by ICML 2025

点击查看摘要

Abstract:Transformers consist of diverse building blocks, such as embedding layers, normalization layers, self-attention mechanisms, and point-wise feedforward networks. Thus, understanding the differences and interactions among these blocks is important. In this paper, we uncover a clear Sharpness Disparity across these blocks, which emerges early in training and intriguingly persists throughout the training process. Motivated by this finding, we propose Blockwise Learning Rate (LR), a strategy that tailors the LR to each block’s sharpness, accelerating large language model (LLM) pre-training. By integrating Blockwise LR into AdamW, we consistently achieve lower terminal loss and nearly 2\times speedup compared to vanilla AdamW. We demonstrate this acceleration across GPT-2 and LLaMA, with model sizes ranging from 0.12B to 2B and datasets of OpenWebText, MiniPile, and C4. Finally, we incorporate Blockwise LR into Adam-mini (Zhang et al., 2024), a recently proposed memory-efficient variant of Adam, achieving a combined 2\times speedup and 2\times memory saving. These results underscore the potential of exploiting the sharpness disparity to improve LLM training.

[NLP-218] Using the Tsetlin Machine to Learn Human-Interpretable Rules for High-Accuracy Text Categorization with Medical Applications

【速读】: 该论文旨在解决医疗文本分类中高准确率与可解释性之间的矛盾问题。传统深度学习方法虽在准确性上取得显著提升,但牺牲了模型的可解释性,难以满足医疗领域对决策透明性的要求。为此,论文首次引入近期提出的有限状态机(Tsetlin Machine),提出一种基于命题逻辑的文本分类新方法:将文本中的词汇表示为命题变量,并利用简单的合取公式(如“若出现‘皮疹’、‘反应’和‘青霉素’,则判定为过敏”)来建模类别特征。该方法不仅能够利用正向词项进行分类,还可通过否定特征(即未出现的词项)增强分类能力。实验结果表明,在20 Newsgroups、IMDb及一个非公开临床数据集上的对比测试中,Tsetlin Machine在准确率、召回率和精确率方面均优于或持平于朴素贝叶斯、决策树、线性支持向量机(SVM)、随机森林、长短期记忆网络(LSTM)等主流方法。此外,其基于GPU的实现相较CPU版本提速5至15倍。因此,该研究的关键在于通过可解释的符号逻辑机制实现了高性能与可解释性的统一,为医疗文本分析等需要透明决策过程的应用提供了强有力的新范式。

链接: https://arxiv.org/abs/1809.04547
作者: Geir Thore Berge,Ole-Christoffer Granmo,Tor Oddbjørn Tveit,Morten Goodwin,Lei Jiao,Bernt Viggo Matheussen
机构: 未知
类目: Machine Learning (cs.LG); Computation and Language (cs.CL); Machine Learning (stat.ML)
备注: 10 pages, 4 figures

点击查看摘要

Abstract:Medical applications challenge today’s text categorization techniques by demanding both high accuracy and ease-of-interpretation. Although deep learning has provided a leap ahead in accuracy, this leap comes at the sacrifice of interpretability. To address this accuracy-interpretability challenge, we here introduce, for the first time, a text categorization approach that leverages the recently introduced Tsetlin Machine. In all brevity, we represent the terms of a text as propositional variables. From these, we capture categories using simple propositional formulae, such as: if “rash” and “reaction” and “penicillin” then Allergy. The Tsetlin Machine learns these formulae from a labelled text, utilizing conjunctive clauses to represent the particular facets of each category. Indeed, even the absence of terms (negated features) can be used for categorization purposes. Our empirical comparison with Naïve Bayes, decision trees, linear support vector machines (SVMs), random forest, long short-term memory (LSTM) neural networks, and other techniques, is quite conclusive. The Tsetlin Machine either performs on par with or outperforms all of the evaluated methods on both the 20 Newsgroups and IMDb datasets, as well as on a non-public clinical dataset. On average, the Tsetlin Machine delivers the best recall and precision scores across the datasets. Finally, our GPU implementation of the Tsetlin Machine executes 5 to 15 times faster than the CPU implementation, depending on the dataset. We thus believe that our novel approach can have a significant impact on a wide range of text analysis applications, forming a promising starting point for deeper natural language understanding with the Tsetlin Machine.

[NLP-219] MolReps: Taking SELFIES Making SMILES Teaching LLM s

【速读】: 该论文旨在解决大语言模型(Large Language Models, LLMs)在面对同一分子的不同语法上异构表示(如SMILES、InChI、Molfile等)时,难以识别和关联这些等价编码的问题。其核心挑战在于现有模型对分子表示的语义一致性理解不足,导致在跨表示任务中表现不佳。解决方案的关键是提出MolReps——一个系统化生成9,600组问答对的流水线,覆盖十种分子表示形式,涵盖转换、归一化及二维结构识别任务。通过在该数据集上进行针对性后训练(post-training),显著提升了文本仅模型(如MolQwen3-4B-Instruct)和多模态模型(如MolQwen3-VL-4B-Instruct)在分子重叠与非重叠测试集上的性能,尤其在分子属性预测基准MolRepBench上实现了27.84%的相对提升,验证了针对分子表示理解进行专门微调的有效性。

链接: https://arxiv.org/abs/2611.20848
作者: Ken Deng,Garrett Morris
机构: 未知
类目: Chemical Physics (physics.chem-ph); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:The same molecule can be expressed in multiple, syntactically distinct representations, yet it remains unclear whether large language models (LLMs) can recognize and relate these equivalent encodings. We introduce MolReps, a pipeline that generates 9,600 question-answer pairs across ten representations for conversion, normalization, and 2D structure recognition. Post-training on MolReps improves text-only and multimodal models, MolQwen3-4B-Instruct (MQ3-4B-I) and MolQwen3-VL-4B-Instruct (MQ3-VL-4B-I), from 4.01% to 75.13% and from 7.11% to 73.14% (InChIKey match) on the molecule-overlap test set, respectively. On the molecule-disjoint test set, post-training increased validity from 20.35% to 58.12% for MQ3-4B-I and from 17.19% to 59.34% for MQ3-VL-4B-I. However, InChIKey match remained low, decreasing from 3.86% to 3.09% and from 6.44% to 2.29% , respectively. Furthermore, the resulting models achieve a relative 27.84% improvement in average over their base model on MolRepBench, a comprehensive benchmark for evaluating molecular property prediction. These results show the effectiveness of targeted post-training for enhancing molecular representation understanding and its potential to improve downstream molecular tasks.

信息检索

[IR-0] ExRel: Learning Directed Reuse Relevance between Rich-Text K-12 Question Images with Executable Relation Contracts

链接: https://arxiv.org/abs/2611.21461
作者: Xuexin Liu,Zengyi Wo,Shiyu Zhang,Suqi Cheng,Daiting Shi,Xiuyi Chen
类目: Information Retrieval (cs.IR)
备注: 12 pages, 6 figures, 6 tables

点击查看摘要

Abstract:In K-12 education, question images combine text, mathematical expressions, diagrams, and diverse layouts. Variations in wording, conditions, answer options, or task scope can substantially alter their reuse utility despite visual similarity. We formulate this problem as directed reuse relevance, grading candidate utility as Strong, Medium, or Weak. However, generic similarity scores overlook task-specific reuse requirements, while natural supervision is dominated by Strong relations and lacks fine-grained coverage. We propose ExRel, an agentic, code-grounded framework for controllable and verifiable visual data synthesis. It reconstructs real questions into editable HTML/CSS seeds and compiles 15 relation subtypes into executable contracts specifying permitted edits, protected invariants, and query-candidate direction. Synthesized pairs undergo deterministic DOM/render validation and target-hidden, evidence-assisted review, with bounded repair before mild photographic degradation. Only rendered images are used for downstream training. Experiments show that ExRel supervision outperforms size-matched natural data and count- and class-matched replay, particularly on Medium relations, while maintaining comparable query-level ranking performance and improving robustness to visual degradation. Construction ablations demonstrate higher valid yield and fewer false admissions than less structured baselines. Human audits confirm 98.0% post-admission validity, while ExRel achieves a lower per-accepted-pair cost than outsourced annotation (0.70 vs. 1.10 RMB). Downstream answer routing further demonstrates the practical utility of the synthesized supervision.

[IR-1] EE-Motion: An EEG-based Web Architecture for Live Affective Music Recommendation

链接: https://arxiv.org/abs/2611.21032
作者: Sabino Roccotelli,Potito Aghilar,Vito Walter Anelli,Alejandro Bellogín,Tommaso Colafiglio,Michelantonio Trizio,Tommaso Di Noia
类目: Information Retrieval (cs.IR); Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Music listening is intrinsically tied to emotion, which drifts continuously within a single session. Although modern browsers interface with wearable brain-computer interfaces, recommendation systems remain structured around discrete logs and lack the infrastructure to integrate continuous physiological signals. Furthermore, existing affect-aware models rely on static user profiles and offline training, failing to serve anonymous newcomers without model retraining. We present a web platform that resolves both challenges. Pairing with a consumer electroencephalogram (EEG) headband over Web Bluetooth, the system captures live EEG, mapping frontal alpha asymmetry and arousal index to normalized valence-arousal proxy coordinates on Russell’s circumplex plane. Driven by these coordinates, an affective adapter unifies structurally diverse recommendation backbones under a shared live scoring mechanism. By parameterizing anonymous visitors through a few favorite tracks, the module recommends tracks conditioned on live proxy coordinates. This demonstrates that real-time EEG-derived control signals can steer recommendations without persistent user tables or model retraining.

[IR-2] Adaptive-Elbow Graph Traversal for Schema Retrieval in Text-to-SQL

链接: https://arxiv.org/abs/2611.20979
作者: Fabio Picariello,Salvatore Cipolla,Susanna Bonura,Vito Morreale
类目: Databases (cs.DB); Information Retrieval (cs.IR)
备注: Presented at the 5th Italian Conference on Big Data and Data Science (ITADATA 2026). To appear in CEUR Workshop Proceedings

点击查看摘要

Abstract:LLM-based natural-language interfaces to relational databases are gated by a schema-retrieval step: the generator can only be as good as the slice of schema it is shown. On a database with hundreds of tables, that slice must be selected and the selection is two-sided—omit a needed table and the query becomes unrecoverable, include too many and the irrelevant ones act as noise that degrades the model’s output. When an agentic retriever that lets the model inspect the schema can be too costly or too slow, the fallback is an embedder that ranks tables by cosine similarity and keeps a top-k. This fallback has two weaknesses: a fixed budget is suboptimal question by question, and embedding similarity overlooks foreign-key-linked tables that are necessary for joins. We present Adaptive-Elbow Graph Traversal (AEGT, ElbowHop for short), a property-graph table retriever that addresses both: an adaptive elbow cutoff replaces the fixed budget with a per-question one under a single aggressiveness parameter, and bidirectional foreign-key traversal deterministically recovers the bridge tables embedding similarity misses. We evaluate the retriever on 3 corpora of increasing candidate-pool size (BIRD Mini-Dev queried per database, 3–13 tables; its eleven schemas merged, 74; and a merged Spider-train graph, 546) and isolate the two mechanisms under a per-question budget-matched protocol. Across these settings the adaptive elbow emerges as the dominant mechanism—at matched recall it roughly doubles precision over a fixed budget on the larger pools, the margin widening as the pool grows—while the foreign-key traversal pays off on the join-critical bridge tables dense retrieval cannot reach, in the single-hop tight-budget case. We characterize the method as a low-cost, recall-robust first stage for schema retrieval on which a downstream selection or generation step can build.

[IR-3] Living Context Graphs: Continual Semantic Learning for Enterprise Data Reasoning NEURIPS2026

链接: https://arxiv.org/abs/2611.20976
作者: Varun Agrawal,Ayush Srivastava,Eshan Gujar,Kaushal Kumar,Ayush Gupta
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Databases (cs.DB); Emerging Technologies (cs.ET); Information Retrieval (cs.IR)
备注: 22 pages, 5 figures. Accepted at the NeurIPS 2026 Workshop on Continual Learning in the Era of Foundation Models and Embodied Agents. Data splits and annotations: this https URL

点击查看摘要

Abstract:Enterprise AI can now perform analytics, reporting, and data-driven workflows, yet the data foundations it relies on remain hard to operationalize: business meaning, relationships, metric definitions, and reporting conventions are absent from database schemas and scattered across undocumented systems and tribal knowledge. We introduce the Living Context Graph (LCG), a continually learned context layer that makes existing enterprise data operational for AI without requiring a copy or transformation of the underlying warehouse. LCG bootstraps base semantics and relationships from the data itself, builds a concept graph that indexes business context as queries are served, writes durable semantic learnings from verified outcomes, and withdraws or re-earns context when later outcomes contradict prior routes. We evaluate LCG through text-to-SQL, a core data-reasoning task, on a transfer-aware split of three BIRD databases comprising 499 questions: 353 form the learning history and 146 are held out and never used to update the context layer. Bootstrapping alone improves overall accuracy by +4.6 percentage points. Learning only from the history raises accuracy on the untouched held-out transfer set by +22.6 points (95% CI [+14.4, +30.8]), exceeding five-shot retrieval from the same history by +12.3 points. Incremental learning reaches the one-shot endpoint within observed evaluation noise on every database. Ablating the graph changes held-out accuracy by at most one question per database but increases per-query cost 2.3-3.3x and median latency 1.8-3.4x: continually learned semantics drive accuracy, while the graph makes that context efficient to serve.

[IR-4] Evidence Diet: When Graph Evidence Helps LLM s Rank Competing Candidates

链接: https://arxiv.org/abs/2611.20837
作者: Yuchong Hu,Zhao Zhang,Ziyao Zhao,Haoyu Wang,Jing Xiao
类目: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:When several candidates have graph support, does it matter which candidate receives which supporter and path counts? We study how LLM rankers use collaborative evidence on Amazon Books and MovieTV. Swapping user-level evidence degrades ranking, while matched item substitutions have little effect. On fixed clean-trained checkpoints, unrelated catalog text lowers Hit@1 from 44.77% to 38.70%, with 99.9% of rankings still valid. Across all three paired seeds, the NDCG@5 difference between count and path prompts was larger after count training than after path training. Degree matching preserves target paths while giving more rivals support. Across the full degree-matched development pools, adding candidate-level graph records to history improves NDCG@5 by 2.28 points on Books and 2.54 on MovieTV. When the target and a rival both have support, path- and count-trained rankers benefit from retaining counts. On these development queries, adding counts numerically to history scores outperforms presenting them as text. For count-trained rankers, exchanging intact count pairs among supported candidates preserves support and token length, yet lowers NDCG@5 by 4.06 points on 141 Books queries and 3.26 on 104 MovieTV queries. A frozen evaluation on 274 previously unused Books users with supported targets and rivals confirms the attribution effect: correct assignment improves pairwise accuracy by 6.07 percentage points (95% CI [4.18, 8.00]).

[IR-5] UniGraph: A Unified Graph-based Index for Dynamic Hybrid Dense-Sparse Vector Retrieval ICDE2027

链接: https://arxiv.org/abs/2611.20758
作者: Keyu Chen,Xiaoyao Zhong,Jiabao Jin,Peng Cheng,Wangze Ni,Jingkuan Song,Heng Tao Shen
类目: Information Retrieval (cs.IR); Databases (cs.DB)
备注: Under ICDE 2027 Review, 14 pages

点击查看摘要

Abstract:Hybrid dense-sparse vector retrieval combines semantic signals from dense representations with lexical signals from sparse representations, typically ranking items by a query-dependent weighted sum of dense and sparse vector distances. Existing systems commonly search separate dense and sparse vector indexes and rerank the union of their candidates, incurring two retrieval passes and potentially discarding items that rank highly only after the semantic and lexical signals are combined. Constructing a graph-based index directly under a hybrid distance can mitigate this premature candidate loss. However, because varying the hybrid weight alters the underlying neighborhood relationships, a single static graph built at a fixed weight cannot accommodate dynamic query-specific weights. Therefore, we propose a Unified Hybrid Graph (UniGraph), which merges neighbor sets computed at multiple hybrid weights into a single bounded-degree graph index, enabling hybrid vector retrieval with different query-specific weights without index reconstruction. We further develop an adaptive weight-guided search framework that chooses between sparse- and dense-guided entry points according to the query weight, together with a dense-first hybrid distance pruning strategy that avoids unnecessary exact sparse distance computations. Extensive experiments show that UniGraph provides a favorable Recall-QPS trade-off across different query weights. In the mixed-weight setting, UniGraph delivers 27.6%~52.3% higher QPS than the fastest baseline on each dataset at Recall@100 greater than or equal to 0.95. UniGraph has been deployed at scale in Ant Group’s production environment, where hybrid retrieval with query-dependent weights is in strong demand.

[IR-6] CARE: Contrastive-Anchored Residual Encoding with Listwise Reconstruction for Retrieval-Oriented Semantic IDs

链接: https://arxiv.org/abs/2611.20721
作者: Yuan Xu(Tencent),Dawei Chen(Tencent),Xiaoguang Lin(Tencent),Xin Chen(Tencent),Yulong Ou(Tencent),Haijun Wu(Tencent)
类目: Information Retrieval (cs.IR)
备注: 18 pages, 8 figures, 5 tables

点击查看摘要

Abstract:Generative retrieval replaces nearest-neighbor search with the generation of a document’s semantic ID (SID), so the SID tokenizer bounds what the retriever can recall. Common SID constructions (residual k-means, RQ-VAE, finite scalar quantization, capacity balancing) optimize criteria defined on document embeddings alone and do not control whether a query can still find its relevant documents after quantization. We propose two retrieval-oriented SID tokenizers. CARE-SID (Contrastive-Anchored Residual Encoding for Semantic IDs) quantizes in a retrieval-aligned space: a zero-initialized residual alignment network is trained with a query-document contrastive loss before quantization, an anchor loss and a reconstruction target fixed to the original embedding prevent drift and collapse, and reference-vector rating residual quantization assigns the codes. CARE-LR (CARE with Listwise Reconstruction) then distills, for each training query, the similarity distribution that the original embeddings induce over a candidate list into the SID reconstructions, through a function-preserving layer-resolved head. Both keep a document-only encoding interface. We judge SIDs without training a generator, by row-budget recall in a bucketing view (the partition) and a reconstruction view (the SID with its reconstruction). In offline evaluation on industrial video search data, both tokenizers achieve higher reconstruction-view recall than residual k-means, and CARE-LR improves further on CARE-SID, mainly through the listwise distillation. A comparison of two document embeddings shows that contrastive alignment helps most when the embedding has not been trained on search data, and that the embedding affects SID recall far more than the quantizer. CARE-SID is deployed in the generative retrieval stage of an industrial video search system.

[IR-7] When the Bill Falls: Reusing the Old Index DefersNever Avoids Re-encoding in Complete Embedding-Model Migration

链接: https://arxiv.org/abs/2611.20702
作者: Nannan Zhao,Jiameng Zhang,Taoyu Zhong,Hao Wang,Bo Wang
类目: Information Retrieval (cs.IR); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Upgrading the embedding model behind a vector index requires re-encoding the corpus. Recent work reuses the old index through adapters or window re-ranking. Across sixteen model pairs on passage and code corpora, plus eleven on Natural Questions (NQ), we find three results. First, reuse never avoids re-encoding: if migration must complete, every strategy eventually encodes every document; reuse only changes when that work is paid and the quality served meanwhile. Second, adapters should be compared with the old model, not only the new one. Of five adapter variants, only same-family LLM bridges beat the old model, and only by a negligible margin. Third, window re-ranking cost can be priced from the old index alone; pre-registered replays on all 8,841,823 passages confirmed the prediction on all three tested pairs. Workload skew also matters: across 36 further pre-registered Zipf replays, the planned family matched on 32. These findings yield three pre-migration rules: whether reuse is worthwhile, which strategy to use, and how deep a window to take. The depth rule priced every window that reached target quality to within 4 parts in 10^5 of the corpus. On NQ, a pre-committed plan selected the family on 10 of 11 pairs and priced every window exactly. Quality inputs transferred less reliably across corpus scale, so the model gap and window quality loss must be measured on the serving index.

[IR-8] MAGE: Meta-path Adaptive Graph Evidence for Transferable Multimodal Graph Learning

链接: https://arxiv.org/abs/2611.20449
作者: Haiyang Jiang,Tong Chen,Xinyi Gao,Hao Wang,Hongzhi Yin
类目: Information Retrieval (cs.IR)
备注:

点击查看摘要

Abstract:Multimodal graphs integrate textual and visual attributes with relational structures, providing a rich representation of complex real-world data. To enable the broad applicability and reuse of graph models, it is crucial for models trained on source graphs to generalize to unseen graph domains. However, this is challenging because cross-modal interactions among neighbors and the neighboring information needed for prediction can vary across target nodes and graph domains. This gives rise to two key challenges: adaptive cross-modal neighborhood fusion, which captures and integrates cross-modal interactions among neighbors to accommodate variations across nodes and domains; and target-specific evidence selection, which identifies the neighboring textual and visual information needed for each target node to support LLM-based prediction. In light of these challenges, we propose MAGE (Meta-path Adaptive Graph Evidence Prompting), a framework for transferable multimodal graph learning. MAGE represents textual and visual node views separately in a heterogeneous graph and introduces two complementary components. Meta-path Adaptive Multimodal Propagation (MAMP) adaptively aggregates neighborhood information along heterogeneous meta-paths and compresses it into graph-aware soft tokens. Evidence-Guided Instruction Tuning (EGIT) selects relevant neighboring evidence for each query and incorporates its graph-aware representations into instructions for a frozen multimodal LLM. Together, these components support transfer from multiple source graphs to unseen target graphs without target-domain parameter updates. Experiments with three source graphs and four unseen target graphs demonstrate strong cross-domain transfer performance on node classification and link prediction, without requiring parameter updates on the target domains.

[IR-9] IntentFlow: An Intent-Driven Real-Time Multimodal Search System with Image-Text Queries at Xiaohongshu

链接: https://arxiv.org/abs/2611.20448
作者: Dongjia Zhao,Yuchen Ma,Chunsheng Shi,Ziqiu Gui,Chao Yuan,Jianing Zhu,Zhaoxing Zhang,Yahui Liu,Daoxin Zhang,Yi Wu,Yao Hu,Cheng Chen
类目: Information Retrieval (cs.IR)
备注: 29 pages, 7 figures. Technical report

点击查看摘要

Abstract:Multimodal search with image-text queries helps users find notes in user-generated content (UGC) communities and serves as a key search tool for multimodal search agents. For such queries, query-note relevance depends on the user’s retrieval intent and cannot be determined by visual or topical similarity alone. Simply reranking visual search results cannot fully address this challenge, as notes relevant to the user’s intent may be missing from the initial candidate set. This motivates explicit intent modeling to align candidate retrieval and reranking under consistent relevance criteria. We present IntentFlow, an intent-driven real-time multimodal search system deployed at Xiaohongshu over an index on the order of ten billion entries. IntentFlow represents retrieval objectives with a hierarchical intent taxonomy and uses intent-specific relevance criteria to guide data construction, representation learning, and reranking. An offline agentic data workflow acquires candidates and verifies their relevance under these criteria to produce training supervision. The embedding model combines intent-aware contrastive learning, graded ranking, and distribution distillation, while a compact reranker assesses query-candidate evidence conditioned on predicted hierarchical intents. Evaluations demonstrate that IntentFlow delivers substantial gains in search engagement while supporting real-time search at industrial scale. In a user-level A/B test on the Xiaohongshu platform, IntentFlow outperforms the existing production system based on visual search followed by reranking, achieving relative gains of 28.17% in Top-4 query click-through rate, 18.96% in query click-through rate, and 12.56% in average time spent on search results pages per user. Production monitoring records an end-to-end P50 latency of 1.07 s.

[IR-10] Overview of the Plagiarism Detection Task at PAN 2026 DATE

链接: https://arxiv.org/abs/2611.20443
作者: André Greiner-Petter,Yun-Ang Wu,Yuka Kitamura,Kon Woo Kim,Maik Fröbe,Bela Gipp,Akiko Aizawa,Martin Potthast
类目: Information Retrieval (cs.IR)
备注: Corrected version of the paper published in CEUR-WS Vol-4283: an error in the one-shot generation dataset was fixed and the one-shot results (Sec. 2.3, Table 3, Figs. 2, 3, 6) were updated; the sequential dataset and its results are unchanged. DOI: https://doi.org/10.5281/zenodo.23528623 (Version v3)

点击查看摘要

Abstract:The generative plagiarism detection task at PAN 2026 aims at retrieving the original scientific sources for given automatically generated textual plagiarism. We created two novel datasets of synthetically generated plagiarism following the premise of a malicious actor trying to generate papers with minimal effort while evading detection of undisclosed re-use. We utilized four widely available LLMs: GPT-OSS 120B, Llama 3.1 8B, Qwen 3.5 27B FP8, and DeepSeek-R1 Distill Llama 8B. In this task overview paper, we outline the creation of this dataset and summarize and compare the results of all participants with three baselines. For this year’s task, we received 26 submissions from six teams that significantly outperformed our baselines.

[IR-11] What Unlinked Aggregates Still Prove: Ranker Comparisons from Exact and Differentially Private Item Counts

链接: https://arxiv.org/abs/2611.20388
作者: Chang Xu,Hao Wang,Gedi Zhou
类目: Information Retrieval (cs.IR)
备注: 10 pages, 4 figures. Accepted for publication in the Privacy and Security of Big Data special session at IEEE BigData 2026. Code available at this https URL

点击查看摘要

Abstract:Privacy controls can prevent an evaluator from linking purchases to the recommendations users received, while a merchant can still release complete item-level purchase counts for the evaluation cohort. We ask when these counts identify which recommender performs better. The counts constrain who could have bought each item. We bound the performance difference over all assignments consistent with the counts. A reduced transportation network handles full catalogs and can certify attainable endpoints. If the interval excludes zero, it identifies the better recommender for that cohort. With differentially private counts, one confidence guarantee applies jointly to all comparisons from the same release. In two e-commerce cohorts of about 20,000 and 100,000 users, exact counts narrowed every interval compared with bounds that ignored the counts. Most decisions involved a recommender that favors popular items. For NDCG@10, which rewards ranking purchased items higher, each setting combined one of three personalized pairs with a pattern and rate of missing links. Exact counts settled 14 of 168 settings in the smaller cohort and none of 81 in the larger cohort. In the smaller cohort, exact counts broken down by user group settled two personalized pairs when 90% of users remained linked, at the cost of revealing more about individual purchases. Whole-cohort private releases with one count per item retained much of the narrowing but settled no personalized pair in either cohort. Grouping less popular items into bins let private counts settle one personalized pair in the smaller cohort.

[IR-12] Straight to the Product: What Online Retailers Miss When Shoppers Ask AI First

链接: https://arxiv.org/abs/2611.20378
作者: Michael Iannelli,Alan Ai
类目: Information Retrieval (cs.IR); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
备注: 5 pages, 3 figures

点击查看摘要

Abstract:When people ask an AI assistant what to buy, what happens next at the store? We link the same people’s assistant conversations to their later browsing at twenty large online retailers. People go to the product the assistant named: among shoppers who had not visited the store in the past week, visits to the named product within a day rise from 0.02 to 2.07 per 100 answers (+2.0 points [1.7, 2.4]), and on days with a product-page visit at the store, 44% show a product from their own answer, against 1.5% for a matched answer given to someone else. The result replicates in a separate five-month period. Few arrive through the assistant’s link: 89% of visits carry no sign of it, and 54% follow a web search, the same share as their usual visits. When a search comes just before the first visit, it names the answer’s product 28% of the time (1% in earlier weeks), as if the search box were an address bar. The chat names the product; the store logs a search. Once there, shoppers keep browsing. We propose product follow-through, the share of answers followed by a visit to the product they named, as a way to evaluate shopping assistants and agents by what people do next: it needs no referral data or provider logs and can be computed for each assistant. Comments: 5 pages, 3 figures Subjects: Information Retrieval (cs.IR); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC) ACMclasses: H.3.3; K.4.4 Cite as: arXiv:2611.20378 [cs.IR] (or arXiv:2611.20378v1 [cs.IR] for this version) https://doi.org/10.48550/arXiv.2611.20378 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[IR-13] How did data sharing evolve in the decade since the FAIR Principles?

链接: https://arxiv.org/abs/2611.20319
作者: Sheeba Samuel,Daniel Mietchen
类目: Information Retrieval (cs.IR); Digital Libraries (cs.DL)
备注:

点击查看摘要

Abstract:A decade after the publication of the FAIR Guiding Principles, what have we learned about data sharing, and what do data availability statements (DAS) and other materials accrued over that period tell us about how data-sharing practices, attitudes, and infrastructures evolved? Here, we present a large-scale longitudinal analysis of over 6.8 million biomedical articles from the PMC Open Access Subset, spanning 2011 to 2025. Using an automated pipeline, we extracted and classified DAS, repository links, accession numbers, and supplementary materials, assigning each article a data-sharing score and independently FAIR-assessing a subset of cited dataset URLs. We find that explicit DAS adoption rose from ~2% in 2011 to 68% in 2025, while “available on request” statements grew in parallel from near zero to over 19%. The corpus growth amplifies both patterns: articles with a DAS containing an explicit repository mention rose from 132 in 2011 to 94,697 in 2025, while “available on request” articles rose from 77 to 173,474 over this span. Structural biology and genomics consistently achieved the highest data-sharing scores, while clinical research and public health ranked lowest. Gold Open Access journals showed a small advantage over subscription journals, while Diamond Open Access journals outperformed both. Internationally collaborative articles scored higher than domestic ones. A gender gap emerged: female first-authored articles’ scores declined more than male first-authored ones. The analysis further covered additional dimensions, from compliance with funder/journal mandates and JATS XML tagging standards to computational reproducibility signals and the co-occurrence of DAS with ethics statements. A decade on, the gap between declaring data availability and actually sharing machine-readable, reusable data persists across fields, infrastructures and demographics.

[IR-14] EchoSpec: Training Search Agents to Weigh Evidence Above Echoes

链接: https://arxiv.org/abs/2611.20034
作者: Shihao Yang,Can Zeng,Xinzhe Li
类目: Information Retrieval (cs.IR)
备注: 24 pages, 5 figures, 15 tables. Code: this https URL

点击查看摘要

Abstract:Search agents trained with reinforcement learning decide what to retrieve, when to search again, and when to answer. In real corpora, one claim is copied and paraphrased across many pages while the primary source settling a question appears once. We show that reinforcement-learned search agents grant echoes the credence of independent sources and lose their ability to abstain. Across five base models from 3B to 32B and seven released RL agents on multi-hop Wikipedia questions with primary evidence intact, eight paraphrased false-claim echoes overturn 58-73% of previously correct answers, while exact-match RL drives deleted-evidence abstention below 15%. Conditioning on retrieval localizes the failure to evidence acquisition: once primary evidence enters context, agents overrule eight echoes as reliably as one, whereas echoes fill top ranks and halt search early, so larger budgets, forced search turns, or passive passage filtering alone fail to recover displaced evidence. We introduce EchoSpec, specifying reliability across five corpus transformations with known target answers and training in two stages. Verification warm-start teaches multi-hop bridge-entity recovery past echoes, after which group RL optimizes an evidence-gated \lambda -reward with an outcome-independent acquisition bonus and scale-preserving advantages retaining the optimal rule \lambda/(1+\lambda) . Across Qwen2.5-3B, Qwen2.5-7B, and Llama-3.1-8B backbones on nine evaluation suites, EchoSpec cuts eight-echo deference by 30-44 points on Qwen2.5-7B down to 29-31%, nearly quadruples eight-echo reach to 26-27% with zero-shot transfer to unseen BGE and BM25 retrievers, and raises standalone QA precision to 47-55%. Code is publicly available at this https URL.

[IR-15] ake a Step Back: A Dual-Granularity Raw Feature Empowered Intermediate Fusion Graph Convolutional Network for Multimodal Recommendation WSDM2027

链接: https://arxiv.org/abs/2611.19974
作者: Jinfeng Xu,Zheyu Chen,Shuo Yang,Jinze Li,Wenhao Yuan,Ziyue Peng,Puzhen Wu,Shujie Li,Jian Chen,Zheng Lin,Edith C. H. Ngai
类目: Information Retrieval (cs.IR)
备注: Accepted by WSDM 2027

点击查看摘要

Abstract:Although multimodal recommendations have shown the ability to learn user preferences and item properties by leveraging explicit multimodal information, they remain constrained by two critical limitations: low-quality representation learning and modality fusion. For the former, existing solutions primarily focus on multi-granularity representation learning; however, we argue that training such representations is difficult due to the inherent data sparsity problem in recommendation scenarios. For the latter, existing solutions often rely on contrastive learning to align representations of different modalities to facilitate modality fusion, yet alignment-only objectives may over-compress valuable modality-specific information. To this end, we suggest \underline``Take a Step Back" and propose a \textbfDual-granularity raw feature empowered \textbfIntermediate fusion \textbfGraph \textbfConvolutional \textbfNetwork (\textbfDIGCN) for multimodal recommendation. DIGCN leverages the multimodal understanding capabilities of Large Language Models (LLMs) to construct dual-granularity raw features directly. Furthermore, DIGCN introduces an intermediate fusion graph convolutional network equipped with dynamic cross-modality weight correction, designed to capture cross-modality signals and achieve effective modality fusion. Comprehensive experiments on four real-world datasets validate that DIGCN consistently outperforms a variety of strong baselines. Additional analyses confirm its superiority under varying levels of data sparsity and cold-start scenarios. Notably, DIGCN demonstrates high extensibility, capable of further benefiting from additional contrastive learning-based alignment tasks.

[IR-16] DetouR: Plugging Association Routes into Search Agents for Multi-hop Question Answering

链接: https://arxiv.org/abs/2611.19967
作者: Tingwei Chen,Yunxiao Shi,Zhendong Chu,Zhiying Deng,Min Xu,Qingsong Wen
类目: Information Retrieval (cs.IR)
备注:

点击查看摘要

Abstract:Search agents answer multi-hop questions by interleaving reasoning with retrieval, deciding at each step what to search next. Existing agents follow semantic routes: they infer what is missing and phrase it as a query, reaching only what the model can name. Yet the decisive page is often one the question never names, frequently just one hyperlink away from evidence already retrieved. Following such association routes reaches more gold pages than semantic search under the same budget. Exploiting them raises two challenges: obtaining the association structure without costly offline graph construction, and keeping linked evidence in the reader’s limited context, where relevance- and fusion-based selection discards it. We propose DetouR, a plug-in that reuses the links arriving with retrieved pages and adds no search or LLM call. Link Rerouting replaces near-repeat queries with routes to linked pages, chosen by a lightweight learned scorer; Chain-aware Context Selection runs personalized PageRank over these links so that evidence chains reach the reader intact. Across four agents, DetouR improves EM by 4.0 points on 2Wiki and 1.4 on HotpotQA on average. We also find that heavier LLM models struggle with this non-semantic evidence, adding little over these simple selectors or even falling behind. In summary, our work reveals association routes as a cheap complement to semantic routes that current search agents overlook: added as a plug-in, they improve every agent we test, in replay and in live runs, at no extra search or LLM cost.

[IR-17] A Demonstration System for Faithful Temporal Question Answering

链接: https://arxiv.org/abs/2611.19926
作者: Zhen Jia,Xinghong Hou,Chenyang Jin
类目: Information Retrieval (cs.IR)
备注: 4 pages, 2 figures. Demo paper

点击查看摘要

Abstract:Temporal question answering (TQA) aims to answer questions that involve temporal constraints or require temporal information. We propose Faith, a novel TQA approach that operates over heterogeneous sources and ensures explainable answers by presenting traceable and faithful evidence. In Faith, we design a new recursive answering mechanism transforming implicit temporal constraints into explicit conditions. Additionally, we visualize the system functionality to track how the answers are computed and the supporting evidence is derived. The code and video of the demonstration system are publicly available.

[IR-18] Attribute or Abstain: Evaluating Evidence-Grounded Retrieval on the Low-Resource African Business Web

链接: https://arxiv.org/abs/2611.19900
作者: Bakary Gibba,Muhammad Abba Mandara,Karim Ismail Abd-Ellah Mohamed
类目: Computation and Language (cs.CL); Information Retrieval (cs.IR)
备注: 8 pages, 3 figures, 4 tables

点击查看摘要

Abstract:Retrieval-augmented systems that research companies are judged by how much they find, yet in markets where public evidence is sparse, a confident answer without support costs more than no answer. We introduce AfriBizClaims, a benchmark of 150 company-research queries across six African markets in English, French and Portuguese, including queries about fictitious companies that a correct system must decline. Each claim a system makes is judged against the page it cites by an AI judge, checked against a second AI judge from another model family and against two human annotators, and abstentions are scored against answerability labels. We compare BM25 over a pooled corpus, standard web retrieval-augmented generation, and a multi-step agent pipeline (Konexium V1), all using the same language model. The two simple baselines are well attributed: 98% of their judged claims are supported by the page they cite, and they decline every field of every fictitious company. Counting claims about a company other than the one asked about as errors, the agent pipeline’s per-query precision is 68%, and it never answers a question about a named company correctly: it substitutes other firms, including for 5 of the 24 fictitious names, and it is the least stable between runs and the most expensive. In the pipeline we evaluated, more steps between the query and the evidence did not buy more grounded answers. We will release the queries, system outputs, labels and short evidence excerpts.

[IR-19] NoLBERT-Calls: Reading Earnings Calls Without Hindsight

链接: https://arxiv.org/abs/2611.19849
作者: Kevin Foley,Jonathan Hartadi,Shivesh Prakash,Hino Wirachapong Suwanphibun,Swapnil Vatsal
类目: Computation and Language (cs.CL); Information Retrieval (cs.IR); Machine Learning (cs.LG); General Finance (q-fin.GN)
备注:

点击查看摘要

Abstract:Language models are now a standard tool for measuring what managers say on earnings calls, but the encoders behind these measures were usually pretrained on text from the same years a backtest covers. Holding out a test period does not remove that hindsight, because it sits in the features. We adapt NoLBERT, an encoder pretrained only on text published between 1976 and 1995, to earnings calls without letting it read anything dated after 31 December 2023. The calls supply the training signal themselves: every analyst question is followed by management’s answer, and the model learns to pick that answer out from the other answers of the same call. We test on every US call from January 2025 to October 2026 (20,323 calls, 254,575 questions) and make every modeling choice on calls from 2024. The adapted model, NoLBERT-Calls-2023, ranks the true answer first for 78.0% of questions (95% interval 77.8-78.2; chance 8.0%), against 26.3% before adaptation. It is 1.0 point ahead of a call-adapted ChronoBERT, 3.4 points ahead of a call-adapted FinBERT, and 18.7 and 19.8 points ahead of the best off-the-shelf retriever and BM25. A year-by-year test over 2013-2026 finds no look-ahead premium: relative to a model adapted only through 2012, the released model does no better on the years it was trained on than on years it never saw. The price of the cutoff is compute: under one shared 150-minute recipe NoLBERT adapts least well of four encoders, and its lead comes from a longer, staged adaptation of about 13 GPU-hours.

[IR-20] Reflection Is Not Always Worth It: Budgeted Reflection Allocation for Recommendation Agents WSDM2027

链接: https://arxiv.org/abs/2611.19847
作者: Jinfeng Xu,Zheyu Chen,Ziyue Peng,Jianheng Tang,Zheng Lin,Wenhao Yuan,Jian Chen,Shujie Li,Victor Leung
类目: Information Retrieval (cs.IR)
备注: Accepted by WSDM 2027

点击查看摘要

Abstract:This work aims to make recommendation agents more efficient by reducing unnecessary LLM reflection while preserving full-catalog ranking and inference reranking. Since BRAVE changes only how training-time reflection is allocated, it does not introduce new user data requirements or additional LLM exposure beyond the base agent protocol. The main potential risk is that a poorly calibrated allocator may assign less reflection to users whose preferences are sparse, niche, or under-represented, which could weaken personalization for those users or indirectly affect item exposure. This risk can be mitigated through deployment-time auditing and conservative allocation design. In practice, reflection budgets should be evaluated not only by aggregate Recall or NDCG, but also across user-history lengths, item popularity groups, and application-relevant cohorts when such analysis is permitted. The allocator should rely on task-side ranking evidence rather than sensitive attributes, and fallback policies can reserve a small portion of the budget for random or fairness-aware coverage when subgroup robustness is important. Privacy risks can be controlled with the same protections required by the base recommendation agent: minimizing user text sent to the LLM controller, anonymizing user identifiers, retaining only necessary interaction evidence, and applying standard access-control and logging policies. Overall, BRAVE is designed as a cost-aware allocation layer, and its ethical deployment depends on validating that the saved computation does not come at the expense of vulnerable users or responsible exposure. Comments: Accepted by WSDM 2027 Subjects: Information Retrieval (cs.IR) Cite as: arXiv:2611.19847 [cs.IR] (or arXiv:2611.19847v1 [cs.IR] for this version) https://doi.org/10.48550/arXiv.2611.19847 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[IR-21] Exact Sparse Assignment with Costly Target Embeddings: Fixed-Realization Certificates and an Empirical Audit

链接: https://arxiv.org/abs/2611.19770
作者: Hamidreza Hasani Balyani,Mahtab Movahhedrad,Seyed Pouyan Mousavi Davoudi
类目: Machine Learning (cs.LG); Information Retrieval (cs.IR)
备注: 7 pages, 2 figures. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:Embedding-based one-to-one association is common in vision and data integration: detections are associated with tracks and records are linked across databases. Such pipelines often solve a sparse optional assignment whose edge scores add a cheap term to an inner product of neural embeddings. When target embeddings are costly, one can acquire them selectively and stop once a certificate proves that the current assignment is optimal. We give such a certificate for unit-norm embeddings. It bounds the unknown part of every reassignment by a support function and maximizes the resulting robust gains with a second sparse assignment; a shared owner-centred bound dominates any fixed-centre triangle bound. We make precise what is certified when network features are not batch-invariant: the realized values of the acquired rows and every unit-ball completion of the rest, not a second forward pass in a different batch. Auditing against a separately batched cache tests a different property; on our first pilot it produced a spurious 2.49\times10^-6 objective failure. We audit the corrected protocol on four public text entity-matching benchmarks with two frozen encoders; no image data is used. All 96 certificate simulations and 120 online selective timings were exact on their own realization (largest objective gap 6.82\times10^-13 ). At group size 8 the shared certificate acquires on average 60% of candidate-active targets, against 74% (mean centre), 69% (MEB) and 60% (zero centre): it beats the mean and MEB centres but not the zero centre, and savings concentrate in one benchmark (7% acquired). The prespecified runtime claim is contradicted: selective execution is +0.510 s slower per run (dataset-cluster 95% interval [+0.248, +0.772]). Uniform-microbatch re-encodings preserved the optimum in only 5 of 8 encoder-dataset cells, and 45 of 120 online runs disagreed with an eager cache at near-ties.

[IR-22] StaleBench: Isolating Stale Retrieval in Agent Memory and Pricing the Fix EACL2027

链接: https://arxiv.org/abs/2611.19752
作者: S M Mahbub Murshed
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Information Retrieval (cs.IR)
备注: Accepted to Findings of EACL 2027. 39 pages, 3 figures. Code and data: this https URL

点击查看摘要

Abstract:An agent that stores facts over time eventually stores several versions of the same fact. Asked a value-agnostic question - “which team does X play for now?” - a similarity retriever cannot prefer the current version, because every version matches the entity the query names. StaleBench isolates this stale-retrieval failure on 892 real updating facts, with currency labels from the source (Wikidata validity years via TempLAMA), never from a retrieval score. Our central result is that retrieval metrics can rank memory policies backwards: a retention policy and its retain-everything control return the identical statement at rank 1 for every test query, yet differ by 29.5 and 65.6 points in reader accuracy on two readers when no date is shown. Contamination@5, the metric we add, ranks them correctly where current@1 and current@5 do not. Withdrawing the lexical anchor that statements carry costs deployed keying mechanisms tens of points of freshness; entity resolution on both the write path and the query recovers most of the loss. Three deployed systems (Mem0, Graphiti, LangMem) and a reimplemented ERASE each handle supersession differently, and no single scoring convention describes all four. We release the data, a frozen evaluator, and reference baselines.

[IR-23] User First Intelligence: Three Measurements of Search and AI-Answer Analytics (SEO and GEO/AEO) and a Knowledge-Graph Framework for Query Discovery Beyond Search Volume

链接: https://arxiv.org/abs/2611.19655
作者: Venkata Pagadala(KL University)
类目: Information Retrieval (cs.IR)
备注: 43 pages, 10 figures. Preprint, not peer reviewed. Data and scripts in the ancillary files; implementation at this https URL

点击查看摘要

Abstract:People say what they want to know in many places besides a search box, and the tools that search and AI-answer optimization (SEO and GEO/AEO) rely on see only part of that demand, through models reported as counts. This paper reports three measurements of such instruments and a framework. On the area-code section of a large US telecom provider’s site, two commercial analytics platforms differed on estimated monthly visits by 7.5x to 27.1x (8.6x to 29.9x worldwide); across twelve sections of the domain the ratio ran from 0.20x to 7.24x with no detectable relation to section age, while they agreed on which section is larger in 60 of 66 pairs. On 37,193 reviews of 378 telecom stores linked to a 42,994-node keyword taxonomy, review demand and search volume agreed on branches to a varying degree (Spearman 0.47 to 0.75; 0.91 post hoc) and barely on topics (0.18), on links two blind AI annotators judged two-thirds correct; on the 372 area codes, a public count of Wikipedia requests for each code and vendor volume agreed roughly on ranking (0.45 over 240 articles, gone once keywords matched per code are held constant) and not on magnitudes, predictions fixed first. In one vendor’s index of AI answers naming or citing the largest companies of each SP 500 sector, of four predictions written first, one held, one in direction only and two failed; video and community sites took a median 21% of citation incidences on Google’s AI Overviews and AI Mode against 0.6% to 9.2% on four other surfaces, in every sector, and in education answers in each of 13 countries read later. Estimates and sources need measuring per instrument and surface. User First Intelligence (UFI), built on that premise, collects what people ask beyond the website, links it to an industry ontology and scores where an organization does not answer; an open implementation is released, its score not evaluated. No causal claim is made.

[IR-24] Keep it Simple! A Multiple-Choice Task can Benefit Generative Recommendation

链接: https://arxiv.org/abs/2611.19320
作者: Wei-Hsiang Huang,Jyun-Yu Jiang,Pu-Jen Cheng
类目: Information Retrieval (cs.IR)
备注:

点击查看摘要

Abstract:Sequential recommendation has recently witnessed growing interest in generative retrieval. One of the most popular approaches is to autoregressively generate the semantic identifiers (SIDs) of items as recommendations. Existing efforts mainly rely on semantic alignment, sophisticated architectural designs, and extensive hyperparameter tuning. However, the conventional learning framework for generative models could have foundational limitations. Specifically, most generative models directly predict target items without explicitly learning negative compositions and relative ranking signals. This is not well aligned with the philosophy of Bayesian personalized ranking (BPR), which emphasizes the importance of relative rankings between positive and negative samples. In this paper, we show that adding a simple multiple-choice auxiliary task into the learning objective can effectively address this issue. With minimal adaptation, we introduce Task Identifiers and Candidate Constructions to synthesize training data for the multiple-choice auxiliary task. In addition, we provide theoretical analysis of the fundamental benefits of the multiple-choice auxiliary task from the perspectives of information theory and compositional generalization. Extensive experiments on four publicly available benchmark datasets show that our approach can improve TIGER, a conventional generative recommendation model, by 4.9% to 79% in NDCG@10. Subjects: Information Retrieval (cs.IR) Cite as: arXiv:2611.19320 [cs.IR] (or arXiv:2611.19320v1 [cs.IR] for this version) https://doi.org/10.48550/arXiv.2611.19320 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[IR-25] Revisiting What Makes a Good Context Subset: A Sparse Latent Basis Perspective ACL

链接: https://arxiv.org/abs/2611.19220
作者: Cen-Jhih Li,Siddhant Chaudhary,Aditya Bhaskara
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Information Retrieval (cs.IR)
备注: 17 pages, 2 figures. The format of ACL rolling review

点击查看摘要

Abstract:Selecting context subsets is increasingly important in modern LLM inference, yet what makes a subset effective remains unclear. We study this question in the controlled setting of few-shot demonstration selection. We examine a hypothesis that demonstrations should collectively provide task-relevant information needed by a target query. Assuming that the query representation provides a useful proxy for these requirements, we formulate selection as sparse reconstruction and instantiate it with Residual Explanation Search (RES), a Matching-Pursuit-style diagnostic procedure. Across seven benchmarks and eight open-weight LLMs, no retrieval objective is universally best. RES outperforms MMR in 34 of 56 model–dataset pairs, with its clearest advantage on sentiment tasks, suggesting that query-conditioned collective reconstruction may capture useful complementarity that pairwise dissimilarity misses. Relevance-based retrieval remains strong on template-heavy reasoning benchmarks, while all retrieval methods underperform zero-shot prediction on average on StrategyQA, where a solution-reconstruction analysis is consistent with the query not exposing the required external knowledge. Target-model hidden states further show that retrieval objectives interact differently with representation spaces: middle-layer representations generally favor similarity retrieval, whereas last-layer representations are often stronger for residual- and diversity-based selection. These findings support a conditional, representation-dependent account of context subset quality rather than a universally superior selector.

[IR-26] LABEL: Literature Asymmetry between Efficacy and Limits in Drug Knowledge and Language Models

链接: https://arxiv.org/abs/2611.19166
作者: S M Rayeed,Charles V. Stewart,Yao Ma
类目: Artificial Intelligence (cs.AI); Information Retrieval (cs.IR)
备注: 27 pages (9 main + supplementary material), 8 figures, 7 tables

点击查看摘要

Abstract:Every medication comes with two types of medical facts: indications, which describe what conditions the drug treats (benefits), and contraindications, which describe who must avoid it to prevent serious harm (safety limits). Although contraindications are critical for patient safety and FDA regulations require listing them as rigorously as indications, we find that contraindications are far less documented than indications in general prose. In this paper, we measure how often these texts contain both facts across 574 drugs from PrimeKG by comparing a drug’s indications against its own contraindications. Across Wikipedia articles and PubMed abstracts, we find that reference prose names a drug’s benefits about twice as often as its safety limits (within-drug odds ratio 2.19). We call this documentation gap LABEL (Literature Asymmetry between Efficacy and Limits). Language models mirror this gap. As we scale Qwen2.5 from 0.5B to 14B, accuracy on indications rises by 18.3 points, whereas accuracy on contraindications remains flat near random guessing. However, when we provide models with a retrieved passage that states the answer, they use safety facts just as effectively as benefits. Furthermore, routing safety questions to official FDA labels boosts accuracy by 15.8 points whenever the label states the fact. Overall, our findings show that models fail on medication safety not because they cannot reason about risk, but because free-text corpora lack safety facts. Clinical AI systems should therefore route medication-safety questions to official regulatory labels and structured databases rather than relying on general public records. Code and data are available at this https URL.

[IR-27] CLEAR: Legible and Steerable Recommendation via Grounded Natural-Language Interests

链接: https://arxiv.org/abs/2611.19110
作者: Qi Guo,Reza Mirghaderi,Iordanis Fostiropoulos,David Fan,Yuchen Liu,Shiliang Tang,Lukasz Korycki,Golnaz Ghasemiesfeh,Xinzhu Bei,Sasha Vorontsov,Hanchao Yu,Brendan Quinlivan,Khoi Pham,Jigar Gada,Hao Yu,Felix Li,Hang Gao,Boyu Fang,Yang Gao,Yale Thomas,Xiaomeng Jin,Avishek Dutta,Hong Hong,Janet Xia,Ethan Zhu,Ellie Pan,Xiang Shen,Jawwad Zakaria,Jinfeng Lin,Jonathan Rush,Ruixue Lian,Noah Kennedy,Brian Vanover,Jas Pyneni,Daniel Thiyagu,Apoorv Saxena,Sanyogeeta Lawande,Huayu Ding,Hima Cherukuru,Yang Liu,Chen Kovacs,Qunshu Zhang,Lizhu Zhang,Benyu Zhang,Kate Chen,Ashish Singal,Danny Kanoff,Zhiyu Liu,Jiayi Liu,Jianyu Wang,Fei Liu,Xiangjun Fan,Abhishek Kumar,Hong Yan
类目: Artificial Intelligence (cs.AI); Information Retrieval (cs.IR)
备注: 18 pages, 6 figures

点击查看摘要

Abstract:Modern recommenders keep learned preferences in representations people cannot inspect or change. We present CLEAR (Controllable and Legible End-to-end Architecture for Recommendation), an end-to-end system deployed at Meta and built around an editable natural-language interest layer. CLEAR makes a ranked subset of inferred interests readable and stores accepted additions and suppressions as authoritative preference state. Behavioral grounding maps each inferred interest back to supporting events in the user’s own history. The resulting interest object is also a serving address: its label and profile-conditioned rewrites drive lexical retrieval, its supporting events seed behavioral retrieval, its projection queries an embedding index, and its stable identity supports ranking, attribution, and exposure control. A fresh user declaration may precede behavioral support, so it enters through the same serving identity and initially draws candidates from topical retrieval. Ranking makes those candidates competitive, while a delivery controller observes attributable exposure and replenishes underexposed interests. Direct user review found that users rated about 88.5% of displayed interests accurate. In an offline benchmark using a shared evaluation protocol, the Specialized 3B generator reached 0.3967 mean-user F1@10, compared with 0.3633 for Gemini 2.5 Pro and 0.1911 for the prior generator. For a newly declared interest, a user-randomized cumulative ablation measured same-day delivery losses of 8.7 percentage points for added-interest ranking terms, 11.1 for reserved top-slot insertion, and 13.7 for interest-conditioned candidate sourcing, conditional on the tested removal sequence. CLEAR components have operated in production for over one year; the system spans four surfaces on Instagram and Facebook, with an interest layer people can inspect, correct, and use to direct recommendations.

[IR-28] Knowledge-Guided Local-to-Global Anomaly Learning for Longitudinal Medication Trajectories

链接: https://arxiv.org/abs/2611.19106
作者: Wendi Ji
类目: Machine Learning (cs.LG); Information Retrieval (cs.IR)
备注: 16 pages, 3 figures

点击查看摘要

Abstract:Detecting suspicious deviations in longitudinal medication trajectories requires distinguishing knowledge-consistent treatment variations from anomalous changes and linking local deviations to account-level decisions. We propose a knowledge-guided local-to-global (L2G) anomaly learning framework that addresses both challenges. It constructs proxy supervision at three granularities: medication occurrences, treatment episodes, and contiguous visit segments. Knowledge-consistent medication and episode replacements provide plausible negatives, while cross-category and cross-account perturbations provide suspicious proxies. A medication-multiset encoder and causal visit-sequence encoder feed a shared scoring head that produces visit-level anomaly logits. Top-(\alpha) aggregation derives account-level scores from these logits, while account-level classification and class-balanced visit-localization objectives jointly optimize the same visit-level evidence. Experiments on two city cohorts compare L2G with native and proxy-adapted anomaly methods. On City-A, L2G outperforms the strongest comparators. Cross-city results show more consistent gains in localization than in account-level detection.

[IR-29] okenFormer V2: Breaking the Resolution-Complexity Dilemma in Unified Recommendation

链接: https://arxiv.org/abs/2611.19074
作者: Kaihui Wu,Guoming Li,Liang Guo,Yifeng Zhou,Yuehong Hu,Junwei Pan,Zhixiang Feng,Chengguo Yin,Wenqi Fan,Haijie Gu,Jie Jiang
类目: Information Retrieval (cs.IR); Machine Learning (cs.LG)
备注: Technical Preprint. 12 pages, 5 figures

点击查看摘要

Abstract:Industrial recommendation faces a resolution–complexity dilemma: dense interactions make detailed modeling of long behavior histories expensive, while compact history summaries restrict the evidence available to prediction. A unified backbone must preserve rich behavioral details within strict serving budgets. We introduce \textbfTokenFormer V2, a unified architecture that explicitly decouples \emphrepresentation resolution from \emphinteraction density. Its group-wise token layout jointly encodes behavior segments with non-sequential context, expands interaction ranges via hierarchical group merging, and coordinates distributed groups through \emphNS Fusion and \emphGlobal Sync modules. Through this \emphlocal–global decomposition paradigm, computation is concentrated on structured interactions while preserving fine-grained evidence, ensuring that every encoded event remains directly accessible to the \emphFull-Resolution Readout module. On the Tencent Advertising Algorithm Competition 2026 (TAAC 2026) benchmark, TokenFormer V2 achieves top-ranking accuracy among unified recommendation baselines while using \textbf5.1 \times fewer FLOPs than leading baselines, sustaining its advantage across long histories and imbalanced behavior domains. Furthermore, in billion-parameter evaluations on WeChat Channels advertising, it consistently outperforms state-of-the-art industrial baselines, yielding up to a \textbf+0.07 pp overall AUC gain. In live online A/B testing against the production model, TokenFormer V2 delivers a statistically significant \textbf2.33% increase in GMV. These results establish TokenFormer V2 as a battle-tested blueprint for scalable, full-resolution behavior modeling in industrial recommendation.

[IR-30] LMEG-Doc: Layout-Grounded Multi-View Evidence Graphs for Multimodal Long-Document Retrieval

链接: https://arxiv.org/abs/2611.18901
作者: Changcheng Fu,Xiao Liu,Qi Yu,Yage Wang,Derek Young,Giri Iyengar
类目: Information Retrieval (cs.IR)
备注: Work in Progress

点击查看摘要

Abstract:Retrieving relevant evidence from multimodal long documents remains challenging, as supporting information is often distributed across different pages and modalities. Existing retrieval methods primarily rank document elements by their individual similarity to a question, often overlooking relationships among complementary evidence. Consequently, relevant figures and tables may receive low retrieval scores even when they contain essential information. To address this limitation, we introduce LMEG-Doc, a layout-grounded multi-view evidence graph framework for multimodal long-document retrieval. We first construct an evidence graph that captures structural and contextual relationships among document elements, and introduce Cross-Page Relation Refinement (CRR) to discover additional connections across distant pages through VLM-based verification. Building on this graph, we combine multimodal retrieval signals with graph-guided relevance propagation to recover connected supporting evidence at both layout and page levels. Experiments on MMDocIR, DocScope, and FinRAGBench-V demonstrate consistent retrieval improvements, including outperforming retrievers fine-tuned on MMDocIR without task-specific training. Furthermore, augmenting existing question-answering pipelines with our retrieved evidence improves answer accuracy by 7.6% on DocScope and 15.8% on FinRAGBench-V, demonstrating the effectiveness of relationship-aware multimodal retrieval.

[IR-31] Label-Free Detection of Structural Bias in LLM -Based Recommender Systems via Randomized Paraphrasing: The Case of Verbosity

链接: https://arxiv.org/abs/2611.18878
作者: Carlos Peláez-González,David Herrera-Poyatos,Francisco Herrera,Andrés Herrera-Poyatos
类目: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI)
备注: 24 pages, 5 figures

点击查看摘要

Abstract:Large language models (LLMs) are increasingly deployed as recommenders in e-commerce and conversational shopping systems, where they select a product from a set of retrieved candidates described in natural language. Their decisions should depend on relevance to the user query, but may also be influenced by surface-level properties of the descriptions, such as their length. Detecting such structural biases is difficult because relevance labels are rarely available. We propose a label-free framework to audit LLM-based recommender systems for structural bias. Candidate descriptions are rewritten by randomly assigned paraphrasing models and presented in random order. We prove that, if the recommender’s choice is invariant to these rewritings, every paraphraser is selected with equal probability, yielding a statistical test that requires no relevance labels. A four-step statistical pipeline tests this property and quantifies the contribution of each description feature while controlling for query-product relevance. We apply the framework to verbosity bias on 20,000 product queries and nine recommender LLMs from five model families. After controlling for the measured relevance signals, description length has a positive effect on selection in all nine models, increasing the odds of selection by up to 38% per standard deviation of length. Relevance remains the primary driver of recommendations, while verbosity acts as a consistent secondary signal whose strength depends on the model family. Since the framework needs only black-box access and no annotations, it can be run as a routine audit of deployed shopping assistants, and we provide an auditing protocol with cost and sample-size guidance.

[IR-32] Intrinsic Attention as Navigator: A Non-Generative Retrieval Method for Multi-Hop Queries EMNLP2026

链接: https://arxiv.org/abs/2611.18617
作者: Linwei Song,Weijie Zhang,WeiDi Chen,Huixia Lai,Shi Zhang,Ruliang Xiao
类目: Computation and Language (cs.CL); Information Retrieval (cs.IR)
备注: Accepted to Findings of EMNLP 2026

点击查看摘要

Abstract:For multi-hop queries in retrieval-augmented generation (RAG), it is crucial to accurately obtain evidence pieces. In this study, we propose a training-free method for multi-hop query retrieval, named Intrinsic Navigating-based Non-Generative Retrieval (IN-Retriever). Instead of generating intermediate reasoning text, the method extracts the bridging clue from the internal attention of the embedding model based on the current context and retrieved candidate documents. During iterative clue extraction and document retrieval, IN-Retriever strips away revealed clues from the embedded query vector and then reranks candidate documents in a targeted manner, thereby steering retrieval toward hidden next-hop evidence. In the experiments, we confirm the feasibility and effectiveness of using intrinsic attention signals to guide external multi-hop retrieval. Experimental results show that our method precisely collects sufficient documents to support multi-hop queries and significantly outperforms sparse and dense retrieval methods.

[IR-33] Automating literature screening and knowledge synthesis with large language models

链接: https://arxiv.org/abs/2611.18588
作者: Zhanzhao Li,Qiyao He,Weiqiang Chen,Kengran Yang,Kai Gong
类目: Computation and Language (cs.CL); Materials Science (cond-mat.mtrl-sci); Information Retrieval (cs.IR)
备注: 15 pages, 6 figures, 1 table

点击查看摘要

Abstract:Scientific literature is expanding rapidly, yet much of its knowledge remains difficult to organize and reuse due to the substantial workload needed to navigate large, heterogeneous literature corpora. Here, we develop a generalizable large language model (LLM)-powered framework that automates literature screening, structured information extraction, and knowledge synthesis within a unified workflow. We benchmarked 11 state-of-the-art proprietary and open-weight LLMs across three materials science screening tasks differing in scope, corpus size, and class imbalance. Seven LLMs achieved balanced accuracies above 0.90 across all tasks, with the best-performing LLM reaching 0.96–0.98, demonstrating the potential of LLMs to accurately screen large, highly imbalanced scientific corpora. The framework also substantially reduces manual effort, screening 10,000 records within four minutes compared with approximately 83 person-hours for manual screening. Beyond literature screening, the framework extracts structured information from full-text articles, enabling content-level quantification of research landscapes that extends traditional bibliometric analysis based on publication metadata. This capability reveals methodological trends and underexplored directions, as shown for machine learning in concrete materials science. We further demonstrate its utility for information and data discovery by identifying publications containing silicate glass density data that are difficult to retrieve using conventional search strategies, which can support downstream hypothesis development, research design, and data-driven scientific discovery. Overall, these results demonstrate the broad applicability of LLM-powered literature analysis across scientific domains and its potential to transform large, unstructured literature corpora into structured, analyzable knowledge at scale.

[IR-34] Leverag ing Binary Hashing for Scalable Index Assignment in Semantic Communications

链接: https://arxiv.org/abs/2611.21260
作者: Leonardo Roque,Kuangyun Zhu,Louis-Adrien Dufrène,Quentin Lampin,Grégoire Lefebvre,Mohamad Assaad
类目: ignal Processing (eess.SP); Information Retrieval (cs.IR)
备注: 6 pages

点击查看摘要

Abstract:Transmitting semantically structured embedding vectors over digital channels is costly in bit rate, while conven- tional vector quantization approaches face limitations in scalabil- ity, encoding complexity, and index assignment. This paper inves- tigates binary hashing techniques from the information retrieval literature as semantic indexing strategies for general-purpose networks. We study random Locality-Sensitive Hashing (LSH) along with Principal Component Analysis (PCA)-, Canonical Correlation Analysis (CCA)-, and ITQ (Iterative Quantization)- based variants in a semantic communication pipeline over a Binary Symmetric Channel. Embeddings are compressed into binary codes, reconstructed through Hamming retrieval from a shared table, and evaluated on downstream classification. Ex- periments on the Massive Text Embedding Benchmark (MTEB) and the Banking77 dataset show that data-dependent hashing improves upon random LSH, while ITQ refinement increases robustness to channel noise. Among the evaluated methods, CCA- ITQ achieves the best classification accuracy across the tested channel conditions. These results indicate that binary hashing is an efficient and scalable alternative to quantization-based embedding compression for semantic communications.

人机交互

[HC-0] Benefits and Barriers to Human-AI Teaming for Crisis Counseling

链接: https://arxiv.org/abs/2611.21558
作者: Ivy Pham,Akshay Swaminathan,Sebastian Garcia,Jaimie Lim,Gloria Ye,Divyanjali Verma,Alan Pham,Chastin Chung,Shreya Shah,Shaked Azzam,Sharang Phadke,Ivan Lopez,William Wang,Angelina Wang,Nigam H. Shah
类目: Human-Computer Interaction (cs.HC); Computers and Society (cs.CY)
备注:

点击查看摘要

Abstract:Background: AI systems are increasingly deployed to support mental health clinicians, yet real-world evidence remains limited. In crisis counseling, where clients often seek human connection, understanding whether AI augmentation preserves or undermines meaningful human engagement is essential. Methods: In a mixed-methods study of a counselor-facing AI copilot on a WhatsApp-based crisis hotline in India (Mar.-Oct. 2025), counselors could request, edit, and send AI messages. We analyzed 20 interviews, 8 shadowing sessions, 56 survey responses, and edits in 27,778 copilot messages from 108 counselors via thematic analysis (Technology Acceptance Model 2, plus inductive coding of emergent factors), and survey statistics. Results: Counselors reported high copilot acceptance (survey): 98.2% found it helpful; 89.3% likely to continue use; 80.4% reported efficiency gains. Interviews highlighted sentence scaffolding, empathetic phrasing, stress reduction, catching missed details, and efficiency. Barriers (survey) included repetitive outputs (23.2%), lack of context (17.9%), and authenticity concerns (10.7%). While 69.6% (survey) reported skill enhancement, interviews surfaced skill atrophy concerns. Counselors frequently accepted copilot suggestions, editing to remove questions, adjust tone, or personalize style. Conclusion: In one of the largest real-world deployments of a counselor-facing mental health AI copilot, counselors reported reduced cognitive load, empathetic communication, and greater efficiency, alongside concerns about skill atrophy, authenticity, and over-reliance. In settings like crisis counseling where clients seek human connection, AI tools must balance operational support with preserving clinical skill and genuine therapeutic engagement. Designs soliciting approval or editing of complete drafts may overweight efficiency at the expense of sustained human engagement. Subjects: Human-Computer Interaction (cs.HC); Computers and Society (cs.CY) Cite as: arXiv:2611.21558 [cs.HC] (or arXiv:2611.21558v1 [cs.HC] for this version) https://doi.org/10.48550/arXiv.2611.21558 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Ivy Pham [view email] [v1] Mon, 16 Nov 2026 18:38:15 UTC (963 KB)

[HC-1] Shaping AIs Beliefs Before It Generates: Human Sense of Agency in Creative AI NEURIPS2026

链接: https://arxiv.org/abs/2611.21405
作者: Zi Wang,Mahima Pushkarna
类目: Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
备注: The Proactive Co-Creator (PaCC) can be accessed at this https URL . Accepted at Neural Information Processing Systems (NeurIPS 2026) Creative AI Track

点击查看摘要

Abstract:AI systems for creative media generation expand human capabilities yet risk diminishing human agency. A core problem is anchoring, where the user’s thinking becomes anchored to a polished, generated artifact, shifting the user’s role from active authoring to reactive editing. We propose pre-generation belief controls that can empower creators to inspect, revise, and steer the AI’s interpretation before any media is generated, so that the initial anchor is set by the human. We formalize four design principles and instantiate them in the Proactive Co-Creator (PaCC), a system for early-stage conceptual ideation and prototyping that proactively surfaces its assumptions and uncertainty through clarifications, belief graphs, and media attributes, inviting users to shape the AI’s beliefs before generating the final outcome. Using PaCC as a design probe in a small-scale, single-arm exploratory user study, we observed that participants reported directionally positive sense of agency and process control relative to their familiar prompt-based AI tools, noting that surfaced assumptions helped them think through details they had overlooked, whereas their feelings about ownership and authorship diverged

[HC-2] Beyond Bilingual: Studying Quadrilingual Code-Switching with Conversational AI

链接: https://arxiv.org/abs/2611.21394
作者: Annika Bush,Melika Minnaar,Salys Sultan,Tasnimul Hasan Tashin,Ankur Bhatt,Meltem Aksoy
类目: Human-Computer Interaction (cs.HC); Computation and Language (cs.CL)
备注:

点击查看摘要

Abstract:Multilingual speakers choose languages by topic and conversation partner. Large language models (LLMs) support these languages unequally. Research on code-switching with conversational AI has mostly examined bilingual pairs of well-resourced languages. In Curaçao and Aruba, many people use Dutch, English, Spanish, and the low-resource Creole Papiamentu/o. In a preregistered study, 42 university students held two five-minute chatbot conversations, one about university life and one about local food and culture. Surveys assessed language background, AI use, and switching comfort. 20 international students in Curaçao formed a comparison sample. Participants switched between turns and rarely mixed languages within a message. English dominated both conversations. Participants switched more often in the cultural conversation, where Papiamentu/o and Spanish rose from 15% to 24% of messages. Self-reports showed no lower comfort switching with the chatbot than with a person. Conversational agents must follow turn-by-turn language changes and handle local low-resource languages well.

[HC-3] Whats the Difference? Voice Design and Confirmation Bias in AI-Assisted Self-Diagnosis: Anthropomorphic vs. Machine-Like Voice Output

链接: https://arxiv.org/abs/2611.21379
作者: Niklas Roethlingshoefer,Emma Winkler,Melanie Loeffler,Emely Rosbach
类目: Human-Computer Interaction (cs.HC)
备注: 11 pages, 6 figures, 3 tables. Student research project, TH Ingolstadt, Summer Semester 2026

点击查看摘要

Abstract:As symptom checkers based on artificial intelligence (AI) are becoming common entry points for health information, their influence on lay users’ diagnostic decisions raises questions about trust and bias. Of particular concern is confirmation bias, the tendency to favour information that supports one’s existing belief while discounting information that contradicts it. One of the central design choices for such systems is the style of voice: whether they should present themselves with a human-like (anthropomorphic) or a clearly machine-like (xenomorphic) voice. This study investigates whether the voice design of an AI symptom checker modulates confirmation bias in self-diagnosis. In a within-subjects online experiment (N = 81 participants, 648 trials), participants selected a diagnosis and rated their confidence, then heard an AI-generated recommendation that either confirmed or contradicted their initial choice, delivered by an anthropomorphic or a machine-like voice, before deciding again. We observed a strong, direction-dependent influence of the AI: participants revised their diagnosis in 45.7% of trials when the AI contradicted them versus 1.2% when it confirmed them (generalised linear mixed model, GLMM, p .001); confidence rose under confirmation but barely moved under contradiction (robust linear mixed model, p .001); and decisions took longer after contradiction (Gamma GLMM, p .001). This effect did not differ between the two voices (no voice or interaction term was significant, smallest p = .282), even though a manipulation check confirmed the voices were perceived as markedly different. The findings suggest that the direction of an AI recommendation, rather than the anthropomorphism of its voice, drives confirmation-related bias in AI-assisted self-diagnosis.

[HC-4] A Crowd-Audited Atlas of Implicit Bias in LLM -Powered Agents

链接: https://arxiv.org/abs/2611.21376
作者: Jason Alexander,Yuxuan Li,Michael Licamele,Wesley Hanwen Deng,Sauvik Das
类目: Computers and Society (cs.CY); Computation and Language (cs.CL); Human-Computer Interaction (cs.HC)
备注: Preprint. 32 pages, 5 figures

点击查看摘要

Abstract:Prior work has shown that large language models (LLMs) express implicit biases in high-publicity, one-off decision scenarios such as hiring and intelligence assessment. However, as people increasingly employ autonomous, LLM-powered AI agents that make decisions on their behalf in doing day-to-day tasks, it is of critical importance that we understand how implicit biases, based on sociodemographic assumptions, emerge across a broad milieu of workaday decision scenarios. We used the 17 domains of daily activity in the American Time Use Survey (ATUS) to generate an atlas of 506 decision scenarios in which LLMs — endowed with personas representing different races, gender identities, and political ideologies — expressed significant implicit bias. Next, in a crowdsourced audit, 500 participants made 8,010 binary choices about which implicitly biased decision scenarios most deserved further review. We found that participants prioritized decision scenarios that had broad relevance (e.g., about food choices), were high stakes (e.g., about health), and where all members of a simulated sociodemographic group made unanimous decisions (because it was a clear indication of stereotyping). They were less swayed by the magnitude of an observed disparity or the specific ATUS domain the decision scenario represented. When examining the rationale behind their prioritization choices, we found that participants often interpreted disparities as traits of demographic groups and judged whether those traits seemed plausible, often based on their own internalized stereotypes. We discuss implications for building personalized AI agents, digital twins, social simulations, and for using crowdsourced bias audits.

[HC-5] Share your Screen and Talk to Gemini : How Young Learners Interact with Multimodal GenAI While Solving Scratch Tasks

链接: https://arxiv.org/abs/2611.21306
作者: Sven Jacobs,Lukas Berning,Natalie Kiesler
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
备注: accepted for: The 21th WiPSCE Conference on Primary and Secondary Computing Education Research (WiPSCE 27)

点击查看摘要

Abstract:Children using Scratch to learn programming can struggle with text-heavy, lengthy explanations and hints during debugging. Verbal interactions based on screen sharing or camera data may be an interesting alternative for supporting learners. We therefore explore the quality, use, and perceptions of fifth- and sixth-graders’ verbal interactions with a GenAI tutor. The goal is to investigate the quality and use of the multimodal GenAI tutor in the given context. First of all, a pilot study of eight tasks conducted without students helped inform the process of the tutor’s classroom evaluation. The classroom study was realized after fifth- and sixth-graders had participated in six Scratch classroom lessons. Then 23 students were introduced to the tutor, worked on the eight tasks with the GenAI tutor, and completed a questionnaire. The analysis of the pilot and classroom study shows that the tutor’s responses were intelligible and actionable, and no single turn revealed a complete solution. However, the tutor missed existing errors in 61.5% of 603 error-related classroom turns, compared to 15.7% in the pilot study. Children’s voice prompts were brief and mostly responded to the tutor. Although a video had explained that responses could be incorrect, 39.1% of the children trusted the tutor’s answers to always be correct. These findings have implications for educators and learners using multimodal GenAI tutors, as they need to be aware that an intelligible and human-like voice does not ensure correct support. For developers, our results suggest that screen sharing alone does not ensure responses that match the program state.

[HC-6] Edit Recommendation: Can LLM s Anticipate Edit Requests from Users and When?

链接: https://arxiv.org/abs/2611.21155
作者: Daisuke Kikuta
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
备注: Blog: this https URL , Code: this https URL

点击查看摘要

Abstract:After a large language model (LLM) agent generates an artifact, users often send edit requests for it. This paper studies whether an LLM can anticipate these requests before they are sent. Such anticipation is useful in practice because it reduces the effort of user review and feedback. We formalize this task as edit recommendation and build EditRecBench, a benchmark for evaluating the performance of LLMs on this task, from real-world conversation logs in WildChat-4.8M. Through benchmarking and failure analysis, we identify a practical challenge that some cases are inherently difficult to predict in real-world settings. To address this, we introduce decision model-based gating to filter out cases where predictions are unlikely to be correct before they are presented to the user. Our trained model outperforms existing decision models and raises pass@5 among shown suggestions by 12–25 points at a gating threshold. Our code is available at this https URL.

[HC-7] KuroshioVis: Hypothesis-Driven Discovery of Mesoscale Air-Sea Interaction through Coarse-to-Fine Exploration

链接: https://arxiv.org/abs/2611.21154
作者: Yuanzhi Chen,Yiwei Lin,Xi Duan,Qingqing He,Ruiliang Xu,Yucheng Wang,Qiong Zeng
类目: Human-Computer Interaction (cs.HC); Graphics (cs.GR)
备注:

点击查看摘要

Abstract:Mesoscale air-sea interaction occurs throughout strong sea surface temperature fronts such as the the Kuroshio Extension, but its strength is highly non-uniform and can be resolved only in kilometre-scale coupled simulations. The response of any single eddy is weak relative to synoptic variability, so evidence supporting a hypothesis of systematic air-sea coupling must be accumulated across many eddies rather than inferred from individual cases. We present KuroshioVis, a hypothesis-driven, coarse-to-fine visual analytics workflow. The coarse stage extracts and aligns mesoscale eddies into composite representations of their air-sea responses, and provides a global view for exploration. The fine stage investigates these responses through three coordinated views: the composite in terface view, the cross-interface flow view, and the evidence matrix view. These coordinated views support air-sea coupling analysis from surface patterns to cross-interface dynamics and quantitative statistical evidence. We demonstrate the workflow using the coupled C1440-LLC2160 simulation of the 2026 SciVis Contest.

[HC-8] Human-AI Co-Ideation Across Socio-Technical Configurations: An Interactional Analysis of Planar and Immersive Workspaces

链接: https://arxiv.org/abs/2611.21125
作者: Orit Shaer,Mary Calabro,Zoe Mitchell,Angelora Cooper,Raffaella Sadun,Andrew L. Kun,Osnat Mokryn
类目: Human-Computer Interaction (cs.HC); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Remote collaborative platforms increasingly integrate generative AI to support group ideation. These platforms are sociotechnical configurations in which spatial mediation, interaction modality, AI output format, and the mechanisms that make contributions visible to a group co-vary as properties of a deployed system. This paper examines how such configurations shape participants’ collaboration and their engagement with AI during creative work. We report a comparative study of two commercially available platforms that instantiate contrasting configurations: a planar shared canvas with text-based AI integration, and an immersive virtual reality workspace with concise AI output embedded in the shared space. Fourteen all-women triads completed a structured ideation task consisting of human-only idea generation, AI-supported idea expansion, and group convergence. We analyzed collaboration patterns, selection of AI-generated ideas, perceived idea quality, workload, presence, and their lexical characteristics. The immersive configuration yielded greater spatial presence and introduced a participation paradox, in which spatial freedom enabled exploration and fragmented group coherence; it was also associated with more balanced role distribution during convergence, at the cost of higher physical demand. In both configurations, AI multiplied the idea pool, and ideas groups authored with AI available were rated above participants’ own. AI output format: about 21 words per idea on the planar canvas and 3 in the immersive workspace. Tracked how participants evaluated AI-generated ideas and how often they selected them unaltered as finalists. Neither configuration made AI contributions visible or attributable to the group: participants invoked AI individually and asynchronously in both. We derive design implications for AI output format, group-visible AI invocation, and spatial support for convergence

[HC-9] SWE-WhatIf: Simulating Real Users of Coding Agents from Recorded Session

链接: https://arxiv.org/abs/2611.21095
作者: Tae Soo Kim,Tongshuang Wu,Juho Kim
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Recorded sessions between users and coding agents offer a way to evaluate how agents complete real-world tasks and interact with users. However, these sessions cannot simply be replayed with new agents, because recorded user replies may no longer apply when an agent takes a different trajectory. We introduce SWE-WhatIf, which replays recorded sessions with new agents by modeling the user’s intents, the criteria for satisfying them, and the conditions under which they arise. Using this model, a user simulator generates replies that adapt to unseen agent turns while remaining grounded in the recording. Across 89 real-world sessions, SWE-WhatIf generated the replies most similar to those of real users and matched the strongest baseline when sessions were rewritten. To test adaptation, we replay sessions with agents instructed to avoid or repeat actions that the recorded user pushed back on. SWE-WhatIf produced the highest win rate for avoiding these actions (66.3%), significantly above chance. We further demonstrate its utility by ranking five LLMs as coding agents, where the highest-ranked model depended on how users specified and developed their intents, and a study with 16 users, where replaying their own sessions changed 14 participants’ rankings and informed their future agent use.

[HC-10] Semantic Validity vs. Graphical Arrangement in Visualization Grammars IEEE-VIS2026

链接: https://arxiv.org/abs/2611.21073
作者: Cynthia A. Huang,Matthew Kay
类目: Human-Computer Interaction (cs.HC)
备注: 5 pages (4 + references). Accepted at the Grand Unified Grammar of Graphics (GUGOG) workshop at IEEE VIS 2026

点击查看摘要

Abstract:Systems for modular, declarative and compositional specification for data visualizations, otherwise known as visualization grammars, often claim Wilkinson’s Grammar of Graphics as motivation, inspiration or foundation. However, despite the shared intellectual lineage, visualization grammars can vary greatly in design scope and focus. In particular, systems that strongly adhere to domain-specific notions of correctness must trade-off expressivity for validity. This trade-off between semantics and graphical flexibility is currently under-articulated, leading to difficulty integrating various GoG-inspired visualization contributions, especially between InfoVis and Statistical Graphics. We propose three new types of visualization contributions to support such integration: Semantic Validity Theories, Semantic Visualization Systems and Graphical Arrangement Systems. The first two prioritize the adherence of data visualizations to correctness conditions, such as statistical or mathematical transformations and data semantics, whilst the third largely leave validity concerns to the user. We illustrate these new contribution types using examples from two recent grammar of graphics inspired papers, and show how they supported more nuanced notions of expressivity and usability in evaluation and design of visualization systems.

[HC-11] owards wearable-scale transition-aware sleep staging for clinical populations

链接: https://arxiv.org/abs/2611.21040
作者: Suvadeep Maiti,Keya Sinha,Raju S. Bapi
类目: Human-Computer Interaction (cs.HC); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Automatic sleep staging models are predominantly developed on healthy sleepers and generalize poorly to clinical populations with fragmented sleep. Across three cohorts (patients with stroke, isolated REM sleep behaviour disorder, and other sleep disorders), we observe 2-3x more awakenings per hour, nearly twice as much N1, and less REM sleep than in healthy sleepers. We propose a label-efficient teacher-student framework for wearable-scale sleep staging. SleepGPT (136M parameters), adapted using labels from 25% of training patients, supervises a causal CNN-GRU (31k parameters; 122 KB) that stages EEG, EOG, or both in real time. We introduce Gcv-TAD, a new black-box transition-aware distillation loss. It emphasizes uncommon stage transitions while discounting those that vary across patients, using a prior estimated from labelled patients. In patient-wise 10-fold cross-validation, Gcv-TAD improves per-patient macro F1 over supervised training on the same labels by +3.2 to +7.6 points on average. Gains are largest on N1, and performance approaches full-label training.

[HC-12] Development of an HPLC Analytical Method Based on Large Language Models :A Case Study of the Eczema External Wash Formula

链接: https://arxiv.org/abs/2611.20994
作者: Kexin Guo,Xueting Sun,Xinyan Zhao,Wei Wang,Xingchu Gong
类目: Human-Computer Interaction (cs.HC)
备注: 22 pages,8 tables,7 figures

点击查看摘要

Abstract:Traditional Chinese medicine (TCM) formulas feature complex chemical compositions, and their HPLC quality control method development is time-consuming and experience-dependent. Large language models (LLMs) provide a promising strategy for intelligently assisting analytical method construction. Taking the Eczema External Wash Formula as the research object, this work constructed a human-machine collaborative workflow involving intelligent recommendation, experimental validation and iterative optimization for HPLC method development. Among five prompt strategies tested, expert-role setting with structured reasoning performed best. After inputting formula information, the LLM proposed five chromatographic schemes. Through experimental screening and optimization, the final HPLC method used a SunShell C18 column, mobile phase of 0.1% aqueous formic acid-acetonitrile for gradient elution, with seven markers separated within 55 min. Twenty components were characterized by LC-MS. The validated method exhibited satisfactory linearity, precision, repeatability, stability and recovery. This indicates LLMs can reduce the workload of chromatographic condition exploration and boost the development efficiency of TCM formula analytical methods, supporting intelligent TCM quality control.

[HC-13] Advancing Cognitive Pairwise Comparisons for Human-Centric Artificial Intelligence and Data Science: A Human-in-the-Loop Framework

链接: https://arxiv.org/abs/2611.20926
作者: Kevin Kam Fung Yuen
类目: Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC); Optimization and Control (math.OC)
备注: 33 pages, 19 tables, 8 figures; interactive web demos available at this https URL ; source code and datasets at this https URL ; archived at this https URL ; earlier preprint version available at SSRN 4048692 ( this http URL or this https URL )

点击查看摘要

Abstract:Human-centric artificial intelligence and data science (HCAD) systems increasingly involve human-in-the-loop (HITL) interventions to overcome the limitations of fully automated machine learning algorithms, such as automation bias and the inability to incorporate tacit domain knowledge, and continuously improve predictive performance. Data transformations, feature selection, feature weighting, and model selection usually require laborious human involvement across iterations of artificial intelligence and data science (AIDS) pipelines, making development a challenging and time-consuming process. The Cognitive Pairwise Comparison (CPC) framework is proposed as a human-centric methodology to elicit preferences and priorities in a structured format. To establish essential facets of human-centricity, this research advances the CPC methodology so that user preference elicitation can be systematically quantified in these challenging AIDS tasks. To demonstrate the usability and validity of these facets, this study presents several applications within movie recommender systems using the MovieLens dataset. It exemplifies how the extended CPC framework can facilitate interactions between human judgments and data across supervised learning, dimensionality reduction, and unsupervised learning algorithms.

[HC-14] Lived Experiences of Power and Agency: Achieving Voice Sovereignty in Assistive Speech Technology for Nonbinary Users INTERSPEECH2026

链接: https://arxiv.org/abs/2611.20858
作者: Maxwell Hope,Juliana Francis,Joakim Gustafson,Éva Székely
类目: Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
备注: Proceedings of Interspeech 2026

点击查看摘要

Abstract:Speech-generating devices (SGDs) are essential tools for many speech-disabled individuals, yet current systems often reinforce normative assumptions about identity and communication. Prior research on gender expansive SGD users has largely focused on gendered voice options, with less attention to broader identity alignment and the technological conditions that shape users’ agency. This study examines how gender expansive SGD users experience power over their voices using thematic analysis of qualitative survey data. Three themes emerged: (1) integration of SGDs into daily life, (2) identity expression, and (3) power and agency. Participants expressed the need for more responsive and reliable technology, inclusive and customizable voices, and greater autonomy in defining how their voices sound and are perceived. Ultimately, their perspectives call for a reimagining of SGD systems as tools for reclaiming the power of one’s voice and the ability to define and enact one’s will through it.

[HC-15] EEGProbe: Do Vision-Language Models Really Understand EEG? A Benchmark for Signal-Grounded EEG Signals Evaluation

链接: https://arxiv.org/abs/2611.20779
作者: Haozhe Zhu,Dan Yuan,Chenghao Gu,Feike Xu,Yue Wen,Shin’ichi Konomi
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Human-Computer Interaction (cs.HC)
备注: 21 pages, 12 figures

点击查看摘要

Abstract:Vision-language models (VLMs) have shown strong multimodal capabilities across scientific and biomedical domains, yet their ability to understand electroencephalography (EEG) signals remains unclear. We introduce EEGProbe, a signal-grounded benchmark for evaluating whether general-purpose VLMs can extract and reason from minimally processed, channel-wise EEG without task-specific feature engineering or EEG-specific adaptation. EEGProbe contains 620 EEG-centered questions across 13 subtasks and three capability categories: EEG Signal Classification, EEG Cross-modal Alignment, and EEG Neurophysiological Reasoning. To reduce shortcut-based answering, EEGProbe preserves the temporal structure of the recordings where possible, minimizes benchmark-induced processing, and applies signal-necessity testing followed by expert verification. We evaluate 11 representative open- and closed-source VLMs under a unified zero-shot protocol. The strongest model, GPT-5.6 Sol, achieves an overall score of only 41.63%, with cross-modal EEG alignment emerging as a major bottleneck. We further examine model behavior from a neurophysiological perspective and analyze how temporal, spectral, spatial, and physiological EEG structure relates to model successes and failures. Inference-time neurophysiological guidance can reshape model predictions by helping models recognize the neurophysiological patterns present in the input and connect them with relevant neurological knowledge, narrowing the gap between general multimodal competence and EEG understanding. EEGProbe provides a testbed for current VLMs and a basis for developing future multimodal models that can interpret EEG evidence more reliably. The EEGProbe benchmark will be publicly available at this https URL.

[HC-16] UIAnchor: Anchoring UI Perception and Action Execution for Reliable Service-Composed Mobile Task Automation with GUI Agents

链接: https://arxiv.org/abs/2611.20753
作者: Wentao Zhou,Sicong Liu,Zimu Zhou,Yimeng Duan,Yongyan Cai,Weiye Wu,Teng Li,Daqing Zhang,Zhiwen Yu
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Mobile task automation aims to streamline multi-step, cross-app interactions on smartphones and in-vehicle systems. Recent LLM-based GUI agents have advanced rapidly, yet reliability remains limited because modern mobile workflows are increasingly service-composed (app switching, forms, pop-ups/permissions, copy-paste), requiring fine-grained operations on dense, dynamic, and heterogeneous UIs, often in distraction-sensitive contexts (e.g., hands-busy or attention-limited use). Through an in-the-wild failure analysis of deployable GUI agents, we identify two dominant bottlenecks: agents often miss or misread actionable UI elements, and they execute actions without verifying target correctness or outcomes. We present UIAnchor, a modular multi-agent system that improves mobile GUI automation by anchoring both perception and execution. UIAnchor combines a two-stage UI parser for high-recall element capture and context-aware semantics with a meta-controller that performs pre-action verification, post-action outcome perception, per-step state tracking, and targeted recovery. We further introduce an L1-L5 task taxonomy based on step length, cross-app scope, and UI granularity, showing that L5 remains beyond today’s practical frontier. On the hardest practical tier, L4 tasks (20-30 steps, multi-app, targets 100 x 100 px, ~5 mm on typical phones), UIAnchor improves success by 31.5% over GPT-4o and 16.6% over Mobile-Agent-v3. With edge/cloud assistance, UIAnchor runs at ~2 s per step, comparable to human operation, while reducing per-step latency by up to 75.5% and energy by 52.4%. It also generalizes to unseen apps and cross-platform GUIs.

[HC-17] Injectra: Retrofitting Extensibility into Web-Based Desktop Applications

链接: https://arxiv.org/abs/2611.20713
作者: Qin Sun
类目: oftware Engineering (cs.SE); Human-Computer Interaction (cs.HC)
备注: 31 pages, 4 figures, 4 tables. Source code: this https URL

点击查看摘要

Abstract:Large language models can generate code based on natural language prompts, opening up new possibilities for users to customize software and extend its functionality. This paper proposes and implements Injectra, a general-purpose code injection and feature enhancement system for existing desktop clients, suitable for applications whose user interfaces are built using embedded browser engines (such as Electron) or system WebViews (such as Tauri). Injectra decouples the host integration logic from the upper-layer enhancement code. Combining both runtime integration and pre-launch bootstrap mechanisms, it injects JavaScript, HTML, and CSS into the application’s internal Web runtime environment through a unified execution channel, thereby enabling modifications to interface rendering, adjustments to interactive behaviors, and client-side functionality extensions. The system uses a command-line interface (CLI) as a unified entry point, allowing external tools, AI agents, and MCP integration layers to reuse its injection and enhancement capabilities. We tested 38 desktop applications already installed on our macOS machine; code injection was successfully completed for all of them. We demonstrated the system’s capability to extend functionality through various enhanced use cases and also successfully validated the Windows version. Based on Injectra, we further developed an LLM-powered enhancement tool driven by selection and annotation: users select interface objects within the original application and describe their modification requirements; the large language model then generates the enhancement code, which is injected, executed, and rendered by Injectra. The system’s source code has now been released as open source: this https URL.

[HC-18] AI Tool Adoption and Loyalty among Product Industrial and UX/UI Designers across the Double Diamond

链接: https://arxiv.org/abs/2611.20519
作者: Sepideh Tajarmakan,Khashayar Hojjati Emami
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Designers share many AI tools, yet their roles produce different work. This survey compared 337 Product, Industrial and UX/UI Designers from 40 countries and regions across the four phases of the Double Diamond. It measured AI adoption, breadth and frequency, tool function and the number of tools named per phase. A tool was loyal when a role picked it in all four phases. Industrial Designers adopted AI less often (73.9% against 88.2% and 87.9%), a gap that went with physical projects once project type and country were held constant. Of 60 AI tools, 25 were common to all three roles and carried at least 90% of each role’s picks. Tool function differed by role, most of all in Develop. A few loyal tools carried about half of the picks, led by ChatGPT. Tool counts fell across the phases in every role. A hurdle model found a Phase x Role interaction for using AI in a phase (p = 0.041), which was not shown for the number of tools (p = 0.080). Predicted adoption among UX/UI Designers fell from 87.2% in Define to 43.9% in Develop. Support for designers needs one shared core, role-specific tools and most attention in Develop.

[HC-19] EVIDENT: Evaluating Evidence Grounding in GUI Agents EACL2027

链接: https://arxiv.org/abs/2611.20488
作者: Yerin Nam,Minkyung Cho,MinKyu Kim,SeungWoo Song,Hangyeol Yoo,Jihyo Kim,KyungTae Lim
类目: Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
备注: Accepted to EACL 2027 Findings. 36 pages

点击查看摘要

Abstract:GUI agents are increasingly evaluated by whether they complete web tasks, but surface success can hide a critical failure: the final answer may not be supported by evidence observed during navigation. We introduce EVIDENT, an evidence-driven benchmark for evaluating whether GUI agents ground their decisions in observable web evidence. EVIDENT contains 220 real-world English and Korean tasks spanning semantic grounding, contextual inference, and visual grounding, each paired with human-curated evidence metadata specifying what must be verified. We further separate surface task completion from evidence-grounded success using three metrics: TSR, ESR, and GSR, which respectively measure outcome success, required evidence observation, and claim-level evidence support. Experiments on commercial and open-source GUI agents reveal substantial gaps between surface success and evidence-grounded success, especially in contextual and visual tasks. These results show that current evaluations can overestimate agent reliability and highlight the need to assess not only whether agents answer, but whether they verify.

[HC-20] PromptPoseFX: Authoring Pose-Driven Visual Effects through Augmented Prompting and Interactive Refinement

链接: https://arxiv.org/abs/2611.20485
作者: Kexin Zheng,Hui Ye,Hongbo Fu
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Human pose-driven visual effects offer a compelling way for enhancing video expressiveness and storytelling. Authoring such effects with traditional video editing tools is a complex and time-consuming task that requires considerable technical expertise and manual effort. While generative AI models can simplify the workflow by reducing manual scripting through language-driven generation, existing approaches suffer from imprecise alignment between visual effects and pose data, failure to preserve original video content, and limited support for fine-grained, iterative refinement. To address these challenges, we present PromptPoseFX, an interactive system that allows creators to author pose-driven visual effects through augmented language prompts and interactive refinement. Our system leverages Large Language Models (LLMs) to translate high-level user intent into parameterized procedural graphics scripts, automatically mapping pose-tracking data to specific effect parameters, allowing precise alignment with human motion. This script-based approach manages effects as independent layers, thereby preserving the original content of the underlying video. To reduce language ambiguity, our interface offers controls across spatial, temporal, and parameter domains, enabling the precise, intuitive, and iterative authoring of pose-driven visual effects. We evaluated PromptPoseFX through a comparative study against a text-only LLM interface, followed by an exploratory study. In the comparative study, participants rated PromptPoseFX significantly higher than the text-only baseline on measures of perceived control and precision. The exploratory study examined usability and creativity support in open-ended authoring tasks.

[HC-21] Supporting Community Members Navigating Reentry with Revisable Future Envisioning

链接: https://arxiv.org/abs/2611.20375
作者: Modhumonty Das,Nathan Morelli,Aakash Gautam
类目: Human-Computer Interaction (cs.HC); Computers and Society (cs.CY)
备注: 23 pages, 1 figure, 1 table, to appear at the ACM international conference on Supporting Group Work (GROUP’27)

点击查看摘要

Abstract:Future-planning tools often treat agency as individual and leave out the institutional conditions that shape how aspirations become action. We examine this tension in the context of reentry from incarceration and recovery from substance use disorder, through a partnership with a nonprofit organization. Building on participants’ physical vision-board practices, we designed a digital future-envisioning tool that helps users represent current circumstances, desired futures, and pathways that connect goals to concrete resources. We draw on data from three workshops involving 18 participants and follow-up interviews with 10 participants to examine how they used future envisioning within a fragmented support system. Participants described reentry and recovery as ongoing processes in which the everyday work of staying stable was central to progress. Envisioning futures supported this work by surfacing constraints and interdependencies, by reframing personal strengths and institutional supports as assets, and by helping participants move from abstract hopes toward articulated goals and contingent plans. We theorize revision as maintenance work and conclude with design implications for community-embedded interactive systems that support revisable, dependency-aware planning.

[HC-22] “Can a Robot Really Care About Me?” How Perceiving a Tutor as Human or AI Affects Computer Science Students Attitudes and Experiences

链接: https://arxiv.org/abs/2611.20367
作者: Tymon Vu(1),Wenxuan “Wendy” Shi(2),Silas Hsu(3) ((1) California Polytechnic State University, San Luis Obispo, (2) California Polytechnic State University, Pomona)
类目: Human-Computer Interaction (cs.HC)
备注: 30 pages, 3 figures

点击查看摘要

Abstract:Conversational artificial intelligence can reproduce many instructional and social behaviors of human tutors, but students may interpret those behaviors differently depending on who they believe is providing them. We examined how perceived tutor identity shapes students’ evaluations of instructional feedback, satisfaction of psychological needs according to Self-Determination Theory, and their communication strategies with tutors. In a within-subjects Wizard-of-Oz study, 27 introductory computer science students completed two text-based tutoring sessions involving Python code-explanation problems. Participants were told one tutor was human and the other was AI, although the same human provided feedback in both conditions. We found participants evaluated both tutors favorably and reported few consistent differences in the instructional value of the feedback. However, participants frequently questioned whether AI could genuinely care about them, which corresponded with the human tutor having higher scores of relatedness need satisfaction. Human presence offered connection and accountability for some participants but introduced anxiety in others. Participants also reported communicating more casually with the human tutor and more explicitly with the AI tutor. These findings suggest complementary roles for human and AI tutoring and highlight the importance of considering psychological needs alongside instructional quality.

[HC-23] Sidewalk Stewards: A Visual Analytics Approach for Interactive Topology Repair of Sidewalk Networks

链接: https://arxiv.org/abs/2611.20330
作者: Stefan Cobeli,Leonardo Ferreira,Logan Loch,Luiza Melo,Peyman Baghershahi,Claudio D. G. Linhares,Nivan Ferreira,Sourav Medya,Fabio Miranda,Maryam Hosseini
类目: Human-Computer Interaction (cs.HC); Computer Vision and Pattern Recognition (cs.CV); Computers and Society (cs.CY)
备注: Under review

点击查看摘要

Abstract:Accurate pedestrian network data is foundational to urban accessibility, routing, and infrastructure planning, yet such data remains scarce and difficult to produce at scale. Automated extraction pipelines have made significant strides by leveraging computer vision to derive sidewalk networks from aerial imagery, and have seen real-world adoption by municipalities seeking cost-effective inventory solutions. Yet, these pipelines consistently produce topological errors (e.g., disconnected components, dangling edges, missed connections, malformed intersections) that undermine the utility of the extracted networks for routing and accessibility analysis. Correcting these errors requires situated human judgment that global automated repair strategies cannot replicate, yet practitioners currently lack dedicated tools to support this work efficiently. We present Sidewalk Stewards, a visual analytics system that bridges the gap between automated extraction and a deployment-ready pedestrian network. Sidewalk Stewards detects candidate topology errors, supports interactive user-guided correction through coordinated spatial views, and provides real-time quality feedback as edits are applied. We evaluate Sidewalk Stewards through a comparative user study, case studies across cities, and a quantitative evaluation. Compared with conventional workflows, the system facilitates error identification and enables faster and more complete correction. The case studies illustrate the workflow across different urban contexts, and the quantitative evaluation shows that a small number of user corrections can enhance multiple sidewalk quality metrics. Sidewalk Stewards is available at this http URL.

[HC-24] Vibe Interventions: Can Users Personalize and Iterate Their Way to Behavior Change?

链接: https://arxiv.org/abs/2611.20315
作者: Sally Ahmed,Ayse Alomar,Tawab Safi,Pattie Maes,Pat Pataranutaporn
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Recent advances in large language models allow a working application to be generated from natural-language description, and is referred to as generative user interfaces. Their application to behavior change is unexplored, and we provide one of the first controlled investigations of it. We distinguish two components of authorship: iteration, the revision of a generated artifact in natural language, and personalization, the conditioning of generation on stated context. In a pre-registered between-subjects experiment (N = 300, 60 per condition), we crossed these two factors and added a static intervention as a control, yielding five conditions; participants stated a goal and received a generated intervention. The personalized generative condition ranked highest on every self-report outcome, including ownership, satisfaction, and intended adoption. Randomized iteration had no detectable effect, and higher revision counts were negatively associated with evaluative ratings. This work establishes which component of authorship carries motivational value in generative behavior change tools.

[HC-25] Reveal: A Provenance-Aware System for Multimodal Urban Data Exploration

链接: https://arxiv.org/abs/2611.20311
作者: Leonardo Ferreira,Joao Rulff,Maryam Hosseini,Aline Paes,Daniel de Oliveira,Fabio Miranda,Marcos Lage
类目: Human-Computer Interaction (cs.HC); Computers and Society (cs.CY); Multimedia (cs.MM)
备注: Accepted to SIBGRAPI 2026

点击查看摘要

Abstract:Urban sensing initiatives increasingly produce multimodal datasets that combine visual and textual information to capture city characteristics and subjective factors such as safety, comfort, and accessibility. Yet, analyzing such data at scale remains difficult, as manual inspection is unfeasible and automated methods struggle to interpret subjective cues and uncover relationships across diverse data modalities. Moreover, existing workflows rarely make exploration steps explicit, hindering reproducibility. To address these challenges, we introduce Reveal, a provenance-aware visual analytics system for exploring heterogeneous urban data. Reveal enables users to build dataflows by combining text and image searches across coordinated views, chaining queries into pipelines, and applying set operations within an exploration diagram while capturing analytical provenance. We evaluate Reveal through two case studies focused on accessibility and walkability.

[HC-26] From XR Sketches to CAD: Preserving Design Intent with AI

链接: https://arxiv.org/abs/2611.20245
作者: Trudi Di Qi
类目: Human-Computer Interaction (cs.HC); Graphics (cs.GR)
备注: Vision paper (6 pages, 1 figure) submitted to the Workshop on AI-Augmented 3D Sketching in Extended Reality (ACE-XR) at IEEE AIxVR 2027

点击查看摘要

Abstract:Extended reality (XR) allows designers to sketch shapes directly in 3D, but converting these sketches into editable, feature-based CAD requires interpreting their intent. A sketch may contain drawing errors, intended features, and design choices that remain open. Correcting every irregularity risks erasing intended features, while automatic completion may impose choices the designer has not made. We argue that AI should add precision without overriding design intent, considering both the resulting shape and how the CAD model behaves during later edits. To guide this process, we propose a conceptual framework that distinguishes errors from intentions in individual sketch properties and relates them to correction, preservation, and exploration in editable CAD. We also outline five research directions for sketch interpretation, CAD construction, and evaluation, with testable hypotheses for selected questions.

[HC-27] DoF-Query Transformer for Adaptive Fusion of Egocentric Vision and Forearm HD-sEMG in Continuous Hand Kinematics

链接: https://arxiv.org/abs/2611.20196
作者: Jiahang Liu,Wenbo Zhang,Junsong Wang,Zhenzhi Ying,Siqi Hou,Naohiko Sugita,Liming Shu
类目: Computer Vision and Pattern Recognition (cs.CV); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Continuous and accurate estimation of human hand kinematics is important not only for intuitive human-machine interaction, but also for capturing fine-grained human manipulation demonstrations that can provide action supervision for embodied-intelligence datasets and dexterous robot learning. However, individual sensing modalities may be insufficient for reliable estimation under varying operating conditions. We present a finger-unencumbered multimodal framework for continuous regression of 20 hand joint angles from egocentric RGB images and 128-channel forearm high-density surface electromyography (HD-sEMG). The framework combines a Vision Transformer-based visual encoder, a Transformer-based EMG encoder, and a degree of freedom (DoF) query Transformer fusion decoder. To account for the varying informativeness of visual and EMG cues across joint DoFs and input conditions, the decoder assigns a separate learnable query to each DoF, enabling DoF-specific and context-adaptive fusion of the two modalities. The framework was evaluated on synchronized multimodal data from 10 participants. Averaged across participants, the proposed fusion model achieved a mean coefficient of determination ( R^2 ) of 0.88 and a mean absolute error (MAE) of 2.80^\circ , compared with 0.82 and 3.21^\circ for the vision-only model and 0.68 and 5.13^\circ for the EMG-only model. It also outperformed the evaluated representative unimodal baselines and alternative fusion designs. DoF-wise analyses showed larger fusion gains for visually ambiguous joints and viewing conditions, while synthetic visual occlusion demonstrated greater robustness than vision-only estimation. These results highlight the complementary value of egocentric vision and forearm HD-sEMG for capturing fine-grained human hand kinematics, with potential applications in human demonstration acquisition and dexterous robot learning.

[HC-28] A Line is Not Just a Line: Visualizing Harmonious Complex Line Sets

链接: https://arxiv.org/abs/2611.20190
作者: Francesca Samsel,Daniel F. Keefe,Gregory Abram
类目: Human-Computer Interaction (cs.HC); Graphics (cs.GR)
备注: 16 pages, many figures

点击查看摘要

Abstract:Line is a foundational building block in visual design for traditional media and computer graphics. Controlled variation in line quality can be used to define form, create abstractions, distinguish categories, control affect, and more. Yet, when it comes to computer graphics visualizations of science, particularly those built upon 3D computer graphics, too often a line is just a line. This portfolio presents a series of visual design studies and lessons learned that explore the diverse character of line that might be useful for 3D visualizations. The imagery begins to show how more sophisticated use of this powerful visual building block could increase the clarity of data visualizations while also bringing a human, handcrafted aesthetic that reinforces the people and impacts behind scientific climate data.

[HC-29] Cyber-Secure Wearable Ultra-Low-Power Networked Sensors for Soldier Health Monitoring: Insights from the CAPSARII project

链接: https://arxiv.org/abs/2611.20187
作者: Luciano Bozzi,Christian Celidonio,Umberto Nuzzi,Massimo Biagini,Stefano Cherubin,Asbjørn Djupdal,Tor Andre Haugdahl,Andrea Aliverti,Alessandra Angelucci,Giovanni Agosta,Gerardo Pelosi,Paolo Belluco,Samuele Polistina,Gaia Strada,Luigi Sossi Ortega,Alexandra-Ioana Albu,Riccardo Volpi,Luigi Malagò,Michael Schneider,Xabier Eguiluz,Haritz Saiz,María Viqueria
类目: Emerging Technologies (cs.ET); Artificial Intelligence (cs.AI); Hardware Architecture (cs.AR); Cryptography and Security (cs.CR); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
备注: Pre-print version

点击查看摘要

Abstract:The CAPSARII project advances a cyber-secure, ultra-low-power wearable platform and Internet of Battlefield Things (IoBT) framework to support real-time monitoring of military personnel and data-driven decision support. The architecture integrates smart-textile sensing, a body/squad area mesh, edge-deployed AI on secure RISC-V microcontrollers, and cloud/federated analytics for cross-mission learning and comparative studies. Two representative use cases–pilot fatigue and warfighters operating in harsh environments–guide requirements, validation, and interface design (including HUD and haptic feedback). As an initial step toward embedded models, we benchmark stress classification on a public wearable dataset: tree-ensemble baselines achieve state-of-the-art performance for accuracy and F1 (e.g., XGBoost: \mathrmAccuracy=0.901!\pm!0.023 , \mathrmF1=0.920!\pm!0.019 ), while a feature-based Transformer attains the highest AUC ( 0.963!\pm!0.025 ), informing feature-engineering and compression choices for edge deployment. Building on these results, CAPSARII employs HW/SW co-design, model compression (pruning, quantization, distillation), and precision tuning to meet stringent energy and memory budgets; it further investigates Processing-in-Memory accelerators for future efficiency gains. A structured gap analysis highlights the remaining challenges: establishing validation standards for physiological signal quality and textile durability; achieving low-latency, scalable BAN/SAN communications under mobility and interference; and deploying hybrid lightweight/post-quantum cryptography with robust device identity and side-channel resilience on constrained endpoints. By consolidating these elements into a coherent, secure IoBT stack, CAPSARII delineates a practical roadmap from laboratory baselines to mission-ready soldier monitoring in both training and operational scenarios. Comments: Pre-print version Subjects: Emerging Technologies (cs.ET); Artificial Intelligence (cs.AI); Hardware Architecture (cs.AR); Cryptography and Security (cs.CR); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG) Cite as: arXiv:2611.20187 [cs.ET] (or arXiv:2611.20187v1 [cs.ET] for this version) https://doi.org/10.48550/arXiv.2611.20187 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Related DOI: https://doi.org/10.1016/j.micpro.2026.105355 Focus to learn more DOI(s) linking to related resources

[HC-30] WL-Headset Law: Projecting Future Headset Evolution from a Decade of Display and Mass Data

链接: https://arxiv.org/abs/2611.20083
作者: Jialin Wang,Hai-Ning Liang
类目: Human-Computer Interaction (cs.HC)
备注: 12 pages, 6 figures

点击查看摘要

Abstract:Tracking a decade of headset display and mass changes can inform forecasts. However, comparing headset resolution requires pixels per degree (PPD), which is often unreported and estimated using unclear assumptions. Therefore, we compile the open-source, evidence-linked HMD Resolution and Mass Database (HRMD) with 241 configurations from 2016–2026 and introduce structure- and axis-aware robust calibration (SARC) to estimate missing central PPD. Compared with official PPD, SARC reduces PPD estimation error from 14.8% for traditional geometric estimation to 3.2%. Based on HRMD, we introduce the Weight-Linked Headset Law (WL-Headset Law), which models PPD by year, mass, and category. The law reveals 7.4% annual PPD growth at fixed mass and category. Meanwhile, head-worn mass fell by 6.5% annually after category adjustment. At an eyeglass-like mass of 30g, the law projects average central PPD of 60 (normal visual acuity) around 2035 and 120 (higher acuity) around 2044 if trends persist.

[HC-31] urning Errors into Successes: Designing LLM Tutoring Around Scaffolded Error Correction

链接: https://arxiv.org/abs/2611.20061
作者: Michael W. Asher,Meng Cao,Paulo F. Carvalho
类目: Human-Computer Interaction (cs.HC)
备注: Includes main text (40 pages, 8 figures, 2 tables) and supplementary material (12 pages, 5 tables)

点击查看摘要

Abstract:Large language models (LLMs) could make personalized tutoring available at scale, but it remains unclear how they can best support the instructional processes that make one-on-one instruction effective. We compared two ways of using LLMs after students make errors: providing personalized explanations of those errors, or engaging students in additional cycles of scaffolded practice and feedback, with a Socratic tutor guiding students toward the correct answer. In a preregistered online experiment, 400 adults practiced setting up constraint equations and, after each attempt, received the correct answer, a generic explanation, a personalized LLM explanation of their error, or Socratic tutoring. Socratic tutoring helped: participants resolved 64% of their errors through the dialogue and scored 0.24 SD higher on practice-aligned posttest questions than those who received generic or personalized explanations. Socratic participants also found the session more interesting, less frustrating, and less costly, even though it took nearly twice as long. In contrast, we found no evidence of a benefit of personalized explanations over generic ones, and neither type of explanation reliably outperformed simply seeing the correct answer. These findings suggest that LLM tutors may add the most value not by explaining errors more precisely but by helping students correct those errors themselves through scaffolded practice. This pattern fits a mastery-experience account of tutoring, in which an effective tutor notices when a student is struggling, steps in with guidance, and works with the student until the struggle resolves into success. AI tutors may be most effective when designed around this principle.

[HC-32] VoiceGuard: An Audio-Native LLM -VLA Pipeline for Wheelchair-Mounted Assistive Manipulation

链接: https://arxiv.org/abs/2611.19984
作者: Bang Xiao,Rodolfo Cossovich
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Wheelchair-mounted robotic arms can help people with limited upper-limb mobility reach and manipulate everyday objects, but operating them typically requires stepwise joystick control and frequent mode switching. Natural-language interfaces could let users express task-level goals, yet they must handle ambiguous requests, distinguish commands from nearby conversation, and support interruption. We present VoiceGuard, a voice interface that separates spoken intent from robot execution through structured request validation and a local stop mechanism. On-device detectors handle the wake phrase and stop commands, an audio-native large language model (LLM) routes each request into a structured function packet, a rule-based ToolDispatcher accepts, rejects, or clarifies the packet, and a SmolVLA vision-language-action (VLA) policy executes accepted tasks on a low-cost SO-101 arm. Preliminary tests showed that the system correctly handled 55 of 60 voice commands, including ambiguous and non-directed utterances, completed 16 of 20 deliveries, and stopped the arm on all 20 stop commands with a median latency of 210 ms. We demonstrate the feasibility of combining flexible spoken input with constrained task execution and local interruption for wheelchair-mounted assistive manipulation.

[HC-33] From Companion to Guardian: Designing an Emergency Support System Using a Quadruped Robot for Older Adults

链接: https://arxiv.org/abs/2611.19932
作者: Chunyang Meng,Eduardo B. Sandoval,Josh Andres,Francisco Cruz,Mari Velonaki
类目: Robotics (cs.RO); Human-Computer Interaction (cs.HC)
备注: 10 pages, 2 figures. Accepted at the Australasian Conference on Robotics and Automation (ACRA 2026)

点击查看摘要

Abstract:As older adults increasingly live alone, emergency support becomes an important function for home companion robots. In Australia, demand for the care and support workforce is projected to nearly double between 2020-21 and 2049-50. Service robots such as robot dogs may support older adults by combining everyday companionship with assistance when human support is not immediately available. We present the design and implementation of a domestic companion-robot system on a Unitree Go1 quadruped platform that combines companionship with emergency awareness and response. The system uses a three-stage interaction model consisting of companion, danger-awareness, and emergency modes. A rule-based escalation policy combines signs detected from visual and speech input to determine the current risk level and adjust how long the robot waits for the user to confirm safety before entering emergency mode. Additional detected signs can shorten this waiting time. Users can confirm safety, request immediate help, stop robot motion, or reset the system. We verified the implemented decision rules and escalation timing through 105 automated trials and 25 staged physical trials. Additional visual evidence shortened the waiting time, and the system entered emergency mode when the waiting time ended without confirmation.

[HC-34] Sangyadhi: An Analytical Framework for Diagnosing and Investigating Collective Care

链接: https://arxiv.org/abs/2611.19843
作者: Shyama Sastha Krishnamoorthy Srinivasan,Mohan Kumar,Pushpendra Singh
类目: Human-Computer Interaction (cs.HC)
备注: 25 pages, 1 figure, 4 tables. Preprint of a manuscript under review

点击查看摘要

Abstract:While HCI has expanded from Personal to Family Informatics, researchers still lack a critical analytical vocabulary to diagnose what truly constitutes a collective care context. We introduce Sangyadhi, an analytical sociotechnical framework that synthesizes Situated Action, Distributed Cognition, the Capabilities Approach, and Actor-Network Theory to investigate multi-actor care ecologies. The framework articulates four interdependent analytical lenses: contextual grounding, distributed responsibility, trust-mediated sharing, and collaborative actionability. To demonstrate its diagnostic utility, we apply Sangyadhi to critically audit a decade of HCI digital health literature across self-tracking, family informatics, and clinical AI—surfacing structural omissions in how shifting authority, dynamic trust, and AI escalation are conceptualized. Ultimately, the goal of Sangyadhi is to equip researchers with an analytical blueprint for diagnosing observed care contexts, structuring field inquiries, and evaluating future technologies.

[HC-35] Developing JEEVHITAA: Designing and Evaluating a Human-Centered AI Ecosystem for Collective Care

链接: https://arxiv.org/abs/2611.19810
作者: Shyama Sastha Krishnamoorthy Srinivasan,Harsh Pala,Shahnawaz,Mohan Kumar,Pushpendra Singh
类目: Human-Computer Interaction (cs.HC); Emerging Technologies (cs.ET)
备注: 38 pages, 5 figures, 7 tables, Preprint of a Submitted manuscript

点击查看摘要

Abstract:Health management is often a collective activity, involving family members and close relations in monitoring, interpreting, and coordinating care. Yet, most digital health technologies remain individual-centric, offering limited support for the distributed responsibilities and negotiated information sharing that characterize collective care. We present JEEVHITAA, a human-centered AI ecosystem designed to address this gap. Drawing on existing literature on collective health practices, we identify key requirements and translate them into a canonical ontology and a versioned data schema that underpin a mobile ecosystem integrating wearable data, role-aware sharing, granular access controls, auditable histories, and AI-assisted health summarization and verification. We first evaluated core system primitives using 1,000 ontology-driven synthetic care circles to test data visibility and information-processing workflows. We then conducted a controlled lab study with 13 subject–care-actor dyads (26 participants), combining task-based interactions, semi-structured interviews, the System Usability Scale, and component-level usability assessments. Findings indicate positive perceptions of usability and highlight the value of granular sharing and transparent information provenance. They also expose tensions between reciprocal transparency and surveillance, and between AI-mediated assistance and potentially intrusive notifications. We discuss implications for designing accountable, context-sensitive technologies that support collective care.

[HC-36] LLM Facilitation of Group Decision-Making

链接: https://arxiv.org/abs/2611.19652
作者: José Córdova,Chun-Wei Chiang,Miriam Schirmer,Ming Yin,Emőke-Ágnes Horvát
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Human groups often fail to integrate information distributed across their members, converging on incorrect decisions. We ask whether large language models (LLMs) can help groups reach the correct decision in such settings. In an online, chat-based experiment, 138 three-person groups (N = 414) completed a hidden-profile hiring task that required selecting the best candidate using public information available to all group members and private information available only to individual members. The randomized human-subject experiment included three conditions: no facilitation, an individual-level LLM that messaged each member privately, and a group-level LLM that messaged the whole group. The LLM facilitation consisted of adaptive interventions that summarized the discussion, nudged quiet members, and played devil’s advocate. The LLM did not know which candidate was correct. Both LLM conditions significantly increased the rate at which groups reached consensus on the correct candidate, from 32.6% without facilitation to 56.5% with LLM facilitation. LLM-facilitated groups exchanged more consensus messages signaling agreement on the group’s choice. Participants in the group-level condition also disclosed significantly more information about the correct candidate, whereas those receiving individual-level facilitation repeated public information less often. Finally, group-level facilitation received more favorable usability ratings than individual-level facilitation. These results show that LLMs can deliver adaptive, theory-informed interventions that improve group decision-making, and that the design of these interventions may shape how people respond to them.

[HC-37] Are Averag e Estimates in Line Graphs Biased Toward Areas of Higher Variability?

链接: https://arxiv.org/abs/2611.19601
作者: Sheng Long
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:When estimating the average value of a data series in line graphs, do viewers systematically overweight regions of higher variability? This is the question that Moritz et al. [14] set out to answer, and across two preregistered experiments they reported evidence for a bias they termed variability overweighting. Using their released data and code, I first reproduce their reported analysis, then ask whether their inference survives an equally valid analytic choice: measuring estimation error against the average of the displayed marks of each stimulus, rather than against a reference series that, for one mark type, is never shown. Under this alternative operationalization, the effect is substantially reduced and, for one mark type, reverses. Splitting the analysis by whether variability appears in the upper or lower region of the graph further reveals an asymmetry that the original direction-corrected analysis collapses away. I argue that the deeper issue is what verbal theories leave unspecified. When a theory is silent about which operationalization it requires, the metric ends up doing the theory’s arguing for it. I close with a discussion of what reproduction and reanalysis are for: not verdicts on truth, but instruments for making analytic contingency visible and for forcing theories to say more.

[HC-38] owards Informed Autonomy Design: Understanding the Interplay of Function Characteristics and Autonomy Preferences in Automotive Agents

链接: https://arxiv.org/abs/2611.19474
作者: Sebastian Zepf,Annika Stampf,Markus Schuster,Enrico Rukzio,Mark Colley
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Proactive, autonomous interactions receive high interest in the field of intelligent assistants, including the automotive domain. A central challenge, however, is identifying appropriate levels of autonomy for different functions while maintaining user acceptance and trust. This work examines how function characteristics relate to suitable autonomy levels and whether these relationships generalize across functions. We conducted an expert workshop to define autonomy levels and identify relevant function characteristics. We then collected data from 126 participants, who rated 14 in-vehicle functions regarding autonomy preferences and function characteristics, from which we trained machine learning models to predict appropriate autonomy levels. Building on our findings, we developed our prototypical Autonomy Design Assistant that supports designers in making autonomy-related decisions, which we evaluated qualitatively with nine experts. Our results suggest that function characteristics can rank functions by their suitable level of autonomy and thereby guide more informed autonomy design decisions, whereas predictions for individual users remain uncertain.

[HC-39] Security Authentication as Protective Friction: Understanding the Dual Effects of Authentication Measures in e-Government Platform

链接: https://arxiv.org/abs/2611.19435
作者: Jianshan Pan,Zongyi Chen,Qiyou Lin,Xiaofang Cai
类目: Cryptography and Security (cs.CR); Human-Computer Interaction (cs.HC)
备注: Pan, Jianshan; Chen, Zongyi; Lin, Qiyou; and Cai, Xiaofang, “Security Authentication as Protective Friction:Understanding the Dual Effects of Authentication Measures in E-Government Platform” (2026).ICEB 2026Proceedings (Beijing, China). 80

点击查看摘要

Abstract:E-Government platforms, such as government credit platforms, handle highly sensitive personal data, requiring stringent security authentication (e.g., facial recognition, multi-factor verification) that protects users but impedes their experience. Drawing on Tomalin’s (2023) e-friction typology and the emerging literature on beneficial design friction, we reconceptualize these authentication measures as protective friction, a theoretically distinct form of non-elective, overt, impeding e-friction that serves a security-enhancing function. We propose a dual-pathway model in which protective friction simultaneously reduces privacy-security risk perception (protection pathway) and increases effort cost (cost pathway), with self-efficacy moderating only the risk pathway. This differential pattern suggests that what matters is not users’ ability to overcome friction, but their comprehension of its protective purpose, a construct we term protective mechanism comprehension. We outline a mixed-methods research design combining a survey of platform users and semi-structured interviews with platform staff at a Chinese municipal credit center to test this model. This research-in-progress contributes to IS theory by extending e-friction research from conceptual typology to empirical investigation, identifying when security friction becomes beneficial rather than burdensome.

[HC-40] Decomposing the Associative Chain: Schema-Based Mediation Supports Expression and Understanding in Human-AI Music Co-creation

链接: https://arxiv.org/abs/2611.19203
作者: Songruoyao Wu,Wanying Wu,Shuyu Li,Chongjun Zhong,Ziyi Huang,Kejun Zhang
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Users cannot reliably translate their creative intents into AI-generated music. We contend that the root of this difficulty lies in associative divergence: associations between narratives and musical features are overly remote, and users lack a visible mediation structure to negotiate them. To address this, we introduce five core schema groups as cognitive mediation and propose a bidirectional schema-based framework. Its forward pathway decomposes narrative-to-music associations into staged mappings, while its reverse pathway provides structured interpretations from music back to narrative. Findings from two user studies(N = 188; N = 60) show that the framework led to clearer creative expression and greater perceived control, improved the comprehensibility of generated music, especially for unconventional pieces. Furthermore, it supported cultural and affective thinking while helping users locate associative divergence. Our work points toward better creative intent alignment in human-AI music co-creation.

[HC-41] ANCHOR: A Port-Hamiltonian Model for Forecasting Team Dynamics in Human-in-the-Loop Systems

链接: https://arxiv.org/abs/2611.19124
作者: Diana Romero,Salma Elmalaki
类目: ystems and Control (eess.SY); Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
备注: 12 pages, 9 pages appendix

点击查看摘要

Abstract:Cyber-physical-human systems close a loop around people by sensing human behavior, predict it and act on it. Human teams are a natural target, because collaboration breaks down in ways that a timely prompt could prevent. Collaborative mixed reality makes this loop practical, since each headset senses who speaks, where its wearer looks and who stands together, and it can deliver a light prompt. In this loop, the headsets are sensors, the prompts are actuators and the team is the plant. Closing the loop needs a plant model that is small enough to learn from few teams, safe to iterate and readable by a designer. We propose ANCHOR, which models a team as a ball in an energy bowl. The bottom of the bowl is where the team rests, and it is set by the team’s recent behavior, its own history and other teams. Formally, ANCHOR is a switched port-Hamiltonian system with 52 parameters, and each part stands for a testable claim from team psychology. Because of this structure, for any fitted parameters, a forecast moves away from rest only when task progress or a prompt pushes it, and prompts of bounded size that are spaced in time move it a bounded distance. On \nGroups four-person teams held out from training, a team that drifts from its usual behavior returns most of the way within half a minute. After 32 s, about 24% of the deviation remains, and ANCHOR predicted 30%. Forecasts are clearly better when the resting point blends its three sources than when it uses one. ANCHOR matches the accuracy of a black-box baseline with about 40,000 parameters. It anticipates conversation collapse with an AUROC of 0.85. In simulation under assumed prompt effects, it removes 2.2 times as much breakdown time per prompt as periodic prompting.

[HC-42] Conditional Endorsement: Simulating How Adults with Different Skills Evaluate Canadas National AI Strategy

链接: https://arxiv.org/abs/2611.19111
作者: Dipto Das,Felix Leduc,Shion Guha
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
备注: This research was funded by the Future Skills Centre, and an earlier version of this pre-print was submitted as a report to them

点击查看摘要

Abstract:Canada’s 2026 national AI strategy, AI for All, commits to free AI training for everyone and to expanding business AI adoption, yet it is difficult to anticipate whether these policies will reach adults with uneven skills and circumstances. We ask how adults with diverse skills might rate these policies on support, effectiveness, and personal benefit, how their skills and circumstances shape these evaluations, and what implementation conditions they raise. We built 3,200 generative agents from a weighted sample of the Programme for the International Assessment of Adult Competencies (PIAAC) Cycle 2 respondents, collected their ratings and explanations for 20 policies, and analyzed dialogues between agents with contrasting ratings. The agents supported nearly every policy but were far less confident that they would improve adult AI literacy or benefit them personally. This gap widened as digital technology use declined: agents modeled on adults who had never used a digital device still endorsed the policies but rarely expected to benefit. Agents traced the gap to participation costs, hard-to-exercise rights, and limited community control, and their dialogues specified these conditions rather than resolving disagreement. We describe this pattern as conditional endorsement and argue that AI literacy policy should be designed and evaluated for adults with the fewest digital skills.

[HC-43] Instructor-designed AI improves learning in a randomized trial across 19 universities

链接: https://arxiv.org/abs/2611.19099
作者: Jun Li,Robert Bray,Sébastien Martin,Nils Van den Steen,Andrew Wu,Xiao Ni,Arielle Anderer,Ekaterina Astashkina,Volodymyr Babich,Jean-Paul Baldwin,Hamsa Bastani,Christian Blanco,Oguzhan Celebi,Christopher Chen,Shi Chen,Benjamin Collier,Yao Cui,Pnina Feldman,Blair Flicker,Brian Han,Yichun Hu,Daniela Hurtado-Lange,Jane Jiang,Zhaohui Jiang,Ismail Kirci,Murat Köksalan,Zhen Lian,Hongyan Liang,Mengshi Lu,Lesley Meng,Ken Moon,Mi Kyong Newsom,Minje Park,Chris Parker,Anyan Qi,Meng Qi,Debjit Roy,Neha Sharma,Alp Sungu,Tom Tan,Necati Tereyagoglu,Hanu Tyagi,Sriram Venkataraman,Guihua Wang,Yixin Wang,Owen Wu,Zonghao Yang,Qiuping Yu,Sophie Yu,Jingwei Zhang,Wenchang Zhang,Ruihao Zhu
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Emerging Technologies (cs.ET); Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Universities are alarmed by recent evidence that artificial intelligence (AI) can undermine learning, with students prescribed generic chatbots or centrally designed AI tools performing worse. In a pre-registered randomized controlled trial across 7,200 students at 19 universities over three semesters, we tested whether AI designed by frontline instructors can improve learning. We encouraged instructors to design AI solutions they believed would best help their students. Drawing on their unique knowledge of their classes and students, the instructors created a wide array of innovations. We show that these instructor-designed AI innovations improved student learning, measured by AI-proof assessments (proctored, in-person examinations). We found no change in student engagement, measured by course activity records, or satisfaction, measured by course evaluations. Our findings demonstrate that AI can improve student learning when its adoption is frontline-driven, with instructors empowered to design and refine AI solutions for their own classrooms.

[HC-44] AdaptType: Accessible and Customizable Multimodal Text Entry in Virtual Reality for People with Upper-Body Motor Impairments

链接: https://arxiv.org/abs/2611.19096
作者: Junhan Kong,Mingyuan Zhong,Sanjana Pavani,Jacob O. Wobbrock
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Text entry in virtual reality (VR) is cumbersome, especially for people with upper-body motor impairments. Existing techniques often demand both fine and gross motor control, yet users’ motor abilities vary widely. To address this range of needs, we present AdaptType, a multimodal VR text entry system that enables users to configure input techniques that best suit their abilities, and scaffolds customization through ability-based recommendations. Through a study with 10 participants, we found that customized multimodal techniques improved text entry speed compared to gaze- or hand-only techniques, and complemented voice for error correction, decreasing failure rate from 15.1% to 2.0%. We found that AdaptType yielded high text entry speed (36.26 WPM) and accuracy (0.36% error rate), with low task load and high usability ratings. We also found ability-based recommendations effectively scaffolded customization. Our work contributes more accessible and personalized VR text entry solutions for people with upper-body motor impairments.

[HC-45] Aspirational Identities: How VTubing Affords New Modes of Digital Embodiment

链接: https://arxiv.org/abs/2611.19055
作者: Wells Lucas Santo,Hibby Thach
类目: Human-Computer Interaction (cs.HC)
备注: 5 pages, 3 figures, Body Politics workshop at CHI 2025

点击查看摘要

Abstract:VTubing, short for “Virtual YouTubing”, is a modality of streaming where a CGI avatar is used in place of the streamer’s analog face and body. As VTubing has become increasingly popular, especially for queer and trans streamers of color, it is worth studying as a way for streamers to express identities that they hope to embody, which we call “aspirational identities”. In this workshop submission, we draw on preliminary findings from qualitative research to argue that VTubing is a potent object of study for understanding digital embodiment and the performance of identity online, especially for questions of race, gender, sexuality, and disability.

[HC-46] B2G-Agent : An LLM -Based Multi-Agent Tool for End-to-End Building-Grid Interactions

链接: https://arxiv.org/abs/2611.19036
作者: Xueyuan Cui,Qian Zhang,Xiaochen Liu,Yi Wang
类目: Human-Computer Interaction (cs.HC)
备注:

点击查看摘要

Abstract:Buildings are major electricity consumers and can interact with distribution network dispatch through flexible power scheduling. These interactions require distribution network operators (DNOs) and building energy managers (BEMs) to reconcile their operational constraints and perform complex cross-field computations. The lack of cross-domain expertise and unified computational workflows could lead to inefficient or physically infeasible interactions. To address this bottleneck, we develop an end-to-end large language model (LLM)-based multi-agent tool, B2G-Agent, that bridges DNOs and BEMs. The tool deploys three collaborative agents: a Translator for bidirectional cross-domain semantic mapping via hybrid retrieval-augmented generation (RAG); an Actor for automated configuration and execution of joint simulation and optimization through directed acyclic graphs; and a Reviewer for memory-augmented reflective failure diagnosis. This workflow abstracts away cross-field modeling complexities and enables both sides to focus on high-level operational decisions. Comprehensive evaluations across 200 synthesized tasks show that B2G-Agent improves the interaction quality score from 66.71 to 85.57 and increases the task success rate from 38.0% to 69.0% over the baseline while satisfying operational constraints. The open-source tool and interface are available at this https URL.

[HC-47] MemoryLane: Supporting Reminiscence Conversations Between Older Adults with Mild Dementia and Lay Caregivers Using a Simulated AI Agent

链接: https://arxiv.org/abs/2611.19013
作者: Soumay Agarwal,Aaryan Gajula,Lingyao Li,Renkai Ma,Jennifer Martin,Krisstina Madan,Ellen Brown,Chen Chen
类目: Human-Computer Interaction (cs.HC); Artificial Intelligence (cs.AI)
备注: 4 pages, 1 figure

点击查看摘要

Abstract:Reminiscence Therapy (RT) is a non-pharmacological intervention that invites patients to discuss and review past life experiences and memories to improve their emotional and mental well-being. In practice, many reminiscence conversations are led by lay caregivers, such as spouses and adult children. While traditional RT often relies on personal photographs, curating them can be tedious and difficult. In addition, caregivers who lack shared life experiences with the patient may struggle to facilitate engaging reminiscence conversations. This paper presents MemoryLane, an AI-integrated tool that helps lay caregivers facilitate reminiscence conversations without curated photographs. MemoryLane supports a triadic conversation among the patient, the caregiver, and a simulated AI agent that acts as the patient’s peer. A shared interface shows the agent and AI-generated images of the memories being discussed, while a private smartphone interface lets the caregiver control the behavior of the simulated AI agent and moderate the AI-generated images. A preliminary focus group study with five dementia-care stakeholders suggested the value of MemoryLane’s AI agent as a peer companion and of AI-generated images, as well as the feasibility of having caregivers moderate images and control the AI agent.

[HC-48] Who Shows Up Who Benefits and Who Wins? Expert-Novice Disparities in Team-Seeking Forums and Competition Performance on Kaggle

链接: https://arxiv.org/abs/2611.18983
作者: Ran Bi,Shiyao Wei
类目: Computers and Society (cs.CY); Human-Computer Interaction (cs.HC)
备注: 22 pages, 8 figures, 7 tables. Accepted for publication in Proceedings of the ACM on Human-Computer Interaction (PACMHCI), GROUP 2027

点击查看摘要

Abstract:Team formation is the first social gate to collaboration in online communities, yet it remains unclear whether public team-formation mechanisms broaden access to collaborators or reproduce existing hierarchies. We examine this question on Kaggle, a data science competition platform whose “Looking for a Team” (L4T) megathreads let participants publicly seek teammates for on-platform competitions. Grounded in Communities of Practice theory, we conduct two complementary studies of 18 competitions held in 2023. Study 1 analyzes the reply network among 671 megathread participants and finds that 89.3% are low-tier users, yet higher-tier users receive significantly more replies (Kruskal-Wallis H=16.46, p=.00246) and achieve higher rates of forum-visible teaming success (52.9% vs. 32.6%; chi-squared=10.53, p=.001), a proxy for eventual team membership. Linking megathread posters to verified team rosters from 16 competitions partially supports this proxy. Twenty-six final teams contained two or more megathread posters, indicating that megathread interaction sometimes, but not always, maps onto realized teaming. Study 2 links participation traces from 219 users to private leaderboard outcomes and finds that productive participation (submission count rho=.435, cross-competition participation rho=.247) is positively associated with performance, whereas megathread-specific discussion is not (rho=-.021). The submission-performance association appears stronger in higher tiers, although the novice subgroup estimate is imprecise. Together, the findings suggest that open access to team-seeking forums does not guarantee equal participation. Visible expertise shapes who attracts attention, and engagement in the community’s core practice is more strongly linked to success than discussion alone. We discuss implications for designing more equitable team formation in online collaboration environments.

[HC-49] Between Rigor and Reality: How AI Safety Benchmark Developers Understand Benchmark Use and Maintenance ACL2026

链接: https://arxiv.org/abs/2611.18930
作者: Megan Li,Ningjing Tang,Christopher Stewart,Hoda Heidari,Hong Shen
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY); Human-Computer Interaction (cs.HC); Software Engineering (cs.SE)
备注: Poster at the Eval Eval Workshop at ACL 2026. 11 pages. this https URL

点击查看摘要

Abstract:Safety benchmarks are cited as the primary tool to understand the risks posed by generative AI (GenAI) systems, yet growing evidence suggests they often fail to meet the needs of real-world safety evaluation. We present findings from five interviews with AI safety benchmark developers in academia and industry. We find a disconnect between perceptions of the intended usefulness of safety benchmarks and their practical usefulness in safety evaluations of deployed systems. Our participants offered perspectives on two challenges contributing to this disconnect: (1) the difficulty of achieving inter-rater reliability on safety constructs, and (2) a lack of clarity regarding how to address persistent threats to external validity. Based on our analysis, we argue that safety benchmarks must not only be grounded in deployment contexts, but actively integrated with them. To facilitate this, we outline pathways for more transparent communication between academic and industry AI safety stakeholders.

[HC-50] Dependable Voice Command for Multi-UAV Swarms Using Small Language Models

链接: https://arxiv.org/abs/2611.18924
作者: Afshin Ameri,Luiz Giacomossi,Håkan Forsberg,Ivan Tomasic,Baran Cürüklü
类目: ound (cs.SD); Human-Computer Interaction (cs.HC); Robotics (cs.RO); Systems and Control (eess.SY)
备注: 10 pages, 4 figures, 3 tables

点击查看摘要

Abstract:Supervisory control of multiple autonomous unmanned aerial vehicles (UAVs) imposes high visual and cognitive demands on human operators, particularly when communicating mission objectives and system constraints. Voice command provides a natural interaction modality that can reduce operator workload; however, its reliability is affected by environmental noise, speech variability, and recognition errors. This paper presents a dependable voice command framework for multi-UAV swarm supervision using Small Language Models (SLMs). The proposed architecture integrates an Automatic Speech Recognition (ASR) module with an SLM-based semantic processing layer to transform spoken operator commands into structured task definitions for downstream multi-UAV planning and execution. To evaluate robustness, an end-to-end assessment pipeline is developed using a Probabilistic Context-Free Grammar (PCFG) to generate synthetic supervisory commands, which are converted into audio samples with varying operator accents and environmental signal-to-noise ratios (SNR). The resulting speech signals are processed by the ASR module, and the SLM performs contextual interpretation and phonetic correction of transcriptions when required. The evaluation isolates the resilience of the voice command pipeline under acoustic degradation and speech variation, demonstrating reliable task generation for multi-UAV supervision in challenging operational conditions.

[HC-51] Small Data in Long-Term Environmental Vision: Annotation Noise Gamification and Label-Budget Scaling ACCV2026

链接: https://arxiv.org/abs/2611.18706
作者: Bryceton Bible,Hairong Qi,Masahiro Toyoura
类目: Computer Vision and Pattern Recognition (cs.CV); Human-Computer Interaction (cs.HC)
备注: 12 pages, 3 figures, 2 tables. Accepted to the LFSD-CV Workshop at ACCV 2026

点击查看摘要

Abstract:Long-running environmental camera networks produce abundant images but few reliable annotations. Using FujiView, a network of webcams around Mount Fuji, we study three linked questions: whether gamification affects annotation participation and agreement, how annotators disagree on ordinal visibility labels, and how label budget affects downstream classification. Across seven nonrandomized sessions with 22 labelers, gamification showed a more promising pattern for sustained participation than for labeling speed or agreement. In the closest remote-session comparison, the session recorded as gamified had longer active labeling time, while throughput and reference agreement remained similar. Annotation disagreement was substantial at fine granularity: crowd votes matched the single-annotator reference on 49.1% of exact visibility classes, versus 90.2% after grouping them into Good/Poor. In five-replicate label-budget experiments, models using 25% of the available label pool achieved held-out exact accuracy only 2.8 percentage points below the full-pool model on average. Together, these findings highlight the importance of designing reliable annotation targets, evaluating engagement mechanisms without assuming causal benefits, and using empirical scaling curves to plan labeling effort.

[HC-52] One Question Four Voices: How Advice for Alzheimers Caregiving Differs Between Caregivers Clinicians and Large Language Models

链接: https://arxiv.org/abs/2611.18604
作者: Congning Ni,Jinkyung Katie Park,Yang Li,Sarvech Qadir,Sean S. Huang,James S. Powers,Julia A. Hiner,Vania Leung,Patricia Commiskey,Lijun Song,Qingxia Chen,Pamela J. Wisniewski,Bradley A. Malin,Zhijun Yin
类目: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Human-Computer Interaction (cs.HC)
备注: Accepted to ACM CSCW 2026. 35 pages, 7 figures

点击查看摘要

Abstract:Online caregiver forums are an important source of emotional and informational support for Alzheimer’s disease and related dementias (ADRD). We compare responses to 85 real caregiver questions from ALZConnected across four responder groups: peer caregivers, physicians, ChatGPT, and a retrieval-augmented CareGPT grounded in peer discussions. Using 28 measures spanning linguistic form, emotional support, informational support, and topic scope, we examine how these voices differ when answering the same questions. Peers provided time-anchored narratives and community grounding; clinicians offered concise, advice-forward, clinically bounded guidance; and LLMs produced a formal, verbose, highly structured consultation style. CareGPT showed minimal measurable separation from ChatGPT, suggesting that retrieval plus prompt injection alone was insufficient to shift communicative stance. LLM outputs scored higher than peers on politeness and emotion-support markers and were comparable to clinicians on informativeness and physician-alignment measures, but were less narrative and less contextually situated. Topic modeling showed that LLM responses concentrated on advice-oriented themes, whereas peers covered a broader range, including emotional processing, legal and financial concerns, and daily care. These findings suggest that AI can complement, but not replace, human contributions to ADRD caregiver support and motivate hybrid systems that combine structured AI delivery with the distinct strengths of human voices.

[HC-53] DriveSafe: A Framework for Risk Detection and Safety Suggestions in Driving Scenarios

链接: https://arxiv.org/abs/2605.16892
作者: Sainithin Artham,Shankar Gangisetty,Avijit Dasgupta,C. V. Jawahar
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Human-Computer Interaction (cs.HC)
备注: 8 pages

点击查看摘要

Abstract:Comprehensive situational awareness is essential for autonomous vehicles operating in safety-critical environments, as it enables the identification and mitigation of potential risks. Although recent Multimodal Large Language Models (MLLMs) have shown promise on general vision-language tasks, our findings indicate that zero-shot MLLMs still underperform compared to domain-specific methods in fine-grained, spatially grounded risk assessment. To address this gap, we propose DriveSafe, a framework for risk-aware scene understanding that leverages structured natural language descriptions. Specifically, our method first generates spatially grounded captions enriched with multimodal context, including motion, spatial, and depth cues. These captions are then used for downstream risk assessment, explicitly identifying hazardous objects, their locations, and the unsafe behaviors they imply, followed by actionable safety suggestions. To further improve performance, we employ caption-risk pairings to fine-tune a lightweight adapter module, efficiently injecting domain-specific knowledge into the base LLM. By conditioning risk assessment on explicit language-based scene representations, DriveSafe achieves significant gains over both zero-shot MLLMs and prior domain-specific baselines. Exhaustive experiments on the DRAMA benchmark demonstrate state-of-the-art performance, while ablation studies validate the effectiveness of our key design choices. Project page: this https URL research/projects/cvit-projects/drivesafe

[HC-54] Giving Meaning to Movements: Challenges and Opportunities in Expanding Communication by Pairing Unaided AAC with Speech Generated Messages

链接: https://arxiv.org/abs/2602.22131
作者: Imran Kabir,Sharon Ann Redmon,Lynn R Elko,Kevin Williams,Mitchell A Case,Dawn J Sowers,Krista Wilkinson,Syed Masum Billah
类目: Human-Computer Interaction (cs.HC); Machine Learning (cs.LG)
备注: To appear in Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26)

点击查看摘要

Abstract:Augmentative and Alternative Communication (AAC) technologies are categorized into two forms: aided AAC, which uses external devices like speech-generating systems to produce standardized output, and unaided AAC, which relies on body-based gestures for natural expression but requires shared understanding. We investigate how to combine these approaches to harness the speed and naturalness of unaided AAC while maintaining the intelligibility of aided AAC, a largely unexplored area for individuals with communication and motor impairments. Through 18 months of participatory design with AAC users, we identified key challenges and opportunities and developed AllyAAC, a wearable system with a wrist-worn IMU paired with a smartphone app. We evaluated AllyAAC in a field study with 14 participants and produced a dataset containing over 600,000 multimodal data points featuring atypical gestures–the first of its kind. Our findings reveal challenges in recognizing personalized, idiosyncratic gestures and demonstrate how to address them using Transformer-based large machine learning (ML) models with different pretraining strategies. In sum, we contribute design principles and a reference implementation for adaptive, personalized systems combining aided and unaided AAC.

计算机视觉

[CV-0] ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization

链接: https://arxiv.org/abs/2611.21605
作者: Huawei Lin,Jinting Liu,Tongyan Wang,Xuyang Chen,Ruihua Shi,Zhichao Xu,Jiaxin Qin,Guoshenghui Zhao,Baijian Yang,Yingjie Victor Chen,Tan Yu,Weijie Zhao
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 74 pages, 11 figures, 26 tables. Code: this https URL

点击查看摘要

Abstract:Visual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existing agents often answer from common sense instead of reading the evidence. Reinforcement learning (RL) on the final answer cannot correct this, because a guess earns the same reward as reading. We propose Cross-Input Relative Policy Optimization (CrossRPO), which trains agents on contrast sets. A contrast set asks one question about several versions of an image that differ only in a small region with the evidence and have different answers, so only reading the evidence solves the whole set. CrossRPO rewards solving the whole set, and we prove that it does not reinforce a set answered from common sense. We build contrast sets for training (ViRA-Train) and evaluation (ViRA-Bench) by editing one small object in images of 21 to 120 megapixels, and we photograph further sets with real changes (ViRA-Bench-Real) to exclude editing traces. We train two Visual Reasoning Agents, ViRA-9B and ViRA-27B, with CrossRPO on only about 3,500 items of ViRA-Train each. CrossRPO achieves 63.3% set accuracy on ViRA-Bench at the maximum resolution, against 59.3% for GRPO. ViRA-9B is the most accurate open model of its size on all benchmarks, and ViRA-27B outperforms Claude Haiku 5.5 and GPT-6-Luna, both at their highest thinking level, on most benchmarks.

[CV-1] FaceJev: Direct Visual Decisions for Face Forgery Detection

链接: https://arxiv.org/abs/2611.21604
作者: Siran Peng,Tianshuo Zhang,Yibo Xue,Ajian Liu,Haoyuan Zhang,Weisong Zhao,Shuhuan Chen,Xiangyu Zhu,Zhen Lei
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Face forgery detection methods often perform well on seen manipulations but generalize poorly to unseen ones. Multimodal Large Language Models (MLLMs) can recognize facial parts and assess their visual plausibility. Existing MLLM-based methods typically express these judgments through textual explanations, but seldom treat individual part decisions as explicit learning targets. We present FaceJev, an MLLM-based method for face forgery detection through direct visual decisions. Following the Jev formulation, FaceJev answers eight independent questions about facial parts using a shared image encoding and reads anomaly probabilities from the native answer logits. A face is classified as forged if any part is predicted to be abnormal, directly linking the detection result to the part decisions. We construct PartForensics, a dataset of approximately 160k images with binary anomaly targets for the eight facial parts, to supervise these decisions. We further propose FS-GRPO, which optimizes each part decision through an exact reward comparison with its alternative, using both detection and part rewards. FaceJev outperforms the strongest baseline by 7.25 percentage points on unseen manipulations. Extensive ablation studies validate our design choices and the effectiveness of the main components.

[CV-2] When Do Histology Foundation Models Need Geometric Test-Time Augmentation?

链接: https://arxiv.org/abs/2611.21565
作者: Lars Chen,Valentino F. Foit,Dorry Segev,Michal Mankowski
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 17 pages, 5 figures

点击查看摘要

Abstract:Test-time augmentation (TTA) aggregates model predictions across multiple transformed views of an input during inference. Because histology images lack a canonical orientation, D_4 TTA—pooling across the 8 rotations and reflections of a square—is a standard default in computational pathology. However, the rise of self-supervised histology foundation models (FMs) with inherent rotational invariance calls its continued utility into question. We benchmark 9 histology FMs and 6 general-purpose models across four patch-level and two whole-slide image (WSI) datasets. We find that D_4 TTA remains a reliable post hoc improvement, but benefits diminish from +1.98 percentage points (pp) for general-purpose backbones to +0.50 pp for histology FMs on TCGA-UT, and shrink further for the largest models. To eliminate computational overhead, our experiments demonstrate that entropy routing—applying TTA selectively to uncertain predictions—recovers most of the gain while augmenting only \sim13% of samples. Finally, we show that patch-level TTA improvements do not transfer to slide-level multiple instance learning (MIL). Code: this https URL

[CV-3] Wildfire-SSR: Geolocation-Aware Multimodal Fusion for Structure-Level Wildfire Risk Assessment

链接: https://arxiv.org/abs/2611.21563
作者: Yu-Hsuan Ho,Yiming Xiao,Hao-En Lai,Ali Mostafavi,Samuel D. Brody
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Structure-level wildfire vulnerability assessment is critical for reducing losses in wildland-urban interface communities, yet existing methods often fail to combine parcel-scale ignition pathways with community-scale geographic coverage. This study introduces Wildfire-SSR (Satellite and Street-view for wildfire Risk), a geolocation-aware multimodal framework that integrates macro-scale satellite context with micro-scale street-view evidence. A Vision-Language Model extracts wildfire-relevant building and site attributes from pre-event street-view panoramas and converts them into structured semantic text. These features are spatially aligned with satellite imagery through a decoupled cross-attention fusion decoder, while a frozen satellite-pretrained DINOv3 encoder captures overhead visual context. Alternative geolocation embedding strategies are evaluated to anchor text-derived features within each satellite patch. The framework is tested on approximately 37,000 structures affected by the 2025 Eaton and Palisades fires in Los Angeles, using wildfire risk labels derived from relative post-event damage patterns to better represent inherent structural vulnerability. Rotary Positional Embedding provides the strongest spatial grounding, achieving a building localization F1 score of 0.6735 and a harmonic mean F1 score of 0.2476 for five-class pixel-level risk classification, outperforming unimodal baselines and an unanchored multimodal configuration. The model also shows a conservative bias in higher-risk classes, reducing severe-risk underestimation, and completes end-to-end assessment of the study area in 37 minutes on a single GPU. Wildfire-SSR reframes cross-view fusion as localized semantic alignment and provides a scalable tool for mitigation prioritization, resilience planning, and insurance-oriented wildfire risk assessment.

[CV-4] Reference Forcing: Long Video Generation with Region-Guided Drift Correction

链接: https://arxiv.org/abs/2611.21559
作者: Sheng Zhao,Haitian Zheng,Weikai Lin,Junjie Zhao,Yuhao Zhu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 17 pages, 12 figures, 3 tables. Project page: this https URL

点击查看摘要

Abstract:Long video generation should keep the subject and background consistent while supporting rich, natural motion. However, autoregressive models are often distilled with distribution matching distillation (DMD) over short windows, which does not train them to keep this consistency over long ranges. As a result, the subject and background can drift over time. Methods that manage the key-value (KV) cache retain or retrieve history to reduce drift. Yet this history may itself have drifted. We propose Reference Forcing, which uses the initial frames as an appearance reference. We first build a teacher that uses attention masks to force each region of a new frame to attend to the same region in the initial frames. The teacher finds these regions with off-the-shelf segmentation models. We then distill this behavior into a separate attention path of the student, which needs no external models at inference. We further introduce value correction, which corrects drifted values in the KV cache toward the matching values of drift-free initial frames. Experiments show that Reference Forcing achieves state-of-the-art subject and background consistency among long-video methods on different base models, and its advantage grows with video length. Ablations show both components are needed to gain consistency without losing motion. More visualizations and code are available at this https URL

[CV-5] BLIP-Live: Balancing Timeliness Firing Precision in Proactive Streaming Video Agents

链接: https://arxiv.org/abs/2611.21553
作者: Kanchana Ranasinghe,Honglu Zhou,Dominick Reilly,Le Xue,Shusheng Yang,Juan Carlos Niebles,Silvio Savarese,Ran Xu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Streaming video understanding is becoming increasingly important for real-world interactive systems. Such systems must be proactive for natural and useful human interaction: they should respond as soon as sufficient evidence becomes available while avoiding premature, redundant, or incorrect responses. We present BLIP-Live, a framework unifying data, metrics, harness, and training around a shared formulation of proactivity as effective causal response decisions balancing the two competing objectives of responding vs silence. First, we construct BLIP-Live-780K dataset containing human annotations of proactive response content and timing, which defines this desired behavior. We then propose PAUC-F, a metric to effectively measure our desired proactive behavior. PAUC-F combines existing response timeliness metrics with firing precision to reward early useful responses without rewarding excessive firing. With our desired behavior defined and measurable, we finally investigate model-agnostic inference harnesses that draw out proactive streaming capabilities of offline vision language models (VLMs), together with training pipelines that learn to further improve their proactive responsiveness. Across both settings, we find that combinations of sliding visual windows, online interaction memory, and structured response outputs consistently yield strongest proactive performance. Our experiments result in two models, with our BLIP-Live-Qwen model achieving state-of-the-art performance across four proactive streaming benchmarks, with gains up to 44% over baselines and outperforming multiple closed API models. Our code and data will be released publicly.

[CV-6] DuoFlow: Towards Synchronized Speech and Motion Generation with Fine-Grained Control

链接: https://arxiv.org/abs/2611.21530
作者: Jun Dong,Ronghui Li,Zhongyuan Hu,Yunqi Zhao,Lijie Guo,Ruizhang Xu,Di Kang,Linchao Bao,Xiu Li
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:An embodied speaker’s expressiveness depends not only on what is said, but also on how the voice sounds and how the body moves. However, existing joint speech-motion generation models are largely driven implicitly by transcript semantics or intermediate speech representations, offering little explicit control over motion hierarchy, body parts, or vocal style. To bridge this gap, we construct MotionSpeech, a structured benchmark for controllable joint speech-motion generation that includes transcripts, voice-style descriptions, and motion conditions at two hierarchical levels. Building on this benchmark, we propose DuoFlow, a dual-stream latent flow transformer that jointly denoises speech and motion under a shared noise level and a unified second-based temporal coordinate. Modality-specific residual connectors inject structured conditions into the corresponding streams, layer-wise bidirectional cross-modal attention coordinates their generation, and a monotonic phoneme pathway combines normalized progress anchors with CTC supervision to strengthen speech temporal modeling. DuoFlow provides a unified interface for fine-grained motion and vocal-style control and demonstrates the ability to synthesize high-quality, synchronized, and instruction-consistent speech-motion sequences.

[CV-7] GHySSeg: Real-Time Geometric Ground and Road Segmentation for Hybrid Solid-State LiDAR

链接: https://arxiv.org/abs/2611.21507
作者: Martín Palos,Fernando García
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: 14 pages, 8 figures, 3 tables. Submitted to IEEE Transactions on Intelligent Transportation Systems. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. HySoLiD-Drive dataset: this https URL

点击查看摘要

Abstract:Hybrid solid-state LiDAR sensors are making their way into production vehicles, yet the data they produce differs structurally from that of the established spinning mechanical sensors behind today’s autonomous-driving datasets: it is ring-free, dense, cluttered, and concentrated in a frontal field of view. The classical geometric methods that autonomous vehicles rely on for drivable space computation and navigation assume the ordered ring-and-column structure of a 360-degree spinning LiDAR and degrade or fail on this data, while learned alternatives are computationally heavy, hard to certify, and trained almost exclusively on spinning-LiDAR data. We demonstrate that the same dense, per-curve geometry is instead an opportunity to close the divide between the industry’s preferred sensors and autonomous-driving research. We present GHySSeg, a purely geometric method that segments ground and drivable road from a single hybrid solid-state LiDAR frame by tracing each scan curve and reading two geometric signatures, a step in height and a fold-back in range, with no training and no graphics processing unit. To evaluate it, we release HySoLiD-Drive, the first public benchmark dedicated to ground and road segmentation on an automotive hybrid solid-state LiDAR, comprising 100 manually labeled and 7000 unlabeled frames recorded by two vehicles in real driving. On this benchmark, GHySSeg surpasses both geometric and learned state-of-the-art baselines across ground and road segmentation, reaching an intersection-over-union of 0.907 for road and 0.896 for ground at over 0.96 precision, while running in real time at about 22 milliseconds per frame on a single core.

[CV-8] Lost in the Hallway: Benchmarking Persistent Spatial Memory Across Revisits

链接: https://arxiv.org/abs/2611.21485
作者: Henry Zheng,Weibo Gong,Yang Yue,Rui Huang,Yang Zhang,Gao Huang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Persistent spatial understanding is essential for embodied agents to operate reliably in environments they repeatedly encounter. It requires recognizing revisited places despite viewpoint changes, distinguishing them from similar-looking rooms, and updating memories as environments change. Yet existing evaluations often leave physical place identity across separated visits implicit. We introduce ReSpaceBench, a benchmark comprising 3,244 questions across 590 synthetic and real-home scenes, organized into 9 tasks spanning room association, persistent state tracking, and downstream reasoning. Controlled walkthroughs define visits by physical room entry and deliberately include same-type room distractors and cross-session object changes. Our evaluation of 32 models and 3 agentic video-understanding frameworks reveals substantial limitations: the strongest evaluated zero-shot model achieves 50.5% overall accuracy and 52.2% on a 450-question subset, compared with 90.1% for human annotators. The evaluated agentic frameworks offer no overall improvement over direct inference with the same base model. Analysis of agent execution traces reveals evidence omissions and incorrect cross-visit room associations as major failure modes, suggesting that iterative evidence seeking alone does not ensure consistent spatial memory. Further analyses expose a gap between recognizing familiar rooms and identifying the correct visited room, while annotation-guided sampling highlights the importance of preserving intervening visits and transitions. We also provide RS-30K, a separate training dataset from synthetic sources with no template, scene, route, or video overlap with the benchmark. Fine-tuning Qwen3-VL-8B raises overall accuracy from 35.5% to 60.4%, with the largest gains in association and improvements extending to real-life scenes.

[CV-9] LynnReal-Hand: Unlocking World Interaction with Hand Control

链接: https://arxiv.org/abs/2611.21479
作者: Shaohao Rui,Chaoyang Zhang,Xiaofeng Mao,Weijie Ma
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Interactive video world models enable exploration through text and camera controls, yet direct object interaction requires finer control. Hands provide a natural interface, but translating continuous hand movements into coherent visual responses remains challenging because of complex hand-object dynamics and demanding real-time computation. To address these challenges, we introduce LynnReal-Hand, a hand-controlled video world model that unifies precise interaction modeling with efficient streaming generation. We curate 100K interaction clips spanning diverse real and synthetic environments and develop hand-aware supervision to improve fine-grained hand motions and object interactions. Building on this foundation, we learn a causal flow map for few-step generation and introduce region-aware score distillation that combines complementary video and image priors to refine interaction dynamics and visual appearance. To further reduce inference latency, we propose Ultra-Fast-VAE, a compact video decoder that preserves high reconstruction fidelity with substantially lower computational cost. Extensive evaluations on HandBench demonstrate superior hand-following accuracy and visual quality compared with existing approaches. Our full-sequence model achieves an FVD of 181.67 and a MANO Dice score of 93.02, while the three-step causal model maintains a MANO Dice score of 92.49. For real-time interaction, 540p streaming on a single H200 reaches 18.03 FPS with 741ms latency and an optimized peak GPU memory footprint of only 34.89GB. We will release the datasets, model weights, and the full training and inference code to support research on hand-controlled video world models.

[CV-10] Event-Triggered Visual Perception for Real-Time Harvest Outcome Estimation in Mechanical Blueberry Harvesting

链接: https://arxiv.org/abs/2611.21472
作者: Nicolas Miroslav Husár,Jelle Driessen,Petra Bosilj,Başak Sakçak
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Mechanical blueberry harvesting requires a balance between removing ripe fruit and avoiding premature removal of unripe berries that could remain on the bush for a later harvest. However, operators generally lack real-time information about the maturity composition of the fruit being collected. This study develops an event-triggered machine-vision system that measures the maturity of harvested blueberries directly on the harvester. When a filled crate is moved by the machine, an RGB image is automatically captured and analysed using a three-class detector that identifies ripe, semi-ripe, and unripe berries. The system operates locally on the existing CPU-only industrial computer and therefore requires neither continuous image processing nor dedicated GPU hardware. During field testing, 79 crate images containing 48,634 visible blueberries in manually verified reference annotations were evaluated. The vision system estimated an overall ripe-fruit fraction of 90.22%, compared with 90.97% determined manually. Per-image ripe-fruit estimates showed a mean absolute error of 0.935 percentage points and a Pearson correlation of r=0.967. Maturity estimates remained similar between the two sides of the harvester for which the light conditions varied despite poorer absolute counting performance under direct sunlight. All field events were processed within the available 14 s per-side crate interval. The proposed approach provides a practical means of quantifying one component of harvest selectivity while the machine is operating. Such information could support operator decision-making and, when combined with measurements of fruit recovery and losses, contribute to future feedback-based control of mechanical harvesting.

[CV-11] Out of Sight Into Memory: Compact Visual Latents Retain Local Evidence in Ultra-High-Resolution Remote Sensing

链接: https://arxiv.org/abs/2611.21457
作者: Meng Tang,Siyuan Du,Yuxin Liu,Shuaishuai Cao,Di Wang,Haonan Guo
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Question answering over ultra-high-resolution (UHR) remote sensing scenes poses a tension for multimodal large language models: global overviews obscure small objects and local attributes, whereas detailed crops consume substantial context. Compressing inspected crops into short sequences of continuous visual latents offers a compact memory once crop tokens are discarded, yet diagnostic studies on natural images find that models often ignore latent content, leaving unclear whether compact latents can retain fine-grained targets lost in downsampled UHR remote sensing scenes. To isolate evidence retention from region selection, we encode high-resolution crops into latent memory tokens for frozen backbones and evaluate both their predictive utility and content dependence. Across two generations of Qwen-VL on LRS-GRO, a 64-token memory achieves 7.6- to 9.9-fold compression of retained crop tokens and improves accuracy by 2.5 to 4.8 percentage points over the downsampled global image alone, recovering most of the high-resolution crop gain with consistent behavior on XLRS-Bench. Substituting memory from another image or another region of the same scene reduces accuracy by 5.8 to 7.4 percentage points and yields lower accuracy than removing memory altogether, demonstrating that predictions depend on the encoded regional evidence. Restoring the source crop within the same model reduces the accuracy loss caused by zeroing or removing memory. These findings establish that compact visual latents retain useful local evidence in UHR remote sensing and identify evidence availability as a key factor governing their contribution.

[CV-12] [CLS] and [EOS] Tokens are not Enough for CLIP-based Class-Incremental Learning

链接: https://arxiv.org/abs/2611.21440
作者: Hao Sun,Tao Hu,Shixiong Kai,Siyuan Xu,Wenqian Zhao,Mingxuan Yuan,Da-Wei Zhou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to TMM. Code is available at: this https URL

点击查看摘要

Abstract:Class-Incremental Learning (CIL) aims to empower models to continuously acquire new knowledge while mitigating catastrophic forgetting. With the advent of vision-language models like CLIP, leveraging their cross-modal alignment capabilities to address CIL challenges has emerged as a dominant paradigm. Existing methods leverage CLIP’s global representations (i.e., the visual [CLS] token and the textual [EOS] token) for cross-modal alignment. While achieving effective classification, they overlook the rich patch-level visual features and token-level textual semantics inherent in CLIP. For example, the [CLS] token aggregates local information to capture the general concept of “cat”, while the patch-level details retain specific localized features essential for recognition, such as the tail, fur, and face. We propose PACE (PAtch-token Cross-modal Enhancement) to uncover the long-neglected local representations within CLIP. First, based on the global representations like [CLS] token and [EOS] token, we identify the top-k most crucial visual patches and textual tokens. Subsequently, to better utilize the extracted key local features, we design a cross-modal fusion module. This module leverages local textual semantics to refine visual representations, while augmenting textual semantics with local visual features. As the fusion module is shared across all tasks, directly fine-tuning it on new tasks would interfere with the knowledge learned from previous tasks, leading to catastrophic forgetting. To alleviate this, we employ a projection-based update strategy to effectively preserve the critical parameter space of previous tasks during the adaptation to new tasks. Extensive experiments demonstrate that PACE significantly outperforms state-of-the-art methods across multiple CIL benchmarks.

[CV-13] Power Line Detection in Point Clouds with Point-SAM and RANSAC

链接: https://arxiv.org/abs/2611.21438
作者: Christoph Dalitz,Oskar Waedt,Steffen Goebbels,Mahdieh Shamsadini Motlagh,Gabriel Gerzabek
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 10 figures, accepted for VISAPP 2027

点击查看摘要

Abstract:This article deals with the problem of detecting the position of power lines in airborne laser scanning point clouds. We address both the problem of segmenting power lines from background like ground and vegetation and of identifying individual lines. For segmentation we compare a rule-based approach with one based on transfer learning (Point-SAM). For individual line detection, we derive the mathematical formulas for RANSAC applied to 3D catenaries and compare it with a non-parametric method (TriplClust). All methods are evaluated on a newly created dataset “PowLiDFR”, which we make publicly available. The samples have been taken from the French project LiDAR HD and cover a variety of terrain layouts and point cloud resolutions. On this dataset, an unsupervised transfer learning outperformed a mere rule-based segmentation, and for individual power line detection, RANSAC was better in following the shape of catenaries than the non-parametric algorithm.

[CV-14] Quality and Similarity Indicators for Video-Based Heart Rate and Respiratory Rate Fusion: A Controlled Evaluation

链接: https://arxiv.org/abs/2611.21419
作者: Nhi Nguyen,Le Nguyen,Manuel Lage Cañellas,Miguel Bordallo López,Constantino Álvarez Casado
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 15 pages, 7 figures, 4 tables; with 14 pages supplementary material included

点击查看摘要

Abstract:Camera-based heart rate (HR) and respiratory rate (RR) estimates vary in reliability from window to window, and a common remedy is to fuse several extraction methods with weights from reference-free signal quality indicators. However, few studies test whether these indicators identify the most accurate method in each window or outperform simpler fusion. We evaluate 11 quality indicators and 8 inter-method similarity measures for fusing 7 HR and 10 RR extractors on four public datasets (75 test subjects). Filters, indicator directions, weighting parameters and a learned ridge baseline are selected on a separate five-subject set before testing. Fusion is compared with equal averaging, a constant typical-rate predictor, the best single method selected in advance and in hindsight, and a per-window oracle with a permutation-based null, using subject-level bootstrap confidence intervals. For HR, quality-weighted fusion reduced mean absolute error by 10.9% [5.2, 18.4] relative to averaging and by 15.6% [8.1, 23.1] relative to the method selected in advance, and was statistically indistinguishable from the best method selected in hindsight. Weighting methods by their agreement did not improve on averaging. For RR, the indicators could not identify the more accurate estimates: sharpened quality weighting increased error (-2.6% [-4.8, -0.8]), fused estimates were uncorrelated with the reference (r = 0.22), and averaging was less accurate than the typical-rate predictor (-3.1% [-6.1, -0.4]). With a second respiratory reference, apparent RR gains reflected estimates moving toward its typical rate. Quality-guided fusion is worthwhile for video HR; for RR, 30-s windows improved estimates but not fusion.

[CV-15] Flexible Multi-Stack Super-Resolution of Clinical Brain MRI Using Hetero-Orientation Variational Encoding: Expert Reader Evaluation in a Nigerian Cohort

链接: https://arxiv.org/abs/2611.21412
作者: Kabir Hamzah Muhammad,Lorena Garcia-Foncillas Macias,Soumya Snigdha Kundu,Emily Guilhem,Jon Cleary,Amal Saleh Nour,Yodit Abraham Yaynishet,Chinasa Kalaiwo,Peter Chibuta,Ugumba Kwikima,Udunna Anazodo,Jonathan Shapey,Tom Vercauteren
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 27 pages, 14 figures

点击查看摘要

Abstract:High-resolution brain MRI is important for neuroradiological assessment, yet routine clinical protocols often rely on fast, anisotropic 2D acquisitions, particularly in resource-constrained settings. These acquisitions may differ in orientation, alignment and anatomical coverage, resulting in partially overlapping image stacks. Existing learning-based super-resolution methods typically rely on single, under-constrained inputs or fixed multi-stack configurations with idealised acquisition geometry. We introduce MS-HVED, a sequence-agnostic framework for flexible multi-stack brain MRI super-resolution. Available axial, coronal and sagittal stacks are independently encoded into multi-scale Gaussian posteriors and fused using a field-of-view-aware Product-of-Gaussians formulation that limits contributions from unsupported regions. Physics-informed domain randomisation simulates heterogeneous clinical acquisition conditions, including obliqueness, partial coverage and imaging artefacts. MS-HVED achieved competitive or superior reconstruction fidelity across synthetic benchmarks and improved downstream tissue segmentation relative to low-resolution inputs. To assess generalisation beyond the training cohorts, we performed external clinical validation on 30 routine brain MRI examinations from a Nigerian cohort, representing an under-represented population in medical imaging research. These examinations formed the largest component of a blinded reader study involving 44 real cases and seven expert neuroradiologists. In the Nigerian cohort, MS-HVED was preferred over the supervised baseline for overall image quality in 187 of 210 paired assessments. Across all clinical cohorts, MS-HVED achieved higher overall-quality scores with all available inputs (5.93 vs. 4.49, p.001), with the preference persisting when axial or coronal inputs were omitted (p.001). Code at this http URL

[CV-16] Portrait Fitness as a Pretext: Probing Implicit Demographic Bias in VLMs WACV2027

链接: https://arxiv.org/abs/2611.21403
作者: Alexandre Fournier-Montgieux,Akram Elbouanani,Aboubacar Tuo,Adrian Popescu,Hervé Le Borgne,Bertrand Luvison
类目: Computer Vision and Pattern Recognition (cs.CV); Computers and Society (cs.CY)
备注: Accepted at WACV 2027

点击查看摘要

Abstract:We audit face-related bias in VLMs using BALPORTRAITS, a new balanced dataset of 9,000 fictional portraits across ethnicity, gender, and age segments. Sixteen VLMs estimate portrait fitness scores for occupations, brands, activities, and banking or insurance products within an advertising-casting framework designed to avoid direct inference of protected attributes. Item-level estimations are aggregated with scenario-specific prestige scores to compare rankings across demographic groups. While the magnitude and direction of these biases vary across scenarios, model families, and demographic dimensions, with age, ethnicity, and gender contributing differently depending on the task, we find demographic dispersion across all evaluated models, with a majority of pairwise segment differences reaching statistical significance. This study introduces a controlled protocol to assess implicit face-related associations and demonstrates that single-scenario or single-attribute audits are insufficient.

[CV-17] VideoNexus: Video Perception Beyond the Reasoning Context

链接: https://arxiv.org/abs/2611.21392
作者: Yancheng Long,Xuanyu Zheng,Wei Chen,Luoyuan Zhang,Haoyang Xu,Bin Wen,Tingting Gao
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Understanding long videos requires locating evidence, resolving visual details, and integrating observations into an answer. Even when relevant evidence is present, performing these operations within long, dense visual sequences burdens both perception and reasoning. Video agents enable selective inspection, but returning frames to the planner still accumulates visual information in its reasoning context. Delegating perception reduces this burden, yet sparse sampling can miss decisive details, and per-frame descriptions can preserve context growth. We introduce VideoNexus, a training-free framework that separates detailed video perception from evidence-driven planning and reasoning. The planner controls where to inspect, what to resolve, and how densely to sample. Perception sub-agents process native audio-video in independent contexts and return compact, temporally grounded evidence for iterative search and verification. Using Gemini 3.8 Flash for both planning and perception, VideoNexus achieves 88.6% accuracy on full LVBench, exceeding same-backbone static ingestion and built-in agentic processing by 3.6 and 1.1 points, respectively, at 44% of the latter’s measured API cost. A paired LVBench subset ablation shows lower planner input and higher accuracy. With Kimi K3 as the shared backbone, VideoNexus improves over the official Video-GAIA ReAct harness by 8.5 points in our evaluation. These results support compact evidence exchange for combining detailed observation with efficient planning and reasoning.

[CV-18] he Mask Is Not the Object: Volumetric Supervision for 3D Gaussian Segmentation NEURIPS2026

链接: https://arxiv.org/abs/2611.21383
作者: Jeonghwan Cho,Minsu Kim,Junyoung Hong,Seon Joo Kim
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: Accepted at NeurIPS 2026. Project page: this https URL

点击查看摘要

Abstract:3D Gaussian Splatting enables object-level operations on Gaussian primitives, yet existing segmentation methods are largely trained and evaluated through rendered views. This mismatch can produce accurate rendered masks from incomplete 3D selections since alpha-blended feature supervision concentrates learning on dominant contributors while leaving occluded or low-contribution Gaussians weakly supervised. We propose VODA, a plug-and-play supervision objective that exploits local depth structure to group per-pixel contributors and directly supervises all Gaussians in selected depth clusters, bypassing contribution-weighted gradient delivery. To evaluate volumetric completeness, we introduce Removal-3D, a synthetic object-removal benchmark, and the Removal Depth Error (RDE), a metric comparing predicted and ground-truth depth changes. Across four rendering-based baselines, VODA consistently improves volumetric segmentation on Removal-3D while maintaining or improving performance on LERF and ScanNet. Further analysis shows that VODA recovers lower-contribution Gaussians, while simple gradient reweighting and spatial propagation fail to reproduce comparable gains.

[CV-19] WeVisForensics: Self-Evolving Agent for Omni AI-Generated Video Detection

链接: https://arxiv.org/abs/2611.21373
作者: Yipeng Du,Yuxi Li,Ying Zhang,Chong Sun,Chen Li,Jing Lyu,Ying Tai
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Recent advances in video generation call for reliable detection grounded in explicit forensic evidence. However, current methods primarily depend on single-pass chain-of-thought reasoning that overlooks brief or localized artifacts, or on predefined multi-agent workflows that offer limited flexibility in selecting relevant inspections. Even with adaptive investigation, optimizing the investigation policy alone does not directly improve the forensic capabilities of fixed experts. We introduce WeVisForensics, an agentic framework in which a main investigator agent coordinates a self-evolving library of forensic inspector sub-agents spanning visual and audio cues. The investigator learns which artifact type to examine, over which window, and when to conclude, and these interactions provide shared feedback for both levels of agents. Our dual-timescale policy optimization strategy, DTPO, closes this loop through fast reinforcement learning of the investigator and slower refinement of the inspectors, guided by annotated artifact intervals, to report subtle artifacts they once dismissed. Furthermore, we curate ForenCue-30K, combining role-specific forensic annotations and interaction trajectories to supervise individual inspections and their coordination, with an accompanying reinforcement-learning pool. Experiments show that WeVisForensics averages 95.2% detection accuracy across in-domain, out-of-domain and agentic benchmarks, 9.0% above the strongest prior detector, while grounding its verdicts in localized evidence.

[CV-20] ALIVE: Making Streaming Video Models Interactive

链接: https://arxiv.org/abs/2611.21366
作者: Jinzhao Li,Yuan Xu,Jinxuan Wu,Yukun Wang,Yijia Lei,Hanyu Ji,Yinuo Chen,Zining Huang,Chi Zhang,Hang Ye,Yiwu Zhong,Honglei Yan,Panwang Pan,Yizhou Wang,Yin Li,Miao Liu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Streaming video LLMs can continuously observe the world, yet often handle user requests in isolation. We introduce ALIVE, an interactive streaming video model supporting proactive monitoring, streaming task coordination, and feedback-seeking interaction. We construct ALIVE-SFT-288K and ALIVE-RL-3K for two-stage training. Contrastive supervised fine-tuning teaches when and how to respond, while reinforcement learning with Behavior-Aligned Credit Assignment (BACA) combines trajectory-level and local advantages to refine decisions, responses, and silence across multi-turn interactions. We also introduce FSI-Bench to evaluate feedback seeking and subsequent task completion. ALIVE achieves the highest overall scores among the compared models on four public benchmarks and FSI-Bench, demonstrating strong streaming understanding and interactive capabilities.

[CV-21] Exploring the link between predictive uncertainty and the sim-to-real gap in deep learning models utilizing synthetic data

链接: https://arxiv.org/abs/2611.21365
作者: Alexander Moriz,Lukas Krebs,Amon Göppert,Robert H. Schmitt
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: A Preprint

点击查看摘要

Abstract:Today, state-of-the-art methods for quality control (QC) are based on Machine Vision, combining deep learning (DL) models with camera-based inspection. Yet, the integration of such solutions in industrial applications is often impeded by data sparsity and trust in model reliability. While synthetic data offers a promising approach to generate the necessary data with perfect annotations, it poses risks related to performance degradation due to the possible occurrence of the sim-to-real gap. This study conducts experiments using the ensemble method to estimate the predictive uncertainty of DL models trained on synthetic data. The results show that predictive uncertainty increases under the sim-to-real domain shift. However, the observed performance drop is dominated by a systematic sim-to-real bias rather than by ensemble this http URL findings contribute to the effective utilization of synthetic data in industrial settings while highlighting the importance of risk assessment for the deployment of DL solutions for QC tasks.

[CV-22] Symmetry Detection with Patch Embedding Models by Means of a Scalar Product Transform

链接: https://arxiv.org/abs/2611.21363
作者: Christoph Dalitz,Jan Beeker
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 7 figures, accepted for VISAPP 2027

点击查看摘要

Abstract:We present a transformation of vector embeddings representing image patches into scores for point reflection symmetry at the crossing points of patches. The transformation is heavily inspired by the Gradient Product Transform, but it uses the vector embeddings instead of the grayscale gradients. We therefore call it the “Scalar Product Transform” (SPT). Unlike deep learning models for symmetry detection, the method does not require a specific training step tailored for symmetry detection, but it works with arbitrary pre-trained patch embedding models. We have tested the method with DINOv3 and Franca. In our tests, the method outperformed the Gradient Product Transform and was even better than EquiSym, a convolutional neural network specifically designed and trained for symmetry detection. Although the runtime of the SPT is greater than that of EquiSym, the SPT has the advantage that it directly yields a bounding box for the symmetric region.

[CV-23] Specificity-Guided Knowledge Diversion for Decomposable Generative Models

链接: https://arxiv.org/abs/2611.21352
作者: Yucheng Xie,Fu Feng,Jing Wang,Xin Geng,Yong Rui
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Generative model pre-training typically optimizes a model as a whole without explicitly organizing its parameters for selective reuse. Decomposable pre-training instead learns reusable components for downstream adaptation while retaining strong generative performance. However, existing knowledge diversion methods incur substantial parameter overhead through weight factorization and rely on fixed component assignments to induce specialization, potentially compromising in-distribution generation quality. We therefore propose Specificity-Guided Knowledge Diversion (SGKDiv), which learns how generative knowledge is shared and specialized for effective generation and selective downstream reuse. SGKDiv constructs Knowledge Units from paired feed-forward channels, preserving native computational structure without SVD-based reparameterization. Calibrated condition-wise routing combines a shared usage prior with conditional adjustments, while complementary regularization shapes their allocation, diversity, and magnitude. Together, these designs replace predefined component assignments with adaptively learned and explicitly regulated knowledge specialization. Experiments across diverse model and training settings demonstrate better in-distribution generation than dense and MoE baselines, alongside stronger downstream adaptation than representative PEFT methods across eight datasets.

[CV-24] MovieRoll: A Systematic Framework for Rolling Out Long-Form Videos

链接: https://arxiv.org/abs/2611.21339
作者: Yunyang Ge,Weijie Kong,Ryan Xu,Yue Wu,Qi Tian,Jiangfeng Xiong,Bin Lin,Mingzhe Zheng,Shenghai Yuan,Zheyuan Liu,Jianwei Zhang,Li Yuan
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Streaming video generation must sustain quality beyond the training horizon while handling self-generated history and growing cache costs. To address these challenges, we present MovieRoll, a base-to-distilled framework that connects streaming fine-tuning, long-rollout distillation, and long-horizon inference. We first establish a robust multi-step base through Window-wise Teacher Forcing (WinTF), which jointly denoises chunks at progressive noise levels with bidirectional attention within each window and causal access to preceding history. To retain historical context efficiently, hierarchical KV-cache compression (HieraKV) exploits layer-wise attention patterns, retaining full history in shallow layers and sink–recent history in deeper layers. This reduces cache memory by up to 78% while improving generation quality. Building on this base, cumulative long tuning with distribution matching distillation (CuLT-DMD) holds parameters fixed throughout each long self-rollout and averages local gradients before updating. Each update therefore combines supervision across temporal positions while retaining only one local computation graph, enabling efficient long-rollout training of a few-step generator. To extend generation beyond the training horizon, KV recomputation bounds context for both models, while guidance annealing (AnnealCFG) reduces guidance toward the clean endpoint to stabilize MovieRoll-Base. Complementing these modeling advances, FBCA attention and PS-VAE decoding support efficient training and streaming output. On HeliosBench, MovieRoll-Base remains competitive at 30s after training on videos of up to 15s, and MovieRoll-Distilled achieves the highest total scores among the compared methods across single- and multi-event evaluations up to 60s, combining sustained naturalness with active motion.

[CV-25] A2A-UAV: A Multimodal Air-to-Air Dataset for UAV Detection and Identification

链接: https://arxiv.org/abs/2611.21311
作者: Hongtao Liang,Yiyao Wan,Kezhi Zhu,Jiayi Zhang,Zequn Jiang,Ziyu Zhou,Jiahuan Ji,Fuhui Zhou,Qihui Wu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 27 pages, 9 figures

点击查看摘要

Abstract:The rapid growth of unmanned aerial vehicles (UAVs) in low-altitude airspace increases the risk of unauthorized intrusions, making reliable detection and identification of non-cooperative UAVs essential for low-altitude airspace safety. However, the existing public UAV datasets are mostly acquired from fixed ground-based viewpoints and are limited to imaging or radio-frequency (RF) observations, which restricts the sensing coverage and deployment flexibility. To address this gap, we present A2A-UAV, a multimodal air-to-air dataset for UAV detection and identification. Specifically, 15 target UAVs of 13 commercial models were recorded at three outdoor sites under daytime and nighttime conditions by using a UAV-mounted platform equipped with an RGB-thermal gimbal camera, a panoramic camera, an RF spectrum analyzer, and a millimeter-wave radar. The dataset contains 105,294 time-aligned samples of RGB, panoramic, thermal, and RF time-frequency images and radar point clouds. Moreover, all 209,539 bounding boxes of the predominantly small UAV targets were manually annotated and cross-checked. Therefore, A2A-UAV provides a valuable resource for air-to-air UAV detection, identification, and multimodal fusion in low-altitude airspace.

[CV-26] Fix Dont Retake: Physics-Simulated Training for Correcting Patient Positioning Errors in Dental Panoramic Radiographs

链接: https://arxiv.org/abs/2611.21304
作者: Sema Helali,Lina Abu Nada,Faleh Tamimi,Rafat Damseh
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: Submitted to ISBI for review

点击查看摘要

Abstract:Patient-positioning errors can distort panoramic radiographs and lead to repeat examinations. This study investigated artificial intelligence for correcting these errors in existing images. A physics-based simulator generated paired correctly positioned and mispositioned panoramics from cone-beam computed tomography volumes, with controlled individual and combined positioning errors. Five approaches were compared: U-Net, pix2pix, conditional diffusion, an error-conditioned model, and a mixture of error-specific specialists. Evaluation used image similarity and geometric metrics across three test sets. All approaches improved mean image similarity and reduced geometric errors, with U-Net and the two error-aware approaches outperforming pix2pix and conditional diffusion.

[CV-27] Déjà-Events: Multi-view 3D Geometry with Event Cameras via Sparsity-Aware Adaptive Gating

链接: https://arxiv.org/abs/2611.21302
作者: Arda Cinar Demirtas,Pablo Varela Bernal,Marc Pollefeys,Sunghwan Hong
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注: 8 pages, 6 figures

点击查看摘要

Abstract:Event cameras keep capturing structure where conventional cameras blur or saturate, which makes them attractive for 3D perception in fast or poorly lit scenes. Event datasets with dense 3D ground truth, however, are far smaller and less diverse than the image collections behind feed-forward 3D reconstruction models. We present Déjà-Events, an event-only model that predicts depth, camera pose and pointmaps from multiple views in a single forward pass. Rather than relying on scarce 3D labels, we distil it from Déjà View, a feed-forward RGB model that refines its tokens by repeatedly applying one shared transformer block. On paired RGB and event recordings, the event student learns to reproduce the frozen teacher’s final tokens, which requires no depth or pose ground truth and trains only about 1% of its parameters. Because events are dense at edges and moving texture but nearly absent in textureless regions, we further introduce sparsity-aware adaptive gating, which adapts the shared block’s gates to events and adjusts them per patch by event density, so that accuracy holds as refinement steps are added. Trained only on simulated data, Déjà-Events achieves the lowest pointmap error among the event-based methods we evaluate, on EventScape and zero-shot on the real MVSEC and DSEC datasets. On DSEC it halves the error of the strongest baseline (0.15 vs. 0.30 AbsRel) with about a tenth of its parameters. It also has the lowest depth error on EventScape and DSEC and the lowest trajectory error on EventScape.

[CV-28] FM-Recon: Exploring Foundation Model Features for Accelerated Cardiac Cine MRI Reconstruction

链接: https://arxiv.org/abs/2611.21290
作者: Donghang Lyu,Marius Staring,Ruochen Gao,Jochen Keupp,Hildo J. Lamb,Mariya Doneva
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Foundation models have demonstrated strong feature extraction capabilities across various medical imaging tasks, yet their potential for MRI reconstruction remains largely unexplored, particularly within unrolled reconstruction networks and under artifact-corrupted inputs. In this work, we explore whether and how features from pretrained foundation models can benefit accelerated cardiac cine MRI reconstruction. To this end, we propose FM-Recon, a framework that refines and integrates extracted features from frozen image encoder of foundation models into an unrolled reconstruction network. Through extensive experiments on the CMRxRecon2023 dataset, we systematically evaluate foundation model features across different reconstruction settings. Our results show that foundation models can extract informative representations from artifact-corrupted inputs and improve reconstruction performance, particularly at moderate acceleration factors. These findings demonstrate the potential of foundation model features for accelerated MRI reconstruction and provide insights into their effective integration into unrolled reconstruction networks.

[CV-29] PLECTA: Overlap-aware instance segmentation of dense strand networks from binary masks

链接: https://arxiv.org/abs/2611.21283
作者: Oday Allan,Igor Ostanin
类目: Computer Vision and Pattern Recognition (cs.CV); Soft Condensed Matter (cond-mat.soft)
备注:

点击查看摘要

Abstract:Many materials and tissues, including carbon-nanotube sheets, collagen and paper, comprise overlapping strands whose arrangement influences their effective properties. Measuring this geometry from micrographs requires individual strands to be traced through crossings and signal gaps, which merge distinct strands or split one strand into disconnected fragments. We present PLECTA (from Latin plecta, a braid), a deterministic method that segments individual strands using only a binary image whose foreground pixels mark strand centrelines. At each junction, PLECTA selects which strand segments continue into one another while allowing segments to terminate. It formulates this decision as an exact maximum-weight partial-matching problem, with an explicit penalty for unpaired segments and no restriction on the number of segments entering a junction. The output permits crossing pixels to belong to multiple strands. On synthetic scenes, PLECTA resolved strand identities more accurately than three released alternatives at every density tested. The same ranking held on an independently published generator run at its own settings. A separate stage uses the corresponding greyscale micrograph to estimate which strand lies on top at each crossing. The complete workflow is demonstrated on scanning electron micrographs of carbon nanotube sheets, from raw images to individual strand instances with estimated crossing order, without manual annotations for training or as pipeline inputs.

[CV-30] QuantileGS: Quantile-Based Conformal Bounds on RGB Rendering Errors for Novel View Synthesis with 3D Gaussian Splatting

链接: https://arxiv.org/abs/2611.21253
作者: Haiyang Xu,Ziye Zhang,Bin Pan,Xia Xu,Zhenwei Shi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Uncertainty quantification for novel view synthesis with 3D Gaussian Splatting (3DGS) assesses rendering reliability at unseen viewpoints. This reliability depends on the discrepancy between rendered and true colors, which cannot be measured without ground-truth images. We therefore quantify uncertainty in this unknown discrepancy through upper bounds on per-pixel RGB mean absolute error. However, learned bounds may underestimate these errors and overstate rendering reliability, motivating calibration to a prescribed coverage level. In this paper, we propose QuantileGS, a post-hoc framework that combines learned rendering-error quantiles with scale-normalized conformal calibration to produce spatially adaptive error bounds. The framework comprises three phases. Phase 1, Rendering-Error Computation, measures RGB mean absolute error between observed images and predictions from a frozen 3DGS model. Phase 2, Ordered Gaussian Quantile Fields, learns two scalar attributes per Gaussian to predict lower and upper rendering-error quantiles. Phase 3, Scale-Normalized Conformal Calibration, calibrates the upper prediction on held-out images, using the quantile gap to scale the correction. Under the stated assumptions, the resulting bounds provide finite-sample marginal coverage. At a nominal coverage level of 90%, calibration raises mean empirical test coverage from 69.6% to 89.6% on LLFF and from 80.1% to 92.1% on Mip-NeRF 360. On LF, coverage remains conservative at 96.3% before and 96.4% after calibration. Applications to active view selection, Gaussian pruning, and RGB-D simultaneous localization and mapping illustrate how these bounds inform downstream decisions.

[CV-31] A Study on Performance Enhancement of a Webcam-Based Real-Time Eye-Tracking System

链接: https://arxiv.org/abs/2611.21233
作者: Minuk Jung,Kyeonghun Kim,Sooyeon Yoon
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 9 pages, 5 figures, 5 tables

点击查看摘要

Abstract:This study proposes a real-time gaze-tracking system that uses a standard RGB webcam to estimate screen coordinates. The system combines ETH-XGaze-based gaze estimation with nine-point calibration and polynomial ridge regression. To reduce errors caused by changes in head position, it also applies a conditional residual correction derived from head and pose features extracted with RTMPose. Whether this correction is used depends on the reliability of the detected features and the characteristics of the calibration data. In a preliminary experiment, one participant completed trials at 12 screen targets. The resulting mean screen-coordinate error was 204.24 pixels, which was 35.83% lower than that of xgaze-js and 57.69% lower than that of gaze_tracker. The system produced valid gaze estimates for all 12 targets. These preliminary findings demonstrate the feasibility of combining user-specific calibration with conditional pose-based correction for webcam gaze tracking. Because the evaluation involved only one participant, further experiments with more participants and a wider range of postures are required.

[CV-32] Oriented Semi-DETR for Semi-Supervised Aerial Object Detection ICASSP2027

链接: https://arxiv.org/abs/2611.21223
作者: Yuqiu Li,Jian Xue,Fan Li,Ke Lu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Multimedia (cs.MM)
备注: Submitted to ICASSP 2027

点击查看摘要

Abstract:Semi-supervised object detection has recently been extended to DETR, where oriented boxes introduce an ambiguity that does not arise for axis-aligned boxes: the map from parameters to rectangles is two-to-one, since (c_x,c_y,w,h,\theta) and (c_x,c_y,h,w,\theta+\pi/2) denote the same rotated rectangle. Losses and matching costs are computed from the parameters, so two boxes that coincide in space can be far apart in the metric that drives training. A correctly localized student box is then penalized by the regression loss, and the inflated matching cost leads the Gaussian mixture model (GMM) filter to discard it. We propose Oriented Semi-DETR, which removes this ambiguity wherever a pseudo box enters training. Swap-aligned Target Selection regresses towards the encoding closer to the student prediction. A Swap-invariant Reliability Cost minimizes over the two encodings inside the matching cost, so the filter decides on geometry alone. Swap-aware Pseudo-label Denoising starts denoising earlier, ramps its loss weight linearly, and weights each query by the teacher score and encoding margin. All three are parameter-free and require no additional forward pass. Across all label ratios of DOTA-v2.0, Oriented Semi-DETR outperforms DETR-based supervised and semi-supervised baselines, with the largest gains at the strict IoU threshold.

[CV-33] RAG E: Residual-Adaptive Gating for Flow-Based Image Editing

链接: https://arxiv.org/abs/2611.21177
作者: Ruofeng Mei,Xiangyu Xu
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Recent inversion-free, flow-based image editing methods evolve the source image toward the target via a differential velocity field. We observe that this velocity field does not explicitly account for which regions should be preserved, leading to unintended drift in non-editing regions. To address this, we formulate source fidelity control as a stochastic optimal control problem in which a selective fidelity term on the editing residual is enforced along the editing trajectory. Solving this problem and approximating its solution yields the Residual Gating Principle: each latent component is pulled back toward the source with a gain that decays exponentially with the local residual magnitude, so that small deviations are suppressed while large, semantically meaningful edits pass through unimpeded. The proposed method, RAGE, requires only one lightweight gated interpolation per sampling step, involves no mask estimation, attention maps, retraining, or architecture-specific modifications, and is applied on top of FlowEdit-based editors. On PIE-Bench, RAGE achieves the best overall rank among 17 methods and substantially improves background preservation on both Stable Diffusion 3 and FLUX backbones while maintaining competitive editing quality.

[CV-34] Beyond Token Stacking: Multi-scale Visual Modulation and Adaptive Fused-feature Calibration for MLLM s

链接: https://arxiv.org/abs/2611.21168
作者: Ye Xiao,Shesheng Song,Yi Ren,Shasha Li,Bin Ji,Xiaodong Liu,Jun Ma,Jie Yu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Current Multimodal Large Language Models (MLLMs) have advanced rapidly and follow two main paradigms: one involves concatenating visual and textual tokens for input into Large Language Model (LLM), while the other injects visual knowledge into Feed-Forward Networks (FFNs) of the LLM. However, the former introduces a large number of redundant visual tokens, substantially increasing computational overhead. The latter reduces sequence length but removes images from the input stage entirely, which may hinder the activation of image-related knowledge. Moreover, naively injecting visual knowledge into the FFNs could induce activation distribution shifts due to cross-modal semantic discrepancies. To address these challenges, we develop Beyond Token Stacking (BTS), a novel visual enhancement framework for MLLMs. Specifically, we propose a Multi-scale Visual Modulation (MVM) mechanism that injects multi-scale visual information at the input stage via modulation without increasing the token sequence length. Furthermore, we develop an Adaptive Fused-feature Calibration (AFC) module that performs per-channel affine calibration to decouple visual injection from distribution alignment in the FFN layers, improving multimodal fusion stability while preserving language priors. Experiments show that our approach achieves state-of-the-art performance.

[CV-35] DRaG AN: Dynamic Graph Attention Network for Multi-Radar Point Cloud Encoding in Automotive Radar Perception

链接: https://arxiv.org/abs/2611.21162
作者: Christof Leitgeb,Thomas Puchleitner,Max Peter Ronecker,Daniel Watzenig
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to 2027 IEEE/SICE International Symposium on System Integration (SII)

点击查看摘要

Abstract:Radar represents a key modality for weather-robust perception. Unlike optical sensors such as camera and LiDAR, radar can penetrate fog, rain and snow due to its longer wavelength, and recent hardware advances have steadily improved its spatial resolution, thus making it increasingly relevant for autonomous systems. While processing raw radar data requires substantial memory computation, most existing work has focused on preprocessed point-cloud data similar to LiDAR. However, the sparse radar points and sensor-relative anisotropy make LiDAR-oriented encoders suboptimal. To address this problem, we present DRaGAN, a dynamic graph attention network for multi-radar point clouds. DRaGAN introduces a sensor-relative Fourier encoding of angle and range to capture fixed-aperture radar characteristics, gated short- and long-range hub nodes along each sensor’s line of sight for efficient local-global message passing, and attention-guided dynamic rewiring that prunes weak edges and adds a bounded number of latent connections restricted to two-hop neighborhoods. This reduces the number of processed edges in deeper layers while preserving representation quality. On TruckScenes, DRaGAN improves test NDS by 10.10 points over the evaluated RadarGNN baseline and reduces the reported model FLOP count by 63.7%. This trade-off includes an increase in model latency from 40 to 53.16 ms and in trainable parameters from 1.2M to 1.7M on the evaluated implementations. Code and pretrained models are available at this https URL.

[CV-36] Model-based neural network for global floodplains analysis

链接: https://arxiv.org/abs/2611.21146
作者: Pierre Audisio(Phys-ENS),Barbara Belletti(EVS),Nelly Pustelnik(Phys-ENS)
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:We address the super-resolution of multispectral remote sensing imagery for long-term floodplain monitoring, enhancing Landsat acquisitions towards Sentinel-2 resolution. Exploiting the period during which both missions observed the same areas, we build DLhasa-L2S , to our knowledge the first paired dataset dedicated to multispectral super-resolution for river analysis. It covers the Lhasa River (Tibet), whose strong anthropogenic changes and diverse hydromorphological profiles make it an ideal study area. We developed two model-based neural networks: a forward-backward (FB) plug-and-play scheme and an unfolded FB scheme. We compared them against learned neural networks baselines (SwinIR, RDN) on synthetic data, on real scenes, and through a statistical test over the whole real paired Landsat-Sentinel dataset. The unfolded network achieves the highest PSNR and SSIM on the real RGB components with fast inference, offering the best trade-off between accuracy and computational cost. For NDWI-based water segmentation, its performance shows no statistically significant difference from that of the best methods.

[CV-37] Region-VAD: Video Anomaly Localization and Detection via Region Selection WACV2027

链接: https://arxiv.org/abs/2611.21132
作者: Giacomo D’Amicantonio
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: This paper has been accepted in the RWS workshop at WACV 2027

点击查看摘要

Abstract:Anomalies are local events, yet weakly supervised video anomaly detection scores them globally. Most methods pool an entire frame into a single vector, rank frames by that vector, and derive a location afterwards as a saliency map of the score. Full-frame pooling mixes the anomaly with its background, so a detector can reach competitive benchmark scores from scene context without being able to show where in the scene the anomaly occurs. We introduce Region-VAD, which first selects a region and then detects the anomaly from that region alone. Region-VAD selects an explicit rectangle over the tokens of a frozen foundation model and scores anomalousness only from the tokens inside the box, through attention pooling followed by a standard temporal head. The rectangle is not learned, it is the box in which the frozen tokens change most between consecutive segments, in all four feature projections of the encoder, and it is found by exhaustive search over the rectangles of the token grid. Only the pooling and the temporal head are trained, from video-level labels without boxes, masks or frame-level annotations, and no trained component reads a token outside the selected region. Experiments on five datasets, with one localizer configuration shared by all datasets and backbones, show that Region-VAD achieves state-of-the-art anomaly localization on UCF-Crime and MSAD at competitive detection performance.

[CV-38] How to Explain your Unexplainable Video Anomaly Detector WACV2027

链接: https://arxiv.org/abs/2611.21131
作者: Giacomom D’Amicantonio,Egor Bondarev
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: This paper has been accepted at the RWS workshop WACV 2027

点击查看摘要

Abstract:Weakly-supervised Video Anomaly Detection (VAD) models assign anomaly scores from high-dimensional features, and their predictions remain unexplained. Saliency maps show where a model looks but not what it relies on, and the descriptions produced by Vision-Language Models (VLMs) are generated by branches that the anomaly score never reads. To address this faithfulness gap, we propose \textbfExplain The Unexplainable (ETU), a classifier-agnostic framework that audits frozen VAD models through counterfactual interventions in their feature space. ETU maps video features to a discrete dependency graph of agents, actions and outcomes ( A_r \to A_a \to A_o ), regenerates the features by graph-conditioned diffusion with selected nodes of the graph masked, and measures the change in the score of a frozen detector; one set of interventions is applied unchanged to every detector. The generator is trained only to reconstruct the input features, so it cannot learn the contrast it is asked to measure, and we validate the audit with experiments that test every alternative explanation of its results, from the cost of regeneration to random deletions of matched size. On UCF-Crime features obtained from a CLIP encoder, we show that substituting the scene semantics results in the largest drop for most detectors, and that removing the outcome chain of the graph ( \Delta_rel ) changes the anomaly score of all six audited detectors far more than removing the same number of random nodes, even though the random deletion perturbs the features about as much. We show that this targeted behavior comes from the divergence objective of the graph embedding. Without it, the two interventions become indistinguishable, and with a divergence objective in the generator as well, the score change of the detector follows the displacement of the reconstruction rather than the semantics that were removed.

[CV-39] IGRecon4D: Dynamic-Object-Aware Feed-Forward Reconstruction of Driving Scenes

链接: https://arxiv.org/abs/2611.21126
作者: Yifei Tong,Ziyang Huang,Yizhao Han,Xiaoxiao Long,Hao Zhu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Feed-forward 4D driving-scene reconstruction enables efficient modeling of dynamic autonomous-driving scenes, but current methods suffer from two complementary failure modes: semantic leakage of traffic participants into the static representation and geometric drift of Gaussian locations across viewpoints. We propose \textbfIGRecon4D, a feed-forward dynamic scene reconstruction framework that addresses these failures with training-time instance-prior distillation and LiDAR-anchored Gaussian positioning. Specifically, an instance-level semantic branch provides a potentially dynamic participant prior, which is distilled into dynamic-mask learning through the Instance-Dynamic Consistency Constraint to encourage high-recall localization of traffic-participant regions prone to dynamic-static leakage. In parallel, sparse LiDAR measurements guide dense depth estimation, and the resulting depths are back-projected to anchor Gaussian positions for improved multi-view geometric consistency. Experiments on the Waymo Open Dataset and zero-shot evaluation on nuScenes demonstrate improved full-image reconstruction on the main benchmark and stronger traffic-participant-region reconstruction under cross-dataset evaluation.

[CV-40] VAE Rhythm: Temporal Decoding Traces for AI-Generated Video Detection

链接: https://arxiv.org/abs/2611.21114
作者: Ke Sun,Shuwei Li,Haiyang Mei,Mike Zheng Shou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 24 pages, 7 figures. Project page: this https URL

点击查看摘要

Abstract:Visible artifacts are an unreliable cue for detecting generated videos, because newer generators learn to remove them. We instead detect a trace left by the generation pipeline itself. Most video generators decode latents with a 3D causal VAE that groups consecutive frames in time, so the fidelity of generated frames rises and falls periodically with their position in a group. We call this periodic trace VAE Rhythm. Real videos are captured frame by frame and do not have it. To reveal it without training, we propose a phase scan. The scan reconstructs a clip with an off-the-shelf VAE from successive start offsets and measures the periodicity of the reconstruction error across these offsets. We find VAE Rhythm in most of the 39 variants of 31 generators we test, including the closed-source Sora, Kling 3.0 and Seedance 2.0. We use the rhythm as a zero-shot score, as a plug-in for existing detectors and as a data engine that pairs real videos with their VAE reconstructions to train our detector CaDet. On DF26, the training-free rhythm score outperforms all 15 retrained baselines, and as a plug-in it improves all 19 tested detectors by 6.7% to 23.7% AUROC. CaDet achieves the best AUROC on AIGVDBench, DF26 and three 2026 closed-source models, and its variant trained without any generated video still exceeds 90% on AIGVDBench.

[CV-41] Source-Free Domain Adaptation for Multi-Label Image Classification: A Framework Baselines and Benchmark

链接: https://arxiv.org/abs/2611.21107
作者: Salih Numan Büyükbaş,Erchan Aptoula,Mamatha Thota,Petra Bosilj
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 21 pages, 5 figures, 6 tables

点击查看摘要

Abstract:Source-free domain adaptation (SFDA) adapts a model trained on a labelled source domain to an unlabelled target domain without access to the source data. Almost the entire SFDA literature assumes mutually exclusive classes, yet many deployed settings are multi-label. This paper formulates SFDA for multi-label classification and establishes the first benchmark for it: five adaptation tasks over six datasets spanning remote sensing, natural, clipart-style and driving imagery, with four single-label SFDA methods extended to provide the first baselines and a published multi-label method run alongside them, over three seeds. The benchmark shows that extending a single-label method is unreliable rather than merely suboptimal: half of the extended baselines fail to improve on the unadapted source model at all, and on two of the five tasks none of them does. The failures are not incidental. Nearest-centroid pseudo-labelling, softmax-based confidence, the diversity term of information maximization and confidence-weighted prototype construction each rely on mutual exclusivity, and each fails differently without it, so replacing the softmax with per-class sigmoids is a valid but insufficient extension. Multi-label Prototype Alignment with Contrastive Teaching (MPACT) is proposed in response, a mean-teacher framework whose objective, prototype construction and alignment term are each reconsidered for overlapping label sets. It attains the best average on every reported metric and is the only method that improves on the unadapted source model on all five tasks. The analysis further introduces a prototype-agreement diagnostic separating feature-space alignment from self-reinforcing agreement with the classifier, and reports a negative result for confidence-based pseudo-label filtering.

[CV-42] CivilBench: Benchmarking Vision-Language Models for Multimodal Information Integration in Civil Engineering

链接: https://arxiv.org/abs/2611.21105
作者: Dan Yuan,Yue Wen,Zhili He,Zhenkui Zhou,Shicen Liu,Ziqiang Zheng
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: 17 pages, 8 figures

点击查看摘要

Abstract:Artificial intelligence (AI), particularly large language models (LLMs) and vision-language models (VLMs), is increasingly applied in civil and structural engineering for disaster assessment, structural analysis, and engineering decision support. However, the extent to which current VLMs possess the reasoning capabilities required for reliable engineering judgments remains unclear. In this work, we introduce CivilBench, a comprehensive benchmark for evaluating VLMs in civil engineering and disaster-related scenarios. CivilBench covers three major task families: Sequential Reasoning, Image-based Structural Understanding, and Open-world Understanding, encompassing numerical reasoning, structural damage assessment, risk inference, causal analysis, and repair planning. The benchmark is constructed through a systematic pipeline of domain-specific data collection, engineering-oriented captioning, expert annotation, question design, and quality verification. Evaluation of representative state-of-the-art VLMs reveals substantial limitations in domain-specific reasoning, with Sequential Reasoning emerging as a major bottleneck, particularly when models must infer numerical relationships from sparse structured observations. Further experiments show that numerical priors and external knowledge can alleviate these limitations, although their effectiveness depends on the model’s ability to reliably interpret and use the available evidence. CivilBench provides a capability-oriented testbed for identifying reasoning limitations and evaluating progress toward reliable AI-assisted engineering decision-making. The benchmark will be publicly available at this https URL.

[CV-43] GRAZE: Text-Free Grasp Synthesis for Unseen Objects from Human Gaze

链接: https://arxiv.org/abs/2611.21104
作者: Jeongho Lee,Boeun Kim,Hyung Jin Chang,Jungchan Cho,Sang-Il Choi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 23 pages, 12 figures

点击查看摘要

Abstract:Text-conditioned hand-object interaction (HOI) generation provides intuitive semantic control, but language is an indirect way to specify where an interaction should occur on an object. Naming the part to grasp requires object-specific words, and naming the approach side requires a reference frame that a prompt lacks. We propose GRAZE (GRAsp from gaZE), which generates hand-object grasp motion from recorded gaze instead of object and part names. Its Geometry-Aware Gaze Encoding (GeoGaze) maps gaze rays to cues on object geometry, and cross-attention conditions a diffusion Transformer on these cues. We also introduce HOT3D-Grasp, a grasp-centered benchmark with object-disjoint folds built on HOT3D recordings. On unseen objects, GRAZE grasps the target part more often than the strongest text-conditioned baseline (49.3% vs. 43.6%), achieves the highest grasp success (66.3% vs. 46.6%), and transfers without fine-tuning to GazeHOI (45.3% vs. 33.8%). Gaze conveys the approach side: when gaze is rotated around the object, the approach turns with it in 59.1% of cases, against 20.5% for a text model retrained with clock words. On the same backbone, a text prompt names parts better, but gaze adds the approach side. Given both, GRAZE keeps the prompt’s part accuracy and follows gaze.

[CV-44] Beyond Family Labels: A Taxonomic Ornstein-Uhlenbeck Prior for Avian 3D Shape Recovery NEURIPS2026

链接: https://arxiv.org/abs/2611.21102
作者: Yunchan Jeon,SooGon Kim,Seungwook Kim,CheolWon Lee,Jongmin Lee
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: NeurIPS 2026 (Oral)

点击查看摘要

Abstract:Recovering 3D pose and shape of birds from single images is bottlenecked by the long tail of avian diversity: out of ~10,000 species, only a few dozen have 3D-labeled training data, and in-the-wild test sets routinely contain species never seen at training. Generalizing across clades is therefore a central problem. The current state of the art, AniMer+, mitigates it with a family-level contrastive loss that conveys only a binary same-family indicator per pair and discards the continuous structure of the species tree. We propose a taxonomic Ornstein-Uhlenbeck prior on the shape code of the AVES parametric bird mesh model: a multivariate Gaussian whose covariance decays exponentially with Linnaean rank distance, with a learnable selection strength and noise scale. The prior generalizes family contrastive into a continuous, generative form, exposing the graded cross- and within-family similarity that a binary indicator cannot, and brings continuous comparative phylogenetics in as a biology-grounded inductive bias for the long tail. Our method improves over AniMer+ on both metrics of the unseen-species CowBird benchmark and on CtrlAVES3D, while remaining comparable on seen-species CUB. A permutation ablation isolates taxonomic structure from generic Gaussian smoothing, and a rank-granularity ablation identifies cross-order distinctions (the structure beyond a family-vs-not indicator) as the dominant source of the unseen-species gain.

[CV-45] MedicalDPO: Multi-Objective Direct Preference Optimization for CT to PET Translation

链接: https://arxiv.org/abs/2611.21098
作者: Huiting Zhang,Yingxin Cai,Siyuan Wang,Xi Yang,Yuyao Yan,Rui Zhang,Kaizhu Huang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:CT-to-PET synthesis estimates functional PET images from anatomical CT scans for auxiliary analysis, representation pretraining, and missing-modality completion. Brownian Bridge diffusion models can produce multiple PET candidates of different quality under the same CT condition, exposing an intra-condition quality hierarchy. Candidate quality spans complementary spatial, structural, and intensity dimensions, while DPO-only fine-tuning may compromise absolute synthesis fidelity. We propose MedicalDPO which converts this hierarchy into preference supervision automatically constructed from paired CT-PET training data. Multiple candidates are ranked across the three fidelity dimensions, and strict cross-metric verification retains preferred-rejected pairs with consistent quality separation. Relative denoising-error preference optimization is jointly trained with the original Brownian Bridge objective to anchor the pretrained CT-to-PET mapping, while condition-balanced multi-pair rotation distributes preference supervision across CT conditions. Experiments on BBDM and CPDM show improvements across LPIPS, MAE, SSIM, and PSNR. Further analyses demonstrate metric complementarity and clear pair separation, while ablations show that the original training objective prevents the substantial degradation caused by DPO-only fine-tuning.

[CV-46] State-Message Interaction Adapter for Cross-Domain Few-Shot Segmentation

链接: https://arxiv.org/abs/2611.21085
作者: Hang Xing,Yan Xia,Yongjiang Liu,Siqi Zhang,Xueming Ding
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 17 pages, 9 figures, 7 tables, including the appendix

点击查看摘要

Abstract:Cross-domain few-shot segmentation adapts natural-image pretrained encoders to medical or industrial domains from a few pixel-annotated support masks. Within frozen DINOv3, intermediate per-head attention messages already carry target-background distinctions that baseline final-layer similarities and predictions do not consistently reflect. We propose the State-Message Interaction Adapter (SMIA) to learn how per-head messages should update each receiving token representation while keeping pretrained attention and FFN parameters frozen. Inserted right after per-head aggregation, SMIA conditions the transformation of each head’s aggregated output (message) on each patch’s pre-attention feature (state), allowing similar contextual messages to produce distinct or opposing updates depending on the receiving position. Because scarce support masks cover only limited state-message combinations, a shared low-rank bilinear map confines the learned interactions to low-dimensional state, message, and output subspaces. Added before each adapted block’s frozen FFN, these updates yield adapted features that are refined by frozen FFNs and deeper blocks for linear readout. With only 0.51M trainable parameters and single-pass, support-free inference, SMIA achieves six-dataset mean mIoUs of 77.40% at 5 shots and 80.99% at 10 shots across 20 medical and industrial tasks, surpassing the strongest prior methods by +3.61 and +3.93 percentage points. In controlled 10-shot comparisons across three benchmarks on the same DINOv3 backbone, SMIA also outperforms parameter-matched AdaptFormer, LoRA, and ParaX with a better precision-recall balance.

[CV-47] MORE: Merge Once Route Early for Training-Free Multi-Subject Composition from Large LoRA Pools

链接: https://arxiv.org/abs/2611.21077
作者: Tian Xia
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Public hubs offer many personalized LoRAs per text-to-image backbone, yet composing two of them in one image remains fragile. In FLUX, keeping a pool of K=10 or 20 subject LoRAs loaded makes their deltas interfere, and even loading only the two requested LoRAs keeps both subjects in fewer than 4% of images: one subject vanishes, the two fuse, or the image splits into a collage. We present MORE (Merge Once, Route Early), a training-free two-stage pipeline for FLUX. Merge once: the whole pool is folded into a single fixed-rank adapter by the subspace-regression router of SSR-Merge, made reliable by ridge, rank and calibration choices, and factored exactly into a shared part and per-subject parts. Route early: during the first 30% of the denoising steps each subject’s part is gated to its layout region; a decaying soft attention bias ties every region to its own subject tokens, and a cyclic swap guidance extrapolates away from the prediction with the subjects’ regions permuted, which counteracts region homogenization. On held-out seeds, MORE keeps both subjects in 94.2% ( K=10 ) and 78.3% ( K=20 ) of images, against 50.8% / 35.8% for FreeFuse and 25.0% / 24.2% for LoRAShop, and reaches an ImageReward of 1.32 (best baseline 0.07), while all K subjects stay in one adapter. With layouts inferred automatically it still keeps both subjects in 70.0% / 55.0% of images.

[CV-48] What Makes a Good Move? Learning to Act from Experience in Visual Worlds

链接: https://arxiv.org/abs/2611.21065
作者: Cehao Yang,Ziyi He,Jing Yi,Jian Guo,Hui Xiong
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Learning visual policies requires turning interaction experience into guidance for individual decisions. Trajectory outcomes obscure action quality, while unguided exploration can revisit familiar situations. We present \textscLightAct, an experience-learning framework that connects skill-guided exploration with policy internalization. Interaction segments are summarized into experience entries and consolidated into a game-specific skill that guides root collection toward broader state coverage. At each root, candidate actions are compared through short rollouts under a shared frozen policy. Environment-reported score gains are averaged across continuations and normalized into relative action rewards, producing reusable supervision for reinforcement learning. Policy training uses a compact dataset of just 1,024 decision roots collected in a single source game and paired with their action rewards. Across 170 GameWorld tasks spanning 34 games, \textscLightAct improves overall progress by 23.6, 29.2, and 13.0 percentage points at the 2B, 4B, and 9B scales, respectively, with our strongest model competitive with reported proprietary-model results. Ablations show broader root-state coverage with skill guidance and consistent gains across three policy optimizers. Case studies show how rollouts distinguish actions with identical immediate gains.

[CV-49] MetricHand: Geometry-Aware Flow Matching for 4D Hand Reconstruction in Metric World Space

链接: https://arxiv.org/abs/2611.21060
作者: Yibo Bai,Likui Zhang,Jinglei Zhang,Zhengbin Pang,Xingshen Song,Xiaoqing Yin,Jinsheng Deng
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Egocentric hand motion provides physical motion supervision for dexterous manipulation and vision-language-action pretraining. Reliable use of this supervision requires accurate and temporally coherent 4D hand reconstruction in metric world space. Existing methods typically lift framewise estimates produced under weak perspective into world coordinates. By ignoring joint-dependent depth variation at close range, this approximation couples errors in translation, orientation, and articulation that persist after world lifting and worsen under occlusion. We present MetricHand, a geometry-aware flow matching framework for 4D hand reconstruction in metric world space. MetricHand encodes depth priors from hand-object interaction regions into joint-associated representations that preserve local spatial structure and metric scale. Fused with visual features, this geometry guides coarse metric estimation and conditions every step of temporal generation. The model jointly recovers world-space translation, global orientation, MANO articulation, and shape, while structural priors and temporal context support estimation under partial occlusion. MetricHand reduces camera-space MPJPE by 49.3% and 52.9%, absolute world-space MPJPE by 28.7% and 21.7%, and wrist localization error by 48.5% and 37.9% relative to HaWoR, respectively. World-space acceleration error decreases by 44.2% and 64.3%, demonstrating improved temporal fidelity alongside more accurate absolute localization.

[CV-50] Dataset Distillation for Handwritten Text Recognition: An Empirical Study

链接: https://arxiv.org/abs/2611.21054
作者: Félix Fuentes-Hurtado,Aric Ayllón,Juan C. Alfaro
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Dataset distillation (DD) replaces large training datasets with smaller synthetic ones to reduce computational costs. While traditionally used for classification, we present, to our knowledge, the first DD study for line-level handwritten text recognition (HTR). Unlike classification tasks, which use discrete labels, HTR employs unaligned sequences from a virtually infinite output space as labels. To adapt classification-based DD methods to this scenario, we define text pseudo-classes by clustering transcriptions based on character n-grams. We adapt gradient matching (GM), distribution matching (DM) and trajectory matching (MTT) to a Convolutional Recurrent Neural Network (CRNN) trained with Connectionist Temporal Classification (CTC). In this process, synthetic line images are generated by optimising only the pixels of real starting lines (either randomly chosen or selected as text medoids), while keeping their original transcriptions fixed. Across four datasets (IAM, George Washington, READ,2016, LAM) and three budgets (1–10,% of the training lines), no distilled set outperforms its real initialisation lines. Random lines prove better than medoids, and further optimisation only degrades recognition. Methodologically, DM lowers its objective without benefit, while GM and MTT barely move theirs. We turn these findings into guidelines for evaluating data compression in HTR.

[CV-51] Body of Evidence: Triangulating 3D Human Pose with Geometry-Conditioned Depth

链接: https://arxiv.org/abs/2611.21044
作者: Erik Brorsson,Kilian Freitag,Josef Bengtson,Kristofer Bengtsson,Lars Hammarstrand,Knut Åkesson,Fredrik Kahl
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Even in multi-camera setups, occlusions can leave individual body joints with only a single usable 2D observation. Such an observation defines a viewing ray but leaves the joint’s depth undetermined. We introduce Body of Evidence (BoE), a multi-view 3D human pose estimator that predicts per-camera joint depths alongside 2D joint locations. To reduce monocular depth ambiguity, we propose geometry-conditioned depth prediction, which combines image features with previously reconstructed 3D body geometry as contextual depth cues. We further introduce learnable depth-aware triangulation (LDAT), which augments differentiable algebraic triangulation with predicted depth constraints, enabling reconstruction even from a single contributing camera while preserving the benefits of geometric triangulation when multiple views are available. The reconstructed 3D poses subsequently guide depth prediction in the next iteration, allowing depth estimates and 3D geometry to progressively refine each other. Experiments on CMU Panoptic demonstrate state-of-the-art in-domain performance and superior average performance across evaluated camera subsets and unseen camera arrangements. BoE is particularly effective under sparse camera coverage, outperforming the previous state of the art by 10.6 AP25 points in the challenging three-camera setting.

[CV-52] ScoutCache: Efficient Few-Step Autoregressive Video Diffusion via Conditional Readout and Selective Caching

链接: https://arxiv.org/abs/2611.21038
作者: Hao Tan,Jinghui Qin,Xuming Hu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Few-step autoregressive video diffusion enables efficient long video generation, but inference still incurs substantial computational costs from repeated denoising and KV cache operations. Existing acceleration methods reduce parts of this overhead, while historical KV reading and cache construction remain costly across video chunks. To address these challenges, we propose ScoutCache, a training-free framework that jointly optimizes denoising, historical KV reading, and cache construction. We observe that shallow feature changes provide useful information for estimating clean latent updates, while historical KV dependence varies across Transformer layers and attention heads but remains relatively stable across video chunks. Based on these observations, we develop Conditional Denoising Readout (CDR) to estimate intermediate denoising updates from shallow feature changes and known noise variations. We further introduce Output Risk Guided Historical KV Reading (ORHR) to allocate historical reading budgets based on attention output perturbations, and Content and Coverage Guided Sparse Cache Construction (CCSC) to reduce cache construction costs by selecting informative spatial tokens while maintaining spatial coverage. Extensive experiments on Self-Forcing and LongLive using VBench and MovieGen demonstrate that ScoutCache achieves up to 2.28 \times and 1.86 \times speedups, respectively, while consistently improving generation quality without modifying pretrained model parameters.

[CV-53] POKE: Probe Offline Keep Efficiently for Vision-Language-Action Models

链接: https://arxiv.org/abs/2611.21033
作者: Jaehyuk Lee,Yeajin Lee,Suhyun Bae,Donghun Lee
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Visual token pruning reduces the inference cost of Vision-Language-Action (VLA) models, but attention-based scores do not directly measure a token’s effect on the predicted action. We define action sensitivity as the magnitude of the first-order action change when a token is replaced by its camera’s mean token and estimate it for all tokens with a single backward pass. Computing it online is costly: scoring on the full (unpruned) graph needs an extra forward pass at every control step. Scoring on the pruned forward pass is cheaper but causes scoring-graph lock-in: removed tokens lose their direct paths through later layers and tend to receive lower scores, making those locations less likely to return. On replayed observations, even with a random conditioning mask, pruned-graph scoring re-selects 59% of the mask’s tokens, versus 6% expected by chance; attention histories show a similar bias. POKE (Probe Offline, Keep Efficiently) computes full-graph action sensitivities offline and distills them into a student with under 0.3M parameters. At inference, the student scores the current observation without backpropagation or previous masks, and score-guided diversity selection retains the tokens. Keeping 32 of 512 visual tokens on LIBERO, POKE achieves 95.2% success on OpenVLA-OFT and 90.8% on \pi_0.5 at 1.85\times and 1.26\times speedups, outperforming attention-, diversity-, and distillation-based pruners, with larger margins on LIBERO-Plus.

[CV-54] Sparse Attention for Visual Generation: Toward Algorithm-System Co-Design

链接: https://arxiv.org/abs/2611.21029
作者: Majun Zhang,Yuan Guo,Decheng Wu,Jiayang Xu,Changhao Pan,Fan Zhuo,Xiaoda Yang,Chenyuhao Wen,Zhou Zhao
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Rapid advances in image and video generation have enabled high-fidelity synthesis at higher resolutions and longer durations, making the quadratic cost of dense self-attention a major bottleneck as visual token sequences grow. Sparse attention seeks to preserve generation quality at lower computational cost by retaining informative query-key interactions and omitting less consequential ones, with end-to-end gains depending on the cost of constructing sparse support and the efficiency with which the system maps the resulting workload to hardware. We survey sparse attention for visual generation through a 4+1+3 analytical framework that links algorithmic design to system implementation. The framework examines how static and dynamic methods use source information and the evidence derived from it to decide which interactions to retain and at what granularity, while accounting for the roles of reuse in reducing repeated computation and correction in compensating for omitted contributions. We relate these algorithmic choices to their implementation on hardware through the framework’s three system dimensions of data organization, execution, and dataflow. We summarize common evaluation metrics and discuss open challenges and research directions, providing researchers and practitioners with a reference for the joint design of sparse attention algorithms and systems.

[CV-55] Digits in Notation Are Not Numbers: A Scoring Pitfall in PDF-to-Markdown Conversion of Japanese Technical Documents (Preliminary Evaluation)

链接: https://arxiv.org/abs/2611.20993
作者: Mitsumasa Wada(Faculty of Engineering and Design, Kagawa University)
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Preliminary evaluation. 26 pages

点击查看摘要

Abstract:Converting PDFs to Markdown is now routine preprocessing for search and language models, yet scoring the numbers in the output is easy to get wrong. We evaluate MinerU, a plain-text OCR baseline (NDLOCR-Lite), and a deterministic post-processing pipeline (anchor insertion, footnote reordering, fixes of look-alike characters in fixed contexts, gated formula repair using two independent OCR partners, and heading reconstruction) on 48 public Japanese pages (three documents), text-layer and rasterized. Our pre-specified number-and-unit F1 favored the OCR baseline (0.575 vs 0.397), and a disagreement check between the two engines seemed to flag about half of the converter’s number errors (54.6%). Both were artifacts. The scorer counted digits in the converter’s notation (image-link paths, HTML attributes), which plain text lacks. With the notation removed (a post-hoc check), the F1 scores tie (0.575 vs 0.594), and the check flags only 13% of number errors, because most errors are misses that both engines share. What remains: the OCR baseline has a lower body character error rate (0.012 vs 0.053 rasterized). The pipeline raises extra-number recall at page level (+0.145; the document-level interval touches 0), mostly by restoring page numbers, but lowers heading-level agreement (0.133 to 0.061) and leaves spurious formulas unchanged; its formula repair never fires. A pre-specified precision gate for its flags (0.80, human-judged) is met at 0.929, yet the post-processing hypothesis fails on its improvement items. We recommend scoring numbers after removing notation and checking critical figures against the source. By the pre-specified criteria, the study stays a technical report.

[CV-56] MedJev: A System One Vision-Language Decision Model for Fast and Effective Medical Image Understanding

链接: https://arxiv.org/abs/2611.20992
作者: Fan Zhang,Hao Wu,Jinyang Wu,Hongyu Qu,Zheng Lian,Dongfu Yin,Qingsong Wen,Pheng-Ann Heng
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Medical image understanding with vision-language models (VLMs) typically relies on autoregressively generating reasoning and answer tokens. These unnecessary tokens make the inference cost scale with the response length and force the answer to be parsed from free-form text, incurring avoidable latency and decision errors. To this end, we introduce \textscMedJev, a \emphsystem one vision-language decision model that maps medical images, questions, and candidate answers directly to decisions in a single multimodal forward pass, without any autoregressive generation process. For fast inference, \textscMedJev removes the upper language layers and the generative output pathway, while keeping the full vision pathway and all visual tokens. On this shortened backbone, a decision head scores each candidate by pretrained query–key matching and refines the scores with a lightweight multilayer perceptron (MLP) residual, coupling the pretrained matching structure with task-specific nonlinear discrimination. Built on Qwen3.5-4B, \textscMedJev-4B improves both accuracy and speed over the original model on each of four medical image classification and visual question answering benchmarks, with an absolute accuracy gain of \textbf7.93% and a \textbf6.04 \times end-to-end speedup. Further comparison with 20 cutting-edge models shows that \textscMedJev-4B achieves state-of-the-art performance and inference speed among 4B–9B general and medical VLMs, indicating that direct multimodal decision making is an effective route to fast and effective medical image understanding.

[CV-57] FAVR: Flow-Aware Visual Residual Alignment for Vision-Language-Action Policies

链接: https://arxiv.org/abs/2611.20971
作者: Qixuan Wu(1),Xiaole Yang(2),Yuguang Yang(3),Canyu Chen(4),Linlin Yang(5),Haixia Pan(2),Guodong Guo(6),Baochang Zhang(1) ((1) School of Artificial Intelligence, Beihang University, Beijing, China, (2) School of Software, Beihang University, Beijing, China, (3) School of Electronic and Information Engineering, Beihang University, Beijing, China, (4) National College for Excellent Engineers, Beihang University, Beijing, China, (5) Communication University of China, Beijing, China, (6) Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, USA)
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 5 figures. Equal contribution: Qixuan Wu and Xiaole Yang. Corresponding authors: Yuguang Yang and Baochang Zhang

点击查看摘要

Abstract:Pretrained video representations provide rich temporal information for robotic manipulation, yet effectively integrating them into existing vision-language-action (VLA) policies remains challenging. We present Flow-Aware Visual Residual Alignment (FAVR), a lightweight framework for flow-aware visual conditioning of frozen VLA policies. FAVR maintains a compact temporal visual state through linear Test-Time Training (TTT) and retrieves action-conditioned visual features throughout flow-based action generation. By exploiting the affine structure of the frozen action head, we explicitly characterize how injected visual features influence predicted action velocities and formulate flow-aware channel selection as a box-constrained optimization problem. We develop a FISTA-based solver for joint gate optimization and derive an efficient closed-form approximation through diagonalization and rescaling. With only 1.74-2.19M trainable parameters and 5,000 adaptation steps, FAVR improves pretrained VLA policies across benchmarks. On RoboTwin 2.0, FAVR improves LingBot-VLA from 87.12% to 90.04% under clean conditions and from 86.88% to 89.38% under randomized conditions. A controlled ablation on five RoboTwin tasks further shows that flow-aware gating increases average success from 72.0% with a gate shared across flow steps to 82.0%, demonstrating the benefit of adapting visual conditioning throughout action generation.

[CV-58] vividRGB: Various visual cue detection in RGB photographs with forced focused attention

链接: https://arxiv.org/abs/2611.20942
作者: Masahiro Ryo
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 32 pages including supplementary material; 6 numbered figures and a graphical summary

点击查看摘要

Abstract:Natural photographs contain dense, overlapping structures that can be weakly expressed in feature maps from whole-image vision encoders. We introduce forced focused attention (FFA) and its Python implementation, vividRGB. FFA is an encoder-agnostic inference-time procedure that applies a frozen encoder to enlarged, overlapping local crops, averages inverse-aligned reflections, and reconstructs a feature field in source coordinates. We evaluate 50 high-resolution photographs from grassland vegetation, fungal networks, airborne forests, coral reefs, and phytoplankton microscopy, using DINOv3 ViT-B/16. Across the five domains, FFA at 5% or 10% focus produces higher mean RGB-edge boundary F1 than direct DINOv3 encoding and automatic SAM3 masks. Domain means range from 0.244 to 0.596 for FFA, versus 0.028-0.076 for direct DINOv3 and 0.020-0.090 for SAM3. Qualitative feature maps also show finer appearance structure. These results concern RGB-edge preservation rather than verified species, traits, or biological segmentation. Narrower focus generally improves boundary preservation, while extraction time and GPU energy use increase sharply and gains diminish at finer settings. FFA may be useful for relatively small, high-resolution ecological image collections where dense visual cues matter. The alpha release vividRGB 0.1.0 and versioned research resources support reproducible cue inspection.

[CV-59] Learning What the True Labels Would Teach: Posterior Analytic Class-Incremental Learning with Unlabeled Data

链接: https://arxiv.org/abs/2611.20937
作者: Wei Tang,Tong Wei,Jiahao Zhang,Min-Ling Zhang,Lijie Hu,Fakhri Karray
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:In class-incremental learning with unlabeled data, each task brings only class names and unlabeled images, so CLIP’s zero-shot predictions replace labels. Existing methods train on hard pseudo-labels, but a wrong one gives the true class zero probability, which self-training often confirms. We argue that each image should carry its label posterior, as an analytic classifier trained on posteriors equals the true-label one in expectation. The zero-shot prediction is not this posterior, as its class proportions confound name bias with true imbalance and it ignores other images. We propose Posterior Analytic Class-Incremental Learning (PACIL), which corrects the class prior with a prior selection rule and refines each distribution by a fixed-prior von Mises-Fisher mixture on whitened features. These soft labels train the classifier and replace labels to select its fusion weight with zero-shot CLIP by closed-form leave-one-out cross-validation. We prove that posterior targets minimize the mean-square gap to the true-label prediction and that their argmax widens this gap only through ambiguous images. The fusion weight selected with soft labels loses at most their distance to the posteriors in expected correct decisions. PACIL has the best average incremental accuracy among methods trained without labels on all 18 settings.

[CV-60] ReAccess: Look Once Stay Focused with Efficient Test Time Memory

链接: https://arxiv.org/abs/2611.20936
作者: Tianhao Chen,Yilong Fan,Yuheng Wu,Xiaobin Hu,Dongman Lee
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: Project Page: this https URL

点击查看摘要

Abstract:Despite recent advances, vision-language models (VLMs) still struggle with fine-grained visual understanding, often failing to effectively utilize critical evidence already encoded in their visual representations. Existing approaches address this limitation through repeated visual processing or explicit reasoning, introducing substantial computational overhead. In this paper, we propose ReAccess, a lightweight framework that preserves and reuses question-relevant visual evidence through compact associative memory. Constructed during a single visual prefill, the memory supports progressive evidence retrieval that steers both attention queries and hidden representations, enabling effective visual grounding without repeated image encoding or explicit reasoning traces. Extensive experiments demonstrate consistent improvements in both general visual understanding and fine-grained perception, while achieving substantially faster inference than existing reasoning-intensive approaches, demonstrating a promising path to efficient multimodal reasoning.

[CV-61] EviBridge: Encoding Predictive Visual Evidence for Proactive Interaction

链接: https://arxiv.org/abs/2611.20916
作者: Zhichao Huang,Changheng Cai,Song Ren,Zhichen Zhao,Hongsen Liu,Chaoyang Zheng,Wenxi Wu,Yichen Long,Yuan Fang,Luya Wang,Gufeng Liu,Yuchen Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Tool-handover interaction calls for preparation before a response is warranted. The model waits until interaction cues appear, then prepares while assessing whether the accumulating evidence supports a response or calls for cancellation. We introduce EviBridge, a framework that integrates predictive visual representations into a vision-language interaction model for joint state and tool prediction. A trainable evidence interface maps features from a frozen V-JEPA 2 encoder into the language decoder alongside native visual inputs, preserving their temporal context and using only observations available at each decision time. The framework has two instantiations. EviBridge-E distinguishes recent from historical evidence and encodes the timing and spatial location of each observation. EviBridge-I preserves the temporal order of pooled V-JEPA 2 features and interaction turns and generates dedicated interaction tokens. We evaluate on FuGIBench, a streaming-video benchmark for preparation, response, and cancellation in tool-handover assistance. The best-performing variants of EviBridge-E and EviBridge-I achieve Overall scores of 0.6448 and 0.5892, compared with 0.2178 and 0.5327 for their respective task-adapted native baselines. These gains span target handovers, long waits, and cancelled requests, demonstrating the value of predictive visual evidence for interaction decisions.

[CV-62] Palmprint Decomposer: Interpretable and Editable Factorization via Synthetic Learning and Instance-Wise Refinement

链接: https://arxiv.org/abs/2611.20913
作者: Licheng Yan,Yuqi Wang,Jingtao Wang,Bob Zhang,Andrew Beng Jin Teoh,Lu Leng,Shuyi Li,Ziyuan Yang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Palmprint recognition and synthesis methods focus on extracting discriminative textures or generating images with varied textures and appearances. However, they do not generally provide an explicit image-level representation that separates palmprint texture from its underlying base appearance for independent analysis or controlled editing. We introduce Palmprint Decomposer, an interpretable framework that represents a palmprint through independently editable factors and reconstructs the image through a multiplicative composition model. A color map and a light map characterize the base appearance, while a structure map and a global texture strength parameter describe texture modulation. However, accurate reconstruction alone does not guarantee separation of palmprint texture from base appearance, as fine palmprint textures may leak into the estimated base appearance. To mitigate this ambiguity, we train the decomposer exclusively on synthetic data with direct factor supervision, using asymmetric decoding paths and decomposition constraints to suppress texture leakage. Instance-Wise Refinement further improves real-image reconstruction by directly optimizing the predicted factors for each input without relying on the trained network. Experiments across seven datasets demonstrate high reconstruction fidelity, with the extracted structure maps generally achieving verification performance comparable to the original images. Compared with existing methods, our approach achieves better factor separation while preserving more discriminative palmprint texture information.

[CV-63] Can Image Generators Hold a Scene Together When Space Is Broken Apart?

链接: https://arxiv.org/abs/2611.20906
作者: Jhen-Ke Lin,Chung-Chun Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 23 pages

点击查看摘要

Abstract:Image generators are good at drawing scenes in which things that belong together appear next to each other. In almost every image they generate, and in almost every benchmark used to evaluate these images, “belonging to the same whole” and “adjacent to each other” go together. Therefore, we cannot know whether the model truly grasps the scene as a whole, or simply makes each part connect with its adjacent parts. We propose GJBench to disentangle the two. It requires the model to draw a scene as a scattered jigsaw puzzle: an image is cut into interlocking pieces, then moved apart and rotated. Thus, the spatial plausibility of the scene exists only in how these pieces collectively point to a whole. The model may have a reference image, a target layout, both, or neither. We separately evaluate three requirements: whether the scene is broken down into a specified number of separate pieces, whether the pieces interlock, and whether the pieces can reconstruct the reference scene. While every open and closed model we evaluated could draw these scenes well as a whole, none could keep them together as scattered puzzle pieces, and the three requirements categorized failures into different patterns. The models draw separate pieces but miscount them in different ways, and unless their outlines are provided, the pieces rarely interlock, and their content does not reconstruct the reference image. We also found that the target layout helps the puzzle pieces fit together, while the reference image improves neither their count nor their fit. These findings show that keeping a scene continuous across space that the canvas no longer shows remains an open problem for image generators.

[CV-64] SciWorldBench: Do Video World Models Really Understand Scientific Experiments?

链接: https://arxiv.org/abs/2611.20899
作者: Xue Yang,Haonan Cai,Xinyu Hang,Zhimeng Huang,Xuanyi Liu,Chuanmin Jia,Siwei Ma
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Preprint. Benchmark paper with paired scientific experiments for evaluating video world models. 8 pages of main text, 16 pages including figures and references

点击查看摘要

Abstract:While the ability of video generation models to accurately depict scientific phenomena has received growing attention, determining whether these models genuinely understand the underlying principles remains a critical challenge. Existing benchmarks typically evaluate video samples in isolation, making it difficult to expose models that simply fit superficial visual priors. In scientific experiments, a small change in one condition produces a markedly different outcome even when the rest of the setup remains fixed. Capitalizing on this sensitivity, we introduce SciWorldBench, a benchmark comprising paired scientific experiments. Each pair varies a single condition in an otherwise fixed setup, yielding two mutually exclusive expected outcomes. Each generation prompt specifies the condition but omits the expected outcome. We evaluate ten text-and-image-to-video (TI2V) models using an independent vision-language model (VLM) that assesses each video without access to the condition text or the label identifying the correct outcome. A pair is scored as correct only when both videos depict their respective scientifically correct outcomes. Across all ten models, we find that the two videos in an experimental pair frequently depict the same outcome even though the conditions require different results. This recurring failure suggests that current video world models struggle to apply the underlying scientific principles reliably during generation. We will publicly release the benchmark dataset and open-source the evaluation code.

[CV-65] Certified Not Explained: Exact Token Erasure in Gated Vision State Space Models

链接: https://arxiv.org/abs/2611.20897
作者: Sadia Parveen,Tehseen Zia
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 20 pages, 4 figures

点击查看摘要

Abstract:Vision models built on Mamba are often explained through their selective step size Delta, on the assumption that a token the scan does not admit cannot influence the prediction. We test that in a concept bottleneck model for lung-nodule malignancy built on a single gated, bidirectional selective state-space block. The answer becomes an interpretability framework: a guarantee, a procedure that checks it, and a pre-registered audit of what it explains. For a fixed gate pattern, masking the input of the causal convolution of the block and giving closed tokens zero weight at readout makes the concept and task logits exactly invariant to the closed tokens. Gating Delta is neither necessary nor sufficient. On LIDC-IDRI, a path-by-path knockout over 15 cross-validated models agrees with this in every run: the largest logit change is exactly 0 whenever the conditions hold and between 4.1 and 6.7 when they do not. The gated model reaches a pooled out-of-fold AUROC of 0.8425 +/- 0.0085, level with an ungated Mamba bottleneck (0.8428). The guarantee belongs to the masked model, not to the gate as an explanation. Post-hoc rankers, which inherit the guarantee on this model, order tokens better than the gate. Gate maps from a model trained on permuted labels are as close to the real maps as two real-label seeds are to each other. Under retraining, the tokens of the gate carry more task information than random tokens and, on shifted crops, more than central ones, but no more than the tokens chosen by hidden attention. We therefore read the gate as a certified support set produced by mask supervision. Splitting the Oracle Impurity Score into own-concept and cross-concept parts reveals differences the aggregate score misses. Partitioning the readout reduced leakage slightly under an earlier recipe, and the final model sits at, rather than below, the level implied by correlations among the true concepts.

[CV-66] ProSE: Probing Reconstruction Error for Semantic-guided Expert Skipping in MoE-MLLM s

链接: https://arxiv.org/abs/2611.20885
作者: Jiameng Li,Minye Wu,Matthew B. Blaschko
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages

点击查看摘要

Abstract:Mixture-of-Experts Multimodal Large Language Models (MoE-MLLMs) offer scalable inference through sparse expert activation, yet substantial computational redundancy remains among the activated experts. Existing expert skipping methods typically calibrate thresholds based on target skipping ratios or final-output divergence, overlooking the accumulation of skipping-induced error across layers and the varying semantic importance of individual tokens. To address that, we propose ProSE, a training-free, semantic-guided expert skipping framework that probes layer-wise reconstruction error to promote representation consistency compared with default activation. Specifically, we construct query-aware semantic masks to identify salient visual tokens and progressively calibrate modality-specific skipping thresholds for each MoE layer. By minimizing masked reconstruction error under a given budget, ProSE maintains semantic consistency after skipping and allows adaptive expert allocation under limited activation budgets. Extensive experiments on 9 benchmarks demonstrate consistent improvements on average over existing expert skipping methods, particularly under aggressive skipping settings. ProSE maintains over 96% performance on KimiVL-A3B-Inst with 83% expert skipping and accelerates prefilling by 1.7\times .

[CV-67] Multi-Rate Flow Matching for Point Cloud Upsampling

链接: https://arxiv.org/abs/2611.20879
作者: Jingkai Shang,Edgar Torres,Duy Nguyen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Flow-matching point-cloud upsamplers refine sparse inputs into dense point clouds by integrating a learned velocity field with a fixed temporal discretization. However, local reconstruction requirements vary substantially across a shape, while a shared solver budget applies the same temporal resolution everywhere. We propose Multi-Rate Flow Matching (MRFM), a patch-adaptive framework that selects an integration budget for each local patch at inference time while keeping the learned velocity field frozen. MRFM relaxes the discrete budget into a continuous variable, optimizes it online through differentiable soft rollouts using an input-derived geometric fidelity objective, and projects the result back to a discrete budget for the final solver rollout. Experiments on PU1K and PUGAN show strong reconstruction quality, while matched-budget permutation and multi-factor analyses demonstrate that the benefit depends on patch-specific allocation rather than on the step distribution alone. MRFM is accuracy-oriented and intentionally uses additional test-time optimization to obtain more suitable local temporal resolutions.

[CV-68] MVSTF: Multi-View Spatio-Temporal Fusion for LiDAR-Based Moving Object Segmentation

链接: https://arxiv.org/abs/2611.20872
作者: Jiaoju Zhou,Chao Ye,Xuebo Yang,Weiyang Lin
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Autonomous mobile robots confront significant chal lenges in recognizing the motion properties of objects in dy namic environments. However, existing methods are hindered by limitations such as boundary ambiguity, occlusion, viewpoint constraints, and training complexity. In this paper, we propose an innovative end-to-end Multi-View Spatio-Temporal Fusion network, named MVSTF, which utilizes sequential point clouds to generate multi-view images and feature differences, and designs a multi-branch network to extract spatio-temporal and dimen sional features in parallel, and realizes high-precision moving object segmentation. First, sequential point clouds are utilized to generate multi-view images and feature differences. Then, an end-to-end multi-branch network including Dual Down-sampling Blocks (DDB), Spatial Attention Blocks (SAB), Multi-Kernel Convolutional Neural Network blocks (MKCNN), Dual Feature Fusion Blocks (DFFB), and the Point Fusion (PF) module, is designed to extract motion features from feature differences and geometric and semantic features from multi-view images. Finally, an auxiliary semantic head is introduced to further improve the algorithm’s robustness and the model’s interpretability. This work is validated on the SemanticKITTI and KITTI-Road bench marks, achieving state-of-the-art results on the SemanticKITTI MOS validation set and the KITTI-Road validation set, and competitive results on the hidden SemanticKITTI-MOS test set. The code is available at this https URL.

[CV-69] LiDSeg: Linear DehazeFormer Guidance for Efficient Semantic Segmentation in Adverse Weather Conditions

链接: https://arxiv.org/abs/2611.20864
作者: Yobsan Bayisa Leta,Christos Kyrkou,Maria K. Michael,Theocharis Theocharides
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Includes supplementary material

点击查看摘要

Abstract:Robust semantic segmentation in adverse weather is essential for safety-critical autonomous systems. Existing restore-then-segment pipelines introduce latency and prioritize pixel-level restoration over semantic discriminability, while domain adaptation methods often face representation bottlenecks and computational complexity. We introduce LiDSeg, a framework that uses reconstruction-informed guidance to improve segmentation under adverse weather. During training, a lightweight guidance encoder is jointly optimized with the segmentation branch using segmentation and auxiliary reconstruction objectives. At inference, the reconstruction decoder is discarded, while the guidance encoder is retained to extract weather-robust features. A lightweight gated fusion module integrates these features into the segmentation branch, preserving clear-weather accuracy while improving adverse-weather robustness. On Foggy Cityscapes, LiDSeg limits the clear-to-dense-fog mIoU degradation to 2.5-2.8 percentage points, reducing the baseline performance drop by more than fivefold. On the official ACDC test server, LiDSeg-PIDNet achieves 75.07% pooled all-condition mIoU, comparable to supervised HRNet-W48 at 75.0%, while remaining substantially smaller. The gains generalize zero-shot to Foggy Driving and Foggy Zurich without training on real-fog images. LiDSeg adds only 1.9-3.1M parameters across several models, with total model sizes of 5.2-39.5M parameters and no inference-time reconstruction overhead. It delivers real-time inference on the NVIDIA Jetson AGX Orin. Code: this https URL

[CV-70] When Is the Expensive Method Worth It? A Content-Driven Analysis of Classical Bayer Demosaicing

链接: https://arxiv.org/abs/2611.20854
作者: Taranum Wasu,Harsh Kashyap,HS Pannu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Preprint. 14 pages

点击查看摘要

Abstract:Demosaicing reconstructs a full-colour image from a single-sensor colour-filter-array (CFA) mosaic, where only one colour is measured per pixel. Classical methods make different assumptions about images, from simple averaging to directional edge-aware interpolation, and the directional methods cost markedly more compute. We ask a practical question that a per-image quality ranking does not answer: for a given image, is the expensive directional method worth its cost? Using a controlled simulate-reconstruct pipeline on 80 benchmark images from four public datasets (Kodak, McMaster, BSD100, Urban100), six classical methods, and image-content descriptors fixed before any results were seen, we show that the quality gain of the edge-aware method (Menon 2007) over the best cheap linear method is not constant. It ranges from -6.6 to +6.8 dB, and on 31% of images it is zero or negative, meaning a cheap method matches or beats the expensive one. The gain is also predictable from content: it correlates with mean image saturation at r=-0.86 ( p10^-6 ). We then build an interpretable depth-2 decision tree that recommends the best cheap method from image descriptors, beating an always-pick-the-most-common baseline by 20.0 percentage points under leave-one-image-out validation. The analysis uses no deep learning, and every step can be inspected.

[CV-71] Vision-Language Model-Based Phase-OTDR Event Classification

链接: https://arxiv.org/abs/2611.20841
作者: Ahmet Bestami Kose,Mehmet Sergen Catal,Muhammet Cagri Yeke,Kivilcim Yuksel,Abdurrahman Gumus
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 20 pages, 7 Figures, 8 Tables

点击查看摘要

Abstract:Phase-sensitive optical time-domain reflectometry (Phase-OTDR) enables distributed vibration sensing, but reliable event classification has traditionally relied on fully supervised models that require extensive annotated data. This study examines the adaptability of a Vision-Language Model (VLM) for Phase-OTDR event recognition through few-shot learning. Raw samples are converted into Short-Time Fourier Transform (STFT) dB-magnitude spectrogram mosaics, color-mapped and resized to RGB images, and processed by a Contrastive Language-Image Pre-training (CLIP) image encoder; few-shot adaptation optimizes only the Tip-Adapter-F cache keys, leaving both encoders unchanged. Under a leakage-controlled protocol (train-side support/validation pool of 512 images per class, fixed held-out test of 120 per class, no test-set access during selection, 20 distinct support-sampling trials), the selected configuration (the ViT-L-14 Vision Transformer encoder with OpenAI weights, 128 shots per class) reaches 96.81 +/- 0.42% accuracy and a macro-averaged F1 score of 96.79 +/- 0.42% on the held-out test set; a label-efficient 16-shot setting, that is, 96 labeled support images at adaptation time across the six classes, already attains 90.17 +/- 0.81%. Tip-Adapter-F clearly outperforms both zero-shot CLIP and the standard training-free Tip-Adapter, and prompt and backbone analyses show that performance is stable across Vision Transformer encoders and insensitive to prompt wording at the selected operating point. Overall, CLIP visual features carry a transferable structure that a lightweight support cache turns into a label-efficient Phase-OTDR event classifier. The code and the corresponding image dataset are made publicly available on GitHub to support further research: this https URL.

[CV-72] Interpretable deep concepts of organoid morphology

链接: https://arxiv.org/abs/2611.20840
作者: Lion Jasper Gleiter,Tingying Peng
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Image-based classification of organoid morphologies can indicate drug response, viability, or successful differentiation. In recent work, learned representations outperform handcrafted features as the basis for downstream classifiers. However, their black-box nature hides the morphological cues a classifier relies on. Based on the foundation models DINOv3 (general-purpose) and ViTAMIn-O (organoid-specific), we train sparse autoencoders (SAEs) to learn interpretable concepts related to distinct organoid morphologies. We annotate concepts with spatial labels (organoid, organoid wall, background) and compute their dataset specificity and task relevance across five tasks. The spatial labels of DINOv3 concepts yield coarse organoid segmentations on out-of-domain datasets. Filtering concepts by spatial label in the DINOv3-SAE improves downstream AUROC for three of the five tasks. Qualitative analysis of the most relevant concepts further reveals a confounding factor introduced by our preprocessing in one task. Through such insights, interpretable concepts could help improve the reliability of classifiers and guide the future development of organoid representation models. The trained SAEs will be made available upon publication.

[CV-73] An Efficient Training Paradigm for Image Restoration via Improved Fréchet Loss

链接: https://arxiv.org/abs/2611.20805
作者: Shanyuan Jie,Wu Shi,Xinqi Lin,Kainan Yan,Zian Wang,Yingxue Pang,Yan Wang,Shijie Zhao,Chao Dong
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 20 pages, 14 figures, and 7 tables

点击查看摘要

Abstract:Diffusion-based image restoration achieves strong perceptual quality but typically relies on iterative sampling, leading to slow inference. Existing single-step methods often use diffusion distillation or adversarial training, complicating optimization. To simplify training, we propose FDIR, a framework for training single-pass generative restoration models through regression-style supervised optimization using directly computable objectives. To realize this paradigm, FDIR employs the Fréchet distance to match the feature distributions of restored and real images in frozen representation spaces, providing explicit distribution-level supervision. To better accommodate image restoration, we incorporate local patch features to capture fine-grained restoration discrepancies. Moreover, we introduce instance-adaptive statistics by combining the feature moments of each SR-HQ pair with synchronized historical statistics, allowing distribution matching to incorporate input-specific correspondence while retaining distribution-level context. FDIR optimizes only the LoRA parameters of the restoration backbone. Experiments demonstrate strong perceptual restoration quality and efficient training, suggesting a simpler paradigm for generative image restoration.

[CV-74] Correspondence-Free Camera Pose Estimation by Matching Moments

链接: https://arxiv.org/abs/2611.20797
作者: Subhransu S. Bhattacharjee,Dylan Campbell,Rahul Shome
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Robotics (cs.RO); Image and Video Processing (eess.IV); Optimization and Control (math.OC)
备注: 65 pages, including references and appendices; under review

点击查看摘要

Abstract:The blind perspective- n -point (PnP) problem is to estimate the pose of a calibrated camera from 2D and 3D point sets with unknown correspondences. We show that, for complete, noiseless observations, fourteen permutation-invariant image moments up to order four identify a fixed camera pose for almost every 3D point set containing at least twenty points. Our learning-free method, MomentPnP, matches these moments between the observed 2D points and the projected 3D points to initialize a camera pose without correspondences and refines the pose and correspondence weights using expectation–maximization. Across the test sets of three benchmarks, our method reduces median rotation error by 5 – 170\times relative to the best learned baseline, both with and without outliers. We recover accurate poses even with as many outliers as inliers. On dense ETH3D scans, pose accuracy improves as the point sets grow, even as individual correspondences become harder to resolve.

[CV-75] RR: A Thermal Interaction Engine for Robot Simulation and RGBT Data Generation

链接: https://arxiv.org/abs/2611.20792
作者: Biwen Yang,Zhe Cao,Jin Zhang,Haoyang Sun,Zonglu Hu,Jintao Zhao,Zhehao Zhang,Kaizheng Chen,Chengyi Liu,Chengyu Wei,Siyuan Wu,Ruiheng Zhang
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:A robot’s actions change the thermal conditions of its next interaction: heat retained after release, carried with objects or stored beneath the surface can change a later contact’s outcome despite similar current temperatures. Training trajectories must capture how these effects connect earlier actions to later observations and task outcomes. We present TRR, a \emphThermal Robot Renderer that attaches spatial and depth-resolved thermal state to both contact partners, updates local heat exchange through contact, transport and separation, and generates synchronized RGB–thermal (RGBT) observations and event-local task checks on one execution clock. Controlled interventions show why retained heat, spatial transport and internal storage matter for subsequent contact outcomes. From the same state, TRR produces a paired six-task benchmark of 2,224 static–dynamic pairs, that is, 4,448 trajectories with synchronized dual-view observations, states and actions. Visible thermal history raises safe contact-site selection from 75.0% to 91.7% on a dedicated learning corpus, and training on dynamic histories raises simulated transfer completion by 50 percentage points under a new heating procedure with shared feedback execution. We demonstrate RGBT observation adapters on MuJoCo and Isaac Sim, report an engineering inventory of ten prepared robot configurations and six demonstrated camera viewpoints, and describe a developer workflow that carries an existing robot scene to exported RGBT records. The benchmark, the workflow and the platform resources described in Section 6 form the first items of a planned public release. Demonstrations and resource updates are available on the project page.

[CV-76] VA-Assembly: View-Aligned Score Guidance for Training-Free 3D Part Assembly

链接: https://arxiv.org/abs/2611.20788
作者: Yeon-Jae Kim,Sangwon Kang,Byung-Kwon Min
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 16 pages, 9 figures

点击查看摘要

Abstract:Image-to-3D foundation models have improved 3D reconstruction, but typically produce monolithic representations without explicit part-level structure. Part-aware generation and learning-based assembly methods often require part-level supervision, category-specific priors, or paired part-layout training data. We instead consider a practical but underexplored setting in which independently prepared or generated 3D parts have no predefined inter-part relationships, shared coordinate frames, or consistent scales, and must be assembled into a reference-consistent object from a single image. We present VA-Assembly, a training-free framework for 3D part assembly through part-level transformation estimation using view-aligned diffusion guidance. Given a reference image and fixed 3D part geometries, VA-Assembly optimizes each part’s translation, rotation, and scale. We condition a frozen view-conditioned diffusion model on same-view renderings of a reference-derived monolithic 3D representation, producing score gradients geometrically aligned with the part configuration. This guidance provides multi-view structural cues under large initial misalignment. By repurposing pretrained visual-structural priors as transformation-level supervision, VA-Assembly requires no annotated part layouts, predefined part taxonomies, category-specific assembly training, or bounding-box-based layout initialization. Experiments compare VA-Assembly with geometric registration and compositional layout baselines using each baseline’s supported transformation components, showing improved part-wise transformation estimation. Ablation studies show that view-aligned conditioning, rather than diffusion guidance alone, is critical for reliable transformation gradients. Updating only part transformations preserves explicit component separability and supports component-level editing, replacement, and reconfiguration.

[CV-77] PhaseCLIP: Phase-Aware Frequency-Domain Adaptation of Vision-Language Models for Few-Shot Remote-Sensing Classification

链接: https://arxiv.org/abs/2611.20787
作者: Mehmet Sergen Catal,Abdurrahman Gumus,Mehmet Zübeyir Ünlü
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 23 pages, 7 figures, 5 tables. Code will be released at this https URL upon publication

点击查看摘要

Abstract:Remote-sensing (RS) image scene classification aims to assign a semantic category to an overhead image, and in practice must often do so from only a handful of labelled examples, since annotation requires expert interpretation and the resulting labels transfer poorly across sensors and regions. Recent progress has come from adapting vision-language models (VLMs) such as CLIP, through learned prompts or feature adapters. Although effective, these methods construct their representation entirely in the spatial domain. RS scenes are dominated by periodic texture and regular geometry, which the Fourier spectrum describes compactly through an amplitude component, carrying texture and contrast statistics, and a phase component, carrying spatial structure. Existing frequency-domain adaptation of VLMs either exploits amplitude alone or does not separate amplitude from phase, leaving the structural component unused. In this article, we propose PhaseCLIP, which inserts phase-capable spectral modules and a feature-gating module into a frozen CLIP visual encoder and trains only 1.30% of its parameters. Because the amplitude and phase pathways share an identical parameter layout, a single weight set is learned on phase and transferred exactly to amplitude, giving a two-stage scheme that adds no parameters. Experiments on five public RS datasets, EuroSAT, PatternNet, UC Merced, LSRS and RSSCN7, against eight baselines under a matched protocol, show that PhaseCLIP attains 84.28% mean accuracy and is the best method in 23 of 30 dataset-by-shot cells, with phase contributing 1.93 points overall and 4.36 points on the most difficult dataset. Because the modules act on the spectrum of image patches and transformer tokens and leave the text encoder untouched, the approach can be extended to other RS sensors and tasks, and to other domains in which class evidence lies in spatial structure.

[CV-78] hermoSight: From Appearance to Internal Thermofluid Fields via Physics-Guided Neural Reconstruction

链接: https://arxiv.org/abs/2611.20786
作者: Yuhao Fang,Chenglin Lu,Hefei Ling,Chun Lou
类目: Computer Vision and Pattern Recognition (cs.CV); Applied Physics (physics.app-ph)
备注:

点击查看摘要

Abstract:Most neural reconstruction methods focus on reproducing scene appearance, yet inferring physical states from visual observations is essential for machine understanding of the physical world. Reconstructing the internal temperature and soot fields of radiating thermofluids such as flames is challenging yet essential for improving energy efficiency and assessing heat exposure and fire spread. In flames, temperature and soot concentration jointly govern visible emission, and soot self-absorption alters the light that reaches each camera. With limited views, changes in the two fields can offset one another in the recorded RGB images, making temperature and soot difficult to disentangle. ThermoSight, a physics-guided neural framework for scientific imaging, jointly reconstructs continuous three-dimensional temperature and soot volume fraction fields from limited calibrated flame images without field-level ground truth during optimization. A differentiable spectral model links these fields to measured RGB through thermal emission, self-absorption, and calibrated camera responses. Temperature-dependent emission spectra and wavelength-dependent attenuation constrain shared fields across RGB channels, reducing temperature-soot ambiguity. Chromaticity consistency, observation-derived spatial constraints, and soot regularization complement intensity fitting. In a synthetic ablation, wavelength-resolved attenuation reduces temperature RMSE in the hot, sooty region from 234 to 112 K relative to a three-channel approximation, with nearly identical RGB PSNR (65.23 vs. 65.24 dB). Measured data yield coherent fields and competitive RGB reconstructions. These results demonstrate that physics-guided neural fields recover interpretable internal physical quantities beyond visually consistent appearance.

[CV-79] Deep Learning for Satellite-Based Mountain Hazard Screening: A Retrospective Failure Analysis on the 2021 Chamoli Rock-Ice Avalanche

链接: https://arxiv.org/abs/2611.20769
作者: Snigdhajyoti Rakshit
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 27 pages, 10 figures, 8 tables

点击查看摘要

Abstract:The approach utilizes 10 bands from Sentinel-2 imagery, along with terrain information, in a deep learning pipeline to classify potential hazard sources in any given mountainous region. The performance of our model has been evaluated against the data collected during the rock-ice avalanche event in the Chamoli region in the Indian Himalayas on 7th February 2021. U-Net performs segmentation of the detected land surface in the satellite image, detects changes in the identify candidate hazard zones, and then the following steps are performed to calculate the potential energy using the Monte Carlo sampling for steepest descent (D8) flow paths, finding the exposed structures in the downstream locations, and ranking the zones depending on the risk score. In comparison to the simple spectral rule, our approach achieved the higher macro F1 score of 0.648 (as compared to 0.519). But the F1 score in case of rock and soil was low, being in the range of 0.3s. In case of optical change detection in winter, it’s only that northeast-facing slope which couldn’t be detected, as it was under 73% of shade prior to the incident. The parameters have been set for this particular scene alone and not independently; hence, this is a retrospective test, and this test locates a failure which has occurred earlier. The primary contribution of this paper is locating the pipeline leak.

[CV-80] UNITE: Consolidating Embodied VLM Specialists via Shared-Anchor One-Bit Factorization

链接: https://arxiv.org/abs/2611.20767
作者: Tao Tao,Anke Tang,Enneng Yang,Zhi Hou,Xiaogang Wang,Dacheng Tao
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Embodied capabilities such as visual grounding, spatial reasoning, navigation, and driving understanding are often developed in separate vision-language specialists fine-tuned from a common pretrained this http URL the specialists separately multiplies storage, whereas merging them into one dense model forces conflicting task updates to share parameters and loses part of each specialist’s this http URL introduce UNITE, a framework that consolidates such specialists into a compact shared-backbone mixture of experts. UNITE keeps the feed-forward block of each specialist as a routable expert and represents the expert bank through a shared-anchor one-bit factorization: a shared floating-point anchor plus expert-specific, row-scaled binary this http URL anchor is optimized jointly with the scales and signs through closed-form alternating updates, using only source weights and no calibration this http URL adaptation then adjusts token-level expert selection with the backbone and experts frozen, updating fewer than 0.6 M parameters ( 0.01% of the 8.8 B active parameters). Built from four Qwen3-VL-8B specialists, the complete UNITE model occupies 20.3 GB, 1.16\times the storage of one dense specialist and 59.6% below the uncompressed MoE. Across 22 benchmarks and R2R Val-Unseen, UNITE matches six static merging methods in mean retention of best-specialist performance, leads all of them on Scan2Cap, ScanQA, and all four navigation metrics, and surpasses every source specialist on all five 3D scene understanding benchmarks (up to +19.1 on Scan2Cap) and on R2R navigation. Relative to the uncompressed MoE, factorization is nearly lossless on 21 of 22 benchmarks; the exception is VSI-Bench, where frame-count experiments suggest impaired use of cross-frame this http URL thus consolidates complementary embodied capabilities near single-model storage.

[CV-81] A Seismic Foundation Model for Dense Feature Extraction

链接: https://arxiv.org/abs/2611.20745
作者: Zhixiang Guo,Xinming Wu,Yimin Dou,Hui Gao,Siyao Zhao,He Huang,Wei Liu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Deep-learning-based seismic interpretation is constrained by scarce high-quality annotations, and most existing models are trained separately for individual tasks. We present SeisDFE, a 7B-parameter seismic foundation model for dense feature extraction, pretrained on 5.34 million 2D images derived from more than 500 seismic volumes. A three-stage pretraining strategy, combining discriminative self-supervised learning, Gram feature anchoring, and multi-resolution adaptation, improves feature discriminability while preserving spatial coherence. With its encoder frozen, SeisDFE supports seismic facies classification, salt detection, fault detection, and relative geologic time estimation. It outperforms existing seismic foundation models on all four tasks and a natural-image-pretrained model on all but fault detection, with the largest gains under scarce supervision: using only two annotated sections of the F3 dataset, it improves whole-volume mIoU by 12-19 percentage points over the baselines. The learned features remain effective with task-specific decoders and extend to synthetic-to-field transfer and sparse 3D geobody interpretation. These results show that large-scale seismic pretraining organizes geological information into dense representations that are directly reusable across interpretation tasks.

[CV-82] Robotic Reward Models in the Foundation-Model Era

链接: https://arxiv.org/abs/2611.20741
作者: Guoqiang Liang,Yiqi Lin,Shangchen Zhou,Wenzheng Zeng,Yanzhe Chen,Zechen Bai,Mike Zheng Shou
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: Project page: this https URL

点击查看摘要

Abstract:Robotic reward models (RRMs) use foundation models to replace per-task reward engineering with reusable, task-conditioned evaluation. Yet under one name, their outputs answer different questions, from whether a task succeeded to which execution is better. This survey organizes 114 methods into five semantic families by the question their native outputs answer: Outcome, Absolute Progress, Relative, Temporal Value, and Hybrid. Within this taxonomy, we review how model design determines what an evaluator can judge, how raw experience becomes supervision, and how existing evaluation practice tests the result. The review exposes a gap: agreement with demonstration order and downstream policy gains do not show whether a reward tracks task fulfillment. We therefore introduce RRM-Bench-1K, 1,117 real-robot videos of closed-loop executions with task-state and camera-evidence labels, and four contrast suites that separate the credited condition from its usual cues. A unified study of 27 evaluators shows that the strongest final scores rank whole executions well enough for selection and filtering. Dense credit, however, follows the progression of an execution more closely than its task state, and no evaluator signals when the decisive view is missing. How much fidelity a reward model needs therefore depends on the decision it drives. We outline directions from real-failure supervision to metrics that separate task state from chronology, and open-source RRM-Hub, a toolbox in which 14 existing RRMs can be readily applied and new evaluators tested on the same capabilities.

[CV-83] CLIFT: Contrastive LiDAR Instance Features for 3D Pedestrian Tracking

链接: https://arxiv.org/abs/2611.20737
作者: German Devchich,Timur Akhtyamov,Denis Fatykhoph,Gonzalo Ferrer
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:A social robot has to keep track of the people around it over time to plan its motion, a task handled by 3D multi-object tracking. For the LiDAR modality, trackers usually sit on top of a detector and link its detections across frames with a separate motion model or re-identification network. We instead show how a 3D LiDAR detector itself can be turned into a tracker by adding an instance-level contrastive objective on its intermediate voxel features. Each embedding is read from the voxels inside a detected box, and boxes are matched across frames by cosine similarity and the Hungarian algorithm, without a motion model. The detector backbone is fine-tuned together with this objective, which learns identity-aware features while leaving detection accuracy essen- tially intact. Because the method uses only the voxel features that Li- DAR detectors already produce, it is backbone-agnostic. On the JRDB 3D tracking leaderboard our LiDAR-only tracker ranks first on HOTA (38.58) and IDF1 (37.99), 13.04 and 9.30 points above the next pub- lished method, and second on MOTA and OSPA. Code is available at this https URL.

[CV-84] Selective Sensor Acquisition for Multimodal Place Recognition

链接: https://arxiv.org/abs/2611.20730
作者: Margaux Edwards,Tobias Fischer,Michael Milford
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 6 figures

点击查看摘要

Abstract:Multimodal Place Recognition (PR) can improve robustness, but most systems rigidly read and process every sensor for every query, even when the primary sensor is sufficient, with implications for energy consumption, second sensor update rate, and domain-specific considerations such as passive stealth. We introduce selective, controllable sensor acquisition, where a user specifies a desired second sensor usage rate and the system acquires the additional sensor only when the primary sensor’s estimated performance is sufficiently poor. An online threshold adjustment achieves the requested usage rate by adapting to recent primary sensor performance, directing additional sensing towards queries where fusion is most likely to help without calibration specific to sensor modality, PR method, or environment. The approach requires only a PR method that produces a match score and applies to both single-frame and sequence-based PR. Our system, across multiple datasets, PR descriptors, and sensor combinations, recovers 55% of the available PR performance gain while using the second sensor on only 25% of queries, and 81% while using it on 50%. We further analyse alternative acquisition decision mechanisms and usage-rate control methods, and examine performance when the second sensor is stronger than, comparable to, or weaker than the primary sensor.

[CV-85] QASK-Edit: Adaptive Sparse Attention Across Queries and Key Sources for Efficient Video Editing

链接: https://arxiv.org/abs/2611.20704
作者: Hengrui Liu,Wenbo Xu,Ziqi Sheng,Wei lu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Diffusion Transformers (DiTs) have achieved strong performance in generation based video editing, but suffer from high inference latency due to the long token sequences formed by conditioning inputs and generated videos. Although training-free sparse attention reduces this cost, existing methods typically apply a fixed retention ratio or a single threshold shared across all tokens. In this work, we view video editing as conditional generation and identify a key insight: attention demand varies significantly both across different key sources (e.g., source videos, reference images, audio, text, and generated videos) and across queries within each source. Motivated by this, we propose QASK-Edit (Query-Adaptive Sparsity over Key Sources for Video Editing), a training-free sparse attention framework that allocates computation independently for each query and key source. Specifically, for each query, QASK-Edit considers both the attention mass and approximation risk of each block. It then evaluates candidate routing thresholds against each key source’s risk budget, choosing exact computation or approximations to dynamically adjust sparsity for each query. Extensive experiments on three video editing models demonstrate that QASK-Edit achieves up to 1.46\times end-to-end speedup over dense attention pipeline while preserving both editing quality and unedited region fidelity. Furthermore, QASK-Edit generalizes directly to Reference-to-Video generation, delivering up to 1.36\times speedup while maintaining generation quality.

[CV-86] Efficient Multi-View Generation with Iterative Global-Local Proxy Attention

链接: https://arxiv.org/abs/2611.20703
作者: Wei Tan,Kunming Luo,Hongyu Yan,Yu Zhang,Zihao Zhang,Xiatian Zhu,Lei Xie
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 6 figures, 5 tables. Supplementary material is available as an ancillary file

点击查看摘要

Abstract:Generating consistent multi-view RGB images and normal maps from a single input image provides complementary appearance and geometric cues for 3D reconstruction. However, maintaining cross-view consistency requires information exchange across viewpoints, and dense cross-view attention becomes computationally expensive as the number of views and spatial resolution increase. In this paper, we propose Global-Local Proxy Attention (GLPA) to reduce cross-view communication costs while preserving fine-grained within-view processing. Specifically, GLPA retains full-resolution interactions within each view and uses compact global proxies for cross-view communication. To further leverage the exchanged information, GLPA incorporates iterative proxy refinement, which rebuilds the proxies from updated features and reuses the same attention parameters for subsequent interactions within a single denoising step. We evaluate GLPA on two structurally different diffusion backbones, FLUX and Stable Diffusion. Experiments on FLUX show that GLPA achieves a 2.65\times speedup in attention computation per denoising step over dense attention, while improving reconstruction IoU from 0.6849 to 0.7019 and reducing NVS LPIPS from 0.086 to 0.076. Results on Stable Diffusion further demonstrate the effectiveness of GLPA, with a 1.80\times attention speedup and improvements across all five quality metrics.

[CV-87] STARS: Stage-Adaptive Responsibility Separation for Shadow Removal in Remote Sensing Images

链接: https://arxiv.org/abs/2611.20698
作者: Gaoliang Song,Chengjie Li,Pengju Liu,Chenchen Chang,Jinhao You,Hongzhi Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 8 figures, 4 tables

点击查看摘要

Abstract:Shadows are a ubiquitous radiative attenuation phenomenon in remote sensing images and can severely affect high-resolution remote sensing interpretation tasks. In shadow removal, although introducing a mask prior provides explicit positional information for the restoration network, it also causes non-shadow regions to be modulated by the shadow prior, producing a systematic brightness shift; notably, this phenomenon is independent of whether the mask is accurate. In remote sensing images, non-shadow regions usually occupy a larger proportion, so a small shift over these regions incurs a much higher cost on whole-image metrics than an equal gain in shadow regions. Adopting a task-decoupling perspective, we therefore decompose shadow removal into two tasks, shadow-region restoration and non-shadow-region identity, within a cascaded architecture, and handle them sequentially. Based on this insight, we propose STARS, a cascaded shadow removal framework that separates regional responsibilities across stages. The method retains the advantages of the shadow mask while resolving the degradation of non-shadow regions, without requiring a complex parallel dual-branch structure. Meanwhile, to introduce large-model prior knowledge into remote sensing tasks, we further design a Feature Selection and Fusion Adapter (FSFA). We conduct extensive experiments on three benchmark datasets, UAV-SC+, ISTD+ and WSRD+ 2026, demonstrating the superiority of STARS. The framework shows that decoupling regional responsibilities can resolve the intrinsic contradiction of mask priors, providing a new direction for image restoration with region heterogeneity.

[CV-88] VIVID: A Visual Harness for Agent ic Video Editing

链接: https://arxiv.org/abs/2611.20688
作者: Wenhan Yu,Zhaoxi Zhang,Zhilong Tan,Weikang Li,Lei Sha
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)
备注:

点击查看摘要

Abstract:Video editing requires agents to make editing decisions grounded in audiovisual observations and precise temporal localization, and to evaluate their effects in rendered outputs. General-purpose coding agents possess strong comprehension, reasoning, and code execution capabilities, while media interaction and evidence management mechanisms tailored to video tasks can further unlock their editing potential. We introduce VIVID, a visual harness that supports multiple native coding agents in video post-production without additional model training. VIVID provides tools for on-demand video inspection, audio analysis, and transcription, and preserves observations alongside their corresponding media versions and temporal coordinates. Agents can actively retrieve these observations as they reason, perform edits using their native coding and shell tools, and inspect rendered outputs to guide subsequent revisions. On AgenticVBench, comprising 100 tasks across Assembly, Repair, Sequencing, and Repurpose, Codex with VIVID achieves an average score of 58.6%, an improvement of 20.2 percentage points over native Codex. The same harness also integrates with Claude Code and OpenCode, achieving average scores of 55.1% and 37.1%, respectively, both substantially exceeding their native baselines. These results highlight VIVID’s potential as a reusable visual harness for extending the capabilities of general-purpose coding agents to video editing.

[CV-89] Room for Supervision: Training-Free Dataset Distillation for Object Detection

链接: https://arxiv.org/abs/2611.20680
作者: Ziang Gan,Qi Zhu,Libao Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Dataset distillation (DD) reduces storage and training costs by compressing large datasets into small collections. Most existing methods focus on image classification, where each image typically has a single category label. In contrast, object detection requires accurate category and bounding box annotations for multiple objects, together with visual content consistent with those annotations. Existing detection distillation pipelines rely on learned models and can alter the spatial relationships between objects and their backgrounds. We introduce GRASP, which packs dense native supervision into limited pixels without network training or inference. GRASP jointly searches over image selections and spatial layouts using convex allocation subproblems that penalize insufficient object resolution, excessive image compression, and padding that source context cannot fill. During composition, each selected image is cropped around all annotated objects and expanded into surrounding context without reducing object scale. The resized regions are then combined into distilled images, preserving the spatial layout within each region and retaining all original class labels and bounding box annotations. We evaluate GRASP on MS COCO and PASCAL VOC with compression ratios ranging from 0.25% to 2%. GRASP constructs the 1% MS COCO dataset in under two minutes. Under a common training protocol using hard labels, GRASP achieves the state of the art with both CNN and ViT backbones. Code is available at this https URL

[CV-90] Competent Autonomous Driving Through Fine Tuning Vision Language Action Models with Symbolic Traffic Rules

链接: https://arxiv.org/abs/2611.20648
作者: Barkin Dagda,Vinal Asodia,Iman Sharifi,Peng Wei,Saber Fallah
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Vision-language-action (VLA) models offer a unified approach to perception, reasoning, and control for autonomous driving, but accurate action prediction alone does not establish compliance with traffic regulations. This paper presents a rule-guided VLA framework that incorporates symbolic traffic-rule feedback into policy optimisation. The framework combines supervised fine-tuning on expert demonstrations with reinforcement learning (RL) refinement using a reward that encourages encoded-rule compliance while penalising deviations from expert actions. Selected provisions of the United Kingdom Highway Code are formalised as executable first-order logic rules using PyReason, covering speed adaptation, following distance, and road-user priority. A frozen vision-language model extracts scene predicates, which are combined with vehicle-state information to assess the projected consequences of candidate control actions. Offline evaluation on held-out data collected in the CARLA simulator compares three group-based RL methods. The results show improvements in both action prediction accuracy and encoded-rule compliance over the supervised baseline. The best-performing variant increases the overall rule pass rate from 92.9% to 97.6%, with zero observed violations of the encoded road-user-priority check in the evaluated samples. These findings support symbolic rule feedback as a means of improving VLA policy alignment with explicit traffic-rule criteria.

[CV-91] Defect-Aware Masked Autoencoders for Unsupervised Clustering of Steel Surface Defects

链接: https://arxiv.org/abs/2611.20645
作者: Shrey Patel
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted at IVCNZ 2026

点击查看摘要

Abstract:Reliable recognition of steel surface defects is essential for automated quality inspection, but supervised methods require manually annotated images that are costly to produce. This creates a need for methods that can learn useful defect representations directly from unlabelled surface images. This work presents Defect aware MAE, a label free masked autoencoder for clustering steel surface defects. Each image is divided into 196 patches, with 147 masked and 49 processed by a Vision Transformer encoder. Texture energy based masking preferentially hides structurally informative patches, while a shallow decoder reconstructs HOG descriptors that capture local texture and edge orientation. Strong augmentation and an intermediate layer readout are used to form the final image features. Pretraining uses no class labels, bounding boxes, detection branch, or contrastive loss. On the held out NEU DET split, UMAP followed by Ward clustering achieves 98.6% Hungarian matched accuracy, 0.968 NMI, 0.967 ARI, and 0.838 silhouette score. Under an identical encoder and training schedule, it exceeds a SimCLR contrastive baseline in accuracy, NMI, ARI and silhouette scores across 5 seeds.

[CV-92] Counterfactual Feature Steering for Adjacency-Preserving Identity Unlearning

链接: https://arxiv.org/abs/2611.20637
作者: FNU Shivam,Layth Sliman,Ramana Reddy,Donald Adjeroh,Binod Bhattarai,Prashnna Gyawali
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Identity unlearning in text-to-image diffusion models is increasingly important for privacy, safety, and ethical deployment, as it aims to remove the model’s ability to generate specific individuals. In this work, we study identity unlearning with \emphadjacency preservation, where the goal is to remove a target individual’s identity while preserving related non-target identities. We propose Counterfactual Feature Steering (CFS), a method requiring one-time SAE training that performs bounded inference-time concept editing in the latent space of a sparse autoencoder attached to the text encoder. CFS derives a target direction by contrasting forget and retain distributions, then applies a per-token projection proportional to each token’s current alignment with this direction, yielding counterfactual representations that suppress the target while limiting collateral damage. To evaluate adjacency preservation, we introduce the Adjacency-Weighted Unlearning Score (AWUS), which measures the tradeoff between target erasure and collateral damage, and extend it to face-specific identity similarity using ArcFace. Across 4,000 prompts derived from Celeb-100, spanning 100 celebrity identities, CFS achieves the strongest evaluated tradeoff between identity erasure of target identities and preservation of adjacent and generic concepts, while remaining stable across intervention strengths where other steering-based methods degrade. Independent ArcFace evaluation and a 10,000-caption COCO study further examine identity and general preservation. Quantitative nudity-suppression benchmarks and qualitative artistic-style examples provide evidence of transfer beyond identities on SD~1.4.

[CV-93] Instructions Before Actions: Prospective and Retrospective Language Supervision for Driving Vision-Language-Action Data

链接: https://arxiv.org/abs/2611.20631
作者: Angel Martinez-Sanchez,Ross Greer
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Driving vision-language-action (VLA) models can accept navigation instructions, yet predicting a driver’s trajectory does not establish that a model follows the instruction. We study how driving-language supervision is collected and evaluated, distinguishing prospective instructions issued before an action from retrospective labels written after observing it. We develop ADVLAT-Engine to synchronize spoken instructions, road-scene video, and vehicle trajectories, and compare three navigation apps on 48 replayed routes. On 87 maneuvers announced by both Google Maps and a passenger, the passenger uses more scene-dependent language. We evaluate four open driving VLAs before and after fine-tuning on prospective navigation-app drives. Fine-tuning reduces passenger-drive trajectory error by 23-59%, but instruction effects remain inconsistent. At real intersections, counterfactual instructions reveal stronger responsiveness: fine-tuned Alpamayo~2 Super gains 31 and 18 accuracy points at 15 and 5~m, while Qwen-Drive follows alternative maneuvers more reliably with a discrete command than with a sentence. We additionally collected retrospective annotations from four people for the driving recordings and describe controlled comparisons of annotation content, timing, and downstream learning; those analyses are pending. Our findings separate maneuver prediction from instruction-following and establish a framework for comparing prospective and retrospective supervision.

[CV-94] Event-Guided Stereo Video Super-Resolution

链接: https://arxiv.org/abs/2611.20629
作者: Hanyu Xuan,Youtao Yao,Cheng Zhao,Kun Li,Xin Du,Fei Wang,Zhiliang Wu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Stereo Video Super-Resolution (SVSR) aims to reconstruct high-resolution stereo videos by exploiting temporal cues across consecutive frames and complementary spatial cues between stereo views. However, existing methods struggle to recover high-frequency texture details under rapid motion and complex illumination, compromising reconstruction quality and stereo consistency. Event cameras provide valuable motion and texture cues with high temporal resolution and wide dynamic range, yet their potential for SVSR remains underexplored. To bridge this gap, we present the first investigation of event-guided SVSR and propose a novel framework. Specifically, we develop an Event-Guided Texture Enhancement (EGTE) module that combines bidirectional temporal feature propagation with event-guided texture refinement to restore high-frequency details in complex motion regions. Subsequently, we introduce a Cross-View Texture Interaction (CVTI) module to model cross-view spatial correspondences and aggregate cross-view complementary texture information. With this design, our method effectively integrates event-driven motion and texture information, inter-frame temporal dependencies, and cross-view spatial relationships to recover motion-affected details and maintain stereo consistency. Extensive experiments on both real-world and synthetic datasets demonstrate that our method consistently outperforms existing super-resolution methods in reconstruction quality and stereo consistency. Project Page: this https URL

[CV-95] PixelBench: Evaluating Language Models for Pixel-Art Creation

链接: https://arxiv.org/abs/2611.20628
作者: Mutian Hong,Jianxuan Yao,Mingkun Yue,Changkai Li,Hower Yang,Haibin Ling
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 38 pages, 12 figures, 11 tables

点击查看摘要

Abstract:Pixel art is widely used in games and digital content creation, requiring readable forms, coherent spatial relationships, and harmonious colors under limited resolution and palette constraints. Language models show growing promise in visual creation, yet their pixel art creation capabilities lack systematic evaluation. We introduce PixelBench, formalizing text-to-image generation and instruction-guided editing for pixel art and providing a benchmark for these tasks. We further present a creation environment in which language model agents invoke drawing tools through code to generate and edit pixel art. We assess the quality of generated artwork through human evaluation and employ a multimodal large language model as a judge, validating the agreement between its automated assessments and human preferences. We compare language models and configurations across diverse subjects and canvas sizes. In our experiments, GPT-6 Astra achieves both the highest average judge score and the best average human rank, highlighting the potential of general-purpose language models for pixel art creation. We release the tasks, all artworks, the human rankings, and the judge protocol.

[CV-96] ASCG: Adaptive Spatial Classifier Guidance for Surgical Concept Suppression in Diffusion Models

链接: https://arxiv.org/abs/2611.20625
作者: Younghwan Kil,Joonhyeong Park,Giung Nam,Juho Lee
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 9 pages, including references and supplementary material

点击查看摘要

Abstract:While text-to-image (T2I) diffusion models achieve remarkable quality, they risk producing harmful content. Existing safety mechanisms often rely on indiscriminate, global interventions that induce unintended semantic drift and degrade image quality. To address this, we propose Adaptive Spatial Classifier Guidance (ASCG), a necessity-driven framework that employs probabilistic monitoring to activate guidance selectively upon harm detection. By leveraging Grad-CAM for localized suppression, ASCG effectively removes harmful concepts while preserving superior semantic fidelity compared to conventional methods. Furthermore, we overcome the limitations of traditional metrics like NudeNet—which fails to assess semantic integrity—by introducing a VLM-based evaluation framework to provide a fine-grained, robust analysis of both safety and context preservation. Our results demonstrate that ASCG not only balances safety and fidelity but also exhibits remarkable plug-and-play synergy with existing frameworks to establish state-of-the-art performance.

[CV-97] Example-Based Spatial Guidance for Training-Free Concept Erasure in Diffusion Models NEURIPS2026

链接: https://arxiv.org/abs/2611.20623
作者: Younghwan Kil,Giung Nam,Joonhyeong Park,Jinwoo Shin,Juho Lee
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted at NeurIPS 2026 (poster). 51 pages, including references, checklist, and appendix

点击查看摘要

Abstract:Text-to-image (T2I) diffusion models often synthesize policy-violating content, necessitating robust safeguards and rigorous evaluation frameworks. However, current approaches remain largely monolithic in both execution and evaluation, limited to coarse-grained categorization within broad unsafe concepts. Such approaches fail to account for the specific constituent factors of a violation, resulting in weak learning signals and allowing model collapse to bypass safety checks. To overcome these limitations, we propose Example-Based Spatial Guidance (EBSG), a training-free method that utilizes user-editable exemplar packs to provide granular spatial guidance for precise concept steering. Unlike existing methods that coarsely steer the entire image away from a broad concept, EBSG decomposes these categories into specific sub-concepts via exemplar text-image pairs, providing explicit, localized steering signals for more robust safety control. Furthermore, we introduce a vision language model (VLM)-based evaluation protocol that provides a fine-grained assessment, avoiding the pitfall of conventional binary evaluators that permit model collapse to bypass safety checks. Empirically, EBSG achieves state-of-the-art erasure performance across diverse safety datasets and concept categories, spanning four nudity benchmarks, seven I2P harmful-concept slices, and four MJA metaphor categories, while supporting multi-concept removal and transferring to a modern backbone, SD3.

[CV-98] From Joint Risk to Decision Invariance: Decision-Preserving Regularization for Exemplar-Free Class-Incremental Learning

链接: https://arxiv.org/abs/2611.20619
作者: Qi Zhu,Ziang Gan,Wanting Zhang,Libao Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Exemplar-free continual learning seeks to minimize joint risk over all seen tasks, yet historical data are unavailable for direct optimization. We observe that, for classification, preserving the model’s decisions on previously learned classes helps retain old knowledge. This motivates Decision-Preserving Regularization (DPR). Penalizing changes in all pairwise old-class logit margins reduces algebraically to a centered relative-logit penalty. We further apply spectral whitening to remove the classifier’s spectral scale bias. Because the global classifier supplies the decision constraints for the next stage, it must remain calibrated across all seen classes despite the drift of stored old-class statistics during representation updates. We therefore fit a closed-form affine map from paired current-task features to transport old-class Gaussian means and covariances into the updated feature space, then calibrate the global classifier using these distributions and current-task features. Across CIFAR-100, Tiny-ImageNet, ImageNet-Subset, and ImageNet-1K under cold-start protocols, comparisons with state-of-the-art methods show that DPR achieves higher accuracy in most settings while remaining highly efficient. Code is available at this https URL.

[CV-99] SCoRe: Support-Calibrated Role Fusion for Training-Free Cross-Domain Few-Shot Segmentation

链接: https://arxiv.org/abs/2611.20582
作者: Abhishek Kathpal,Adarsh Jagan Sathyamoorthy,Utsav Patel
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:We present SCoRe (Support-Calibrated Role fusion), a framework to utilize two complementary vision foundation model-based (VFM) segmenters for training-free cross-domain few-shot segmentation (CD-FSS). Our approach is based on the observation that current VFMs for CD-FSS possess capabilities where one model locates the correct object, while another model computes the better contour. Our approach uses: 1. a method to identify a VFM segmenter to lead Few-Shot Segmentation (FSS) task for a given concept, 2. a novel deterministic boundary correction method using bands of near- and far-pixels at the lead segmenter’s predicted boundary of an object, and 3. a region correction mechanism for the other segmenter to conservatively prune components, fill holes, and bridge short gaps on the lead segmenter’s predicted masks to improve accuracy. We evaluate SCoRe using RSRM and FSS-SAM3 across 3,314 matched five-shot episodes from 10 datasets. SCoRe achieves 72.738% foreground mIoU averaged across datasets, a gain of 5.210 percentage points over RSRM (paired 95% CI [4.618, 5.972]), with per-dataset gains up to 8.461 points over the strongest standalone method and the best point estimate on 7 of 10 datasets.

[CV-100] PerturbJEPA: Joint-Embedding Predictive Learning of Cellular Perturbation Responses

链接: https://arxiv.org/abs/2611.20575
作者: Jiangshuan Pang,Youzhe He,Yiting Deng,Wangyang Tang,Huinan Xu,Qi Lu,Shiping Liu,Zhenkun Zhuang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 23 pages, 6 figures, 14 tables. Code: this https URL

点击查看摘要

Abstract:Identifying the drug and dose that produced a cellular response requires preserving treatment effects within each cellular background. A joint-embedding predictor can match its targets while ignoring the intervention if its encoder preserves differences across backgrounds but erases differences between treated and control populations within each one. We call this failure interventional contrast collapse. PerturbJEPA encodes the difference in mean gene expression between treated and control populations as an effect vector. It aligns the displacement between their latent centroids with this vector and regularizes the vectors to preserve variation across population pairs. We show that the measured displacements cannot all coincide within a batch when effect-vector variation exceeds the root-mean-square alignment error. Trained with separate drug and dose ranking objectives, the model predicts candidate responses from a common control population and matches them to the observed response without cell pairing or expression reconstruction. We evaluate 3,592 Tahoe-100M queries on held-out combinations of known interventions and cellular backgrounds, each with over 700 drug-dose candidates in a fixed, background-specific catalog. PerturbJEPA achieves 91.46% macro joint drug-dose Recall@1, exceeding STATE by 3.40 percentage points (95% drug-cluster bootstrap interval of 1.72 to 5.07 points). Controlled diagnostics show that treatment-control differences can shrink even when representations remain globally dispersed. With the drug ranking objective fixed, adding effect-vector regularization to alignment raises joint Recall@1 from 18.57% to 87.47%.

[CV-101] OrthoMark: Place Recognition with Point-Anchored Orthographic Height Mark

链接: https://arxiv.org/abs/2611.20570
作者: Haichuan Li
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:LiDAR place recognition provides the loop closures of Simultaneous Localization and Mapping(SLAM). Perspective from Z-axis, a gravity-aligned scan is an orthographic map of the scene, and a revisit from another position and heading moves this map rigidly in the plane, a motion that a polar Fourier spectrum ignores and phase correlation recovers. Along the vertical, the heights of the returns describe the structures, but which heights are observed depends on the range, because the sensor’s vertical field of view cuts nearby structures low; height thus misleads a global key, yet is decisive once two scans are aligned and compared cell by cell where they overlap. We present \method, a training-free method built on one principle: two LiDAR scans can only be compared through features computed from the same geometric surfaces in both scans. \method retrieves candidates with a translation- and rotation-invariant polar spectrum of the bird’s-eye view (BEV), 1.9,KB per keyframe, and verifies each candidate by estimating its planar pose through phase correlation and scoring it with \emphpoint-anchored height columns: every BEV cell keeps a 16-bit mask of the heights of its own returns; within a session, the score is weighted by the expected overlap of the two scans’ sensing ranges, computed from their distance. We show that this vertical link pays off in verification but hurts in retrieval, where what a height slice contains depends on the range to the structure. For single robot, \method has the highest AP of the training-free methods without 3D registration on the KITTI development sequences; across robots, on all multi-robot sequences, it matches or exceeds RING++ at 1/350 of its descriptor size and closes more inter-robot loops than the learned BEVPlace++ on all tests.

[CV-102] Vision-Language Grounded Semantic Navigation for Autonomous Mobile Robots

链接: https://arxiv.org/abs/2611.20558
作者: Ravindran Vasoor Saravanan,Chittin Varunav Ravi,Madhumita Singh,Thirulok Sundar Mohan Rasu
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: 7 pages, 4 figures

点击查看摘要

Abstract:Language-guided navigation enables mobile robots to interpret natural language commands and navigate toward user-specified targets in unstructured environments. However, integrating open-vocabulary visual perception with classical robotics pipelines for mapping and navigation remains a challenging problem. In this work, we present a complete ROS2-based system that combines frontier-based exploration, open-vocabulary object detection using Grounding DINO, depth-based 3D localization, and semantic mapping to enable end-to-end vision-language grounded navigation in simulation. The system accepts natural language commands such as ``go to the yellow box,‘’ detects candidate objects from RGB images using text prompts, estimates their positions in the world frame, and generates navigation goals executed by the Nav2 stack. In a Gazebo-based indoor environment with a TurtleBot3 robot, the system autonomously explored the room, detected three of the five target objects, and successfully executed both natural language navigation commands tested, without task-specific training. The results indicate that open-vocabulary grounding models can be integrated into classical navigation pipelines, though performance is limited by detection uncertainty, depth-based localization error, and incomplete exploration coverage. These findings highlight both the feasibility and current limitations of deploying vision-language grounded navigation systems using lightweight, CPU-only architectures.

[CV-103] AsymEEF: Bimanual Manipulation with Asymmetric End Effectors

链接: https://arxiv.org/abs/2611.20557
作者: Xiaokang Sun,Ningyue Wang,Kunqi Xu,Guangqi Jiang,Jianglong Ye,Yong Wu,Xueyan Zou
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Teleoperated demonstrations are costly to collect and rarely reusable across end-effectors with different kinematics and action spaces. This burden grows in bimanual manipulation, where each end-effector pair may require new data. We present AsymEEF, a framework for executable cross-end-effector demonstration generation and unified bimanual policy learning. Given a small source dataset, AsymEEF transfers demonstrations to heterogeneous combinations of two-, three-, and five-fingered end-effectors without additional teleoperation. Kinematic teacher anchors provide geometric supervision, while sparse real-world X-Calib anchors correct discrepancies in grasp execution. A shared encoder maps wrist poses and actuator states to a morphology-agnostic latent space, and a morphology-conditioned decoder reconstructs target commands. Generated trajectories undergo replay filtering for task-level executability. The representation also supports a unified bimanual visuomotor policy trained across multiple pairs. Across four simulated and six real-world tasks, AsymEEF improves replay success over adapted transfer baselines, achieves a mean replay-valid expansion gain of 16.6 times in simulation, and is evaluated through downstream policy training on real robots. Project website: this https URL.

[CV-104] Learning a Task-Adaptive Low-Dimensional Semantic Space for Improved Visual Classification NEURIPS2026

链接: https://arxiv.org/abs/2611.20553
作者: Nikalal Kaldera,Wanqing Li,Chris Bunn,Philip O. Ogunbona
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 20 pages, 2 figures. Accepted at NeurIPS 2026

点击查看摘要

Abstract:Semantic visual classification commonly aligns visual embeddings with semantic spaces induced by pretrained text encoders. However, such spaces are typically high-dimensional, weakly aligned with visual feature geometry, and insufficiently discriminative for downstream classification tasks. We propose TaSS, a framework for learning a task-adaptive low-dimensional semantic space that preserves semantic structure while explicitly optimizing class separability. TaSS is constructed through a parameterized geometry transformation of pretrained text embeddings that enlarges inter-class margins via distance reshaping and improves intra-class compactness through prototype-preserving semantic constraints. The resulting semantic space maintains meaningful semantic relationships while producing a discriminative geometry tailored for classification. Frozen visual encoder features are subsequently aligned to TaSS using a lightweight MLP, enabling improved discriminative learning without fine-tuning pretrained backbones. We further provide theoretical analysis showing that TaSS preserves semantic consistency while enforcing enhanced inter-class separation and controlled intra-class variance. Extensive experiments on skeleton-based human action recognition, video classification, and image classification benchmarks demonstrate consistent improvements over existing semantic alignment approaches, with gains exceeding 3 percentage points on average and up to 8.5 points on challenging datasets.

[CV-105] A Gaussian Primitive Data Management Framework for Efficient 3DGS Trajectory Rendering

链接: https://arxiv.org/abs/2611.20546
作者: Haolun Ma,Ziyi Liu,Zizhuo Xu,Chuhan Hu,Lei Li
类目: Databases (cs.DB); Computer Vision and Pattern Recognition (cs.CV); Graphics (cs.GR)
备注:

点击查看摘要

Abstract:Trajectory rendering, which renders novel views of a reconstructed 3D scene along a camera path, is a core operation in digital twins and autonomous inspection. Structured 3D Gaussian Splatting (3DGS) renders a single view in real time by decoding sparse anchors into Gaussians on demand, but it renders a trajectory view by view: every view projects all anchors to select them, re-decodes anchors already requested by neighboring views, and runs all steps on the GPU while the CPU idles. We treat this as a spatial data management problem, where anchors are spatial objects, views are range queries, and decoded Gaussians are shared derived data, and propose GDM-GS, which removes redundancy at three levels. Within a view, an anchor index and an occluder index answer each view as a holed-frustum range query, so out-of-view and provably hidden anchors are never decoded. Across views, shared materialization decodes each requested anchor once per group of consecutive views while preserving each view’s selection. Across processors, a hybrid pipeline balances selections between the CPU and the GPU before grouped decoding under bounded memory. Experiments on large-scale public scenes show that GDM-GS substantially reduces trajectory rendering latency while preserving image quality.

[CV-106] SmoothShift: Disrupting Visual Adversarial Attacks on LVLMs at Test Time

链接: https://arxiv.org/abs/2611.20537
作者: Xiaoqing Chen,Jiayi Ouyang,Dongrui Wu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Visual adversarial perturbations can substantially impair large vision-language models (LVLMs) despite introducing only subtle input changes. To understand how these small changes cause substantial errors, we investigate their optimization and propagation through the model, revealing two key findings. First, stronger attacks rely on finer iterative optimization, while subtle input perturbations induce representation discrepancies that are amplified layer by layer through the vision encoder and persist in the language model. Second, disrupting the alignment between within-patch input patterns and the model’s computation reduces clean-adversarial representation discrepancies and substantially weakens attacks. Together, these findings suggest that adversarial effectiveness depends on both fine-grained perturbation patterns and their coupling with the model’s computation. Motivated by these findings, we propose SmoothShift, a lightweight, training-free defense that combines local smoothing and spatial transformations. Smoothing weakens fine-grained perturbation patterns, while spatial transformations disrupt their alignment with the encoder’s patch grid. SmoothShift requires no model updates, auxiliary networks, or additional model inference. Experiments cover three LVLM families on four datasets and robustly fine-tuned Contrastive Language-Image Pre-training (CLIP) models on eight classification datasets. Evaluations across eight attack methods and multiple perturbation budgets demonstrate improved robustness while largely preserving clean performance. SmoothShift also provides robustness gains under defense-aware attacks using mixed-gradient BPDA with EOT.

[CV-107] AFRNet: Adaptive Frequency Redistribution Network for Image Deraining

链接: https://arxiv.org/abs/2611.20522
作者: Dezhi Zhang,Shangrong Yang,Husheng Guo,Zhijie Shen,Wenjian Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 8 figures

点击查看摘要

Abstract:Image deraining aims to recover rain-free images from rain-degraded observations. Recent frequency-aware deraining methods have demonstrated promising performance by removing rain artifacts in the spectral domain. However, since rain artifacts and background textures often appear in several correlated subbands, prior frequency-domain deraining models still enhance frequency subbands separately. Such band-wise processing limits information exchange across subbands. Without context from other subbands, rain and clean content that appear together in one subband cannot be separated. This leads to either residual rain streaks or over-smoothing of fine details. To address this limitation, we propose AFRNet, an Adaptive Frequency Redistribution Network for image deraining. First, we introduce an Adaptive Frequency Recombination Module (AFRM) that performs cross-band modulation among wavelet subbands, enabling information interaction and adaptive recombination of degraded frequency components. We then develop a Subband Compensation Block (SCB) that compensates for decoder subband degradation by leveraging encoder subband features through inter-subband attention in the wavelet domain, alleviating detail loss during reconstruction. Extensive experiments demonstrate that AFRNet achieves competitive performance and consistent results across diverse deraining benchmarks.

[CV-108] What the Averag es Hide: Monocular Metric Depth Models Still Fail in Everyday Scenes Mostly on Scale

链接: https://arxiv.org/abs/2611.20518
作者: Hualie Jiang,Yujie Luo,Ziyang Song,Ximeng Li,Zhiqiang Lou,Rui Xu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: this https URL

点击查看摘要

Abstract:Zero-shot monocular metric depth models are evaluated by averaging metric and scale-aligned scores across datasets. Averages cannot show how many images still have severely wrong metric depth, whether depth is under- or overestimated, or whether different models fail on the same images. We propose an evaluation built on per-image results: the failure rate (metric \delta_1 0.5); scale-aligned metrics for structure, and the bias, dispersion and absolute error of the alignment scale for scale; and a cross-model comparison that separates shared from model-specific failures. We apply it to 11 recent architectures with and without ground-truth intrinsics (18 model evaluations) on the 7 everyday-scene benchmarks (4,406 images) used by MoGe-2. Metric depth estimation is not solved in everyday scenes, and the remaining difficulty is mainly scale: the best of the 18 evaluations still has a failure rate of 7.7% (datasets weighted equally); pooled over all failures, 78.8% have intact structure, and after scale correction 98.3% no longer fail. Images failed by at least 6 architectures make up 7.3% of the data, concentrated on HAMMER and DIODE; even when selected only by architectures of other families, they hold 34% to 40% of the error of the top 6 models, against 14% to 18% for as many random images from the same datasets. Adding objects reduces the scale error of most models; ground-truth intrinsics and, in most families, larger models reduce the failures but do not remove them. We recommend reporting failure rates, structure and scale separately, and results stratified by how many models fail. We will release the code, the per-image results with shared-failure labels, and two diagnostic sets built from TUM RGB-D.

[CV-109] VoxSAOR: Semantic-Aware Occupancy Robustification via Aleatoric Uncertainty Modeling

链接: https://arxiv.org/abs/2611.20507
作者: Mengfei Duan,Yuheng Zhang,Guoqiang Zhao,Yuanfan Zheng,Kailun Yang,Zhiyong Li
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO); Image and Video Processing (eess.IV)
备注: The source code will be made publicly available at this https URL

点击查看摘要

Abstract:Vision-based 3D occupancy prediction provides a dense semantic representation of scenes for autonomous driving and embodied perception, yet remains vulnerable to occlusion, depth ambiguity, and projection errors that induce spatially varying reliability in voxel features. Existing methods typically map voxel representations to semantic predictions without explicitly modeling such observation-induced feature variations. To address this limitation, we propose VoxSAOR, a semantic-aware occupancy robustification framework that models aleatoric uncertainty through feature-space proxies. It first quantifies Local Semantic Consistency (LSC) to derive a voxel-wise structural stability prior and uses cross-view semantic consensus to filter unreliable semantic correspondences. Within each semantic category, spatially matched stable and ambiguous voxels form stable-to-ambiguous feature transitions, which are modeled by a class-conditional diagonal Gaussian distribution, capturing both feature displacement and dispersion. Guided by these proxy distributions, Aleatoric Regime Ensemble Learning (AREL) trains independently parameterized prediction heads using feature perturbations at progressively increasing scales, with perturbation strength modulated by structural reliability. At inference, proxy modeling and sampling are disabled, and all heads process the same observed features; their output probabilities are then averaged to produce a deterministic consensus prediction without requiring sampling. Experiments across multiple datasets and occupancy backbones demonstrate the effectiveness of VoxSAOR in improving semantic occupancy prediction, long-tail recognition, and predictive calibration. The source code will be made publicly available at this https URL.

[CV-110] How Reliable Is Visual Token Selection? Rethinking Pruning in Multimodal LLM s EACL2027

链接: https://arxiv.org/abs/2611.20505
作者: Chaehee Song,Seokil Ham,KiHyun Seol,Jihyun Lee,Yubin Choi,Changick Kim
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to Findings of the Association for Computational Linguistics: EACL 2027

点击查看摘要

Abstract:Visual token pruning improves the efficiency of multimodal large language models (MLLMs) by removing redundant visual tokens, yet the reliability of existing selection signals has not been systematically examined. We analyze token selection under semantics-preserving perturbations and across representative pruning criteria, revealing two limitations. First, attention-based importance shifts non-uniformly under perturbations, so magnitude alone cannot distinguish stable evidence from unstable saliency (local instability). Second, attention- and diversity-based methods select substantially different subsets despite comparable performance, showing that useful evidence is not confined to a unique subset (global inconsistency). These findings motivate reliable token-set construction. We instantiate this perspective in CoRePruner, a training-free framework combining instruction-aware diversity with reliability-aware attention. Across image benchmarks, multiple MLLM architectures, and token budgets, CoRePruner improves performance, especially under aggressive pruning, and remains competitive on video benchmarks. Our results establish reliability as a central consideration in visual token selection.

[CV-111] Pointillist Maize: 3D Reconstruction of Field-Grown Maize Using RGB UAV Imagery

链接: https://arxiv.org/abs/2611.20502
作者: Dewi Endah Kharismawati,Toni Kazic
类目: Computer Vision and Pattern Recognition (cs.CV); Graphics (cs.GR)
备注: 26 pages, 8 figures, submitted to Frontiers in Plant Science

点击查看摘要

Abstract:Three-dimensional reconstruction can extend field phenotyping to individual maize organs, but thin, flexible, occluding structures and plant movement complicate reconstruction. We evaluated whether consumer-grade UAV RGB video processed with Structure from Motion and multi-view stereo could recover recognizable field-grown maize structures, and examined the effects of acquisition trajectories and frame sampling. COLMAP reconstructions preserved recognizable leaves, stems, tassel branches, ears, silks, and brace roots. For an equal-frame comparison, we reconstructed one plant from two sets of 295 frames from the same video: a single orbit sampled at 5 frames/second and three orbits at different altitudes (ribbon skirt) sampled at 2 frames/second. After ground removal, the reconstructions contained 520,601 and 375,318 points, respectively. The single orbit yielded greater point concentration within plant structures and fewer visible gaps, while the ribbon skirt yielded fewer scattered points and slightly lower proportions of isolated points. In densely planted nursery rows, visible voids occupied 17.57% of the rendered scene with one orbit at 1 frame/second, decreasing to 5.07% with additional serpentine passes and 2.68% when the original orbit was sampled at 3 frames/second. Additional observations also increased point dispersion associated with plant movement. A curlicue trajectory tested in a different field in 2024 produced distinguishable individual plants without visible duplicated structures, although substantial voids remained, mainly on the soil surface. These results demonstrate the feasibility of recovering recognizable maize organs from consumer UAV imagery and reveal trade-offs between acquisition strategies, reconstruction completeness, and point dispersion. Independent geometric validation is still needed before quantitative phenotyping.

[CV-112] Beyond Seeing: Fine-Grained Preference Validation and Severity-Aware Alignment for Mitigating Visual Hallucinations in Vision-Language Models

链接: https://arxiv.org/abs/2611.20478
作者: Abdulaziz Alqahtani,Alwaleed Alharthi,Muzammil Behzad
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 9 pages, 2 figures, 6 tables. Includes a 4-page supplementary appendix with experimental details, equations, prompt templates, and additional comparisons

点击查看摘要

Abstract:Large vision-language models can produce fluent responses containing objects, attributes, or relations unsupported by the image. We study an implemented alignment pipeline that normalizes released fine-grained hallucination supervision, validates chosen-rejected preference pairs with vision-language judges, repairs rejected chosen responses, and trains LLaVA adapters with standard or severity-aware direct preference optimization. The study uses 16,143 normalized feedback records and 8,386 released preference pairs, comparing direct training with zero-shot and two-shot judged severity-margin variants. The zero-shot variant reduces Object HalBench CHAIRs from 53.00 to 36.00 and CHAIRi from 15.72 to 10.41. Across aligned models, AMBER-Hal decreases from 35.90 to 25.10-26.90, accompanied by reduced visual coverage. In contrast, the base model retains the highest valid POPE adversarial F1, 84.18 versus 83.87-83.91 after alignment. These results support benchmark-dependent reductions in generative hallucination rather than uniform superiority of severity-aware alignment. We distinguish implemented pipeline components from the preference paths used in the reported experiments and discuss the limits of attributing gains to judging, repair, or the training objective individually.

[CV-113] Improving X-Ray Baggage Instance Segmentation with Shared Point-Based Mask Refinement

链接: https://arxiv.org/abs/2611.20476
作者: Abdulaziz Alqahtani,Qasim Umer
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 12 pages, 8 figures, 11 tables

点击查看摘要

Abstract:Overlapping projections and complex object contours make prohibited items difficult to delineate in X-ray baggage images. This study investigates shared point-based mask refinement to improve instance segmentation within Cascade Mask R-CNN. The framework combines a ResNet-101 backbone with an attention-based feature pyramid and progressively updates uncertain mask locations, increasing mask resolution from 28 \times 28 to 112 \times 112 . A single refinement module is shared across cascade stages and refinement levels, with transformer and multilayer perceptron (MLP) variants used to examine the role of interactions among selected points. Ten configurations were evaluated on the official PIXray split across three training seeds using the original target-instance annotations for evaluation. The transformer and MLP refiners achieved mean mask average precision (AP) of 71.79 and 72.47, respectively, compared with 70.50 for the unrefined detector, yielding gains of 1.29 and 1.97 AP points. The MLP achieved the highest mean mask AP among the evaluated configurations. Both variants also improved boundary agreement on commonly detected objects, supporting improved mask delineation, while detection diagnostics did not establish a corresponding positive gain. These findings demonstrate that shared point-based refinement can improve segmentation quality in the tested cascade architecture using existing target-mask annotations, without additional occluder supervision. Evaluation on independent X-ray datasets is needed to determine how well these improvements generalize.

[CV-114] LiveView: Real-Time Camera-Controlled Re-filming of Streaming Videos

链接: https://arxiv.org/abs/2611.20461
作者: Youcan Xu,Xinjie Huang,Jiaxin Shi,Hengye Lyu,Zhen Wang,Jun Xiao,Long Chen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Real-time camera-controlled video re-filming aims to synthesize target views from a streaming source video under online user-specified camera trajectories. This task poses two key challenges: maintaining synchronization and content consistency across views, and achieving low-latency rendering with long-horizon stability. Existing approaches are predominantly designed for offline full-sequence processing and rely on bidirectional attention or costly depth-based geometric preprocessing, making them unsuitable for low-latency streaming scenarios. To address these limitations, we develop a teacher-to-student framework that distills a high-fidelity bidirectional re-filming teacher into a few-step causal autoregressive student. We introduce Cross-Frame Injection to interleave each incoming source frame with its synchronized target latent, thereby establishing direct cross-view correspondence and improving content consistency. To ensure stable long-horizon generation, we further propose Shared Relative RoPE, which bounds temporal indices within the pretrained range, and Cycle Streaming Tuning, which exposes the student to extended cyclic trajectories during streaming training. These designs enable low-latency online video re-filming with precise camera control and sustained spatiotemporal consistency over long streaming rollouts.

[CV-115] VISTA: A Framework for Vision-Informed Semantic Traffic Awareness

链接: https://arxiv.org/abs/2611.20458
作者: Murat Arda Onsu,Poonam Lohan,Burak Kantarci,Aisha Syed,Matthew Andrews,Sean Kennedy
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Networking and Internet Architecture (cs.NI)
备注: 8 pages, 4 figures, Accepted for publication in IEEE Internet of Things Magazine, 2026

点击查看摘要

Abstract:The Internet of Things (IoT) is transforming intelligent transportation systems by connecting roadside cameras, edge computing units, vehicles, and traffic infrastructure into a shared sensing and decision-making ecosystem. However, transmitting raw traffic video over vehicle-to-everything (V2X) links is bandwidth-intensive, while deploying large multimodal models directly on resource-constrained vehicles is often impractical. To address these limitations, this article proposes VISTA (Vision-Informed Semantic Traffic Awareness), a multimodal vision-language edge-intelligence framework that combines predictive video representation learning with lightweight language-based reasoning to provide communication-efficient traffic situational awareness. In the proposed architecture, roadside surveillance units employ the Video Joint Embedding Predictive Architecture (V-JEPA) and an attentive-pooler module to convert traffic video clips into compact future-scene embeddings. These embeddings are transmitted through V2X links to vehicle-mounted systems, where a TinyLlama small language model, enhanced with a custom visual adapter, interprets the embeddings alongside natural-language (NL) traffic queries. VISTA enables vehicles to obtain interpretable traffic information, reason about upcoming scene dynamics, and identify potential collision risks without receiving raw video streams. Experimental results in a digital-twin urban traffic environment show that VISTA achieves 97.77% accuracy and an 89% F1 score for future collision prediction while reducing the communication payload by several orders of magnitude compared with raw RGB video transmission. These results demonstrate the feasibility of embedding-based multimodal intelligence for low-latency, communication-efficient, and safety-oriented IoT-enabled transportation systems.

[CV-116] LEDGE: Separating Critic and Editor Failures in Text-to-Image Self-Correction

链接: https://arxiv.org/abs/2611.20446
作者: Vyom Agarwal,Mokshda Gangrade,Samet Ayhan,Giacinto Paolo Saggese
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Text-to-image (T2I) self-correction combines a vision-language model (VLM) critic that diagnoses errors and proposes instructions with an image editor that executes them. End-to-end evaluation obscures whether failures arise from incorrect diagnoses or an editor unable to execute valid instructions. We introduce LEDGE (Learning Error Diagnosis in Generation and Editing), a framework that separates these bottlenecks. LEDGE compiles oracle edit instructions from GenEval while holding the generated images and the editor fixed, isolating the editor’s correction capability from the critic’s diagnostic accuracy. On Stable Diffusion 3.5 Large, oracle instructions raise GenEval from 72.0% to 81.2%, versus 74.1% for the baseline VLM critic. Yet 435 of 644 failed images remain unrepaired, exposing a substantial editor gap. We also use oracle labels from disjoint training prompts to supervise the critic. Oracle supervision raises exact diagnosis from 33.3% (the same VLM with our structured output, untrained) to 70.8%. With the same editor, the resulting critic improves GenEval from 42.2% (uncorrected) to 59.4% on Stable Diffusion 1.5, against 50.9% for the baseline critic, and from 72.0% to 76.0% on Stable Diffusion 3.5 Large, despite training only on Stable Diffusion 1.5. Improvements transfer to unseen object classes, a second editor, and independent T2I-CompBench judges and prompts. A small human evaluation confirms the ranking in estimated net corrections but suggests a smaller gain than detector-based evaluation. LEDGE thus distinguishes two limits of self-correction: knowing what to fix and being able to fix it.

[CV-117] ReScene: Incremental Scene Chronology from Sparse Revisits

链接: https://arxiv.org/abs/2611.20433
作者: Wei Qin Chuah,Zeji Hui,Amirali Khodadadian Gostar,Alireza Bab-Hadiashar
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 22 pages

点击查看摘要

Abstract:Scene chronology aims to reconstruct a scene across visits separated by hours or days, during which objects may change in position, presence or shape. Such reconstruction must retain every earlier state, yet each revisit often provides only a few images. Storing a complete model for each visit duplicates the unchanged scene, whereas updating a single shared model alters earlier states. We present ReScene, which stores each visit as an immutable compact update, so that any earlier visit can be recovered exactly by applying the stored updates in order, without retaining past images. To localise each update, ReScene attributes the disagreement between the preceding reconstruction and the new images to individual Gaussians through their compositing weights. This attribution guides the suppression of outdated Gaussians and the refinement of persistent ones, while newly visible surfaces are initialised from learned multi-view depth. We also introduce Multi-State, a benchmark of non-rigid deformation and partial material removal, including synthetic scenes with ground-truth geometry. ReScene achieves quality higher than sequential 3DGS checkpointing with 2.5–4.8 \times less uncompressed history storage. At comparable PSNR, it also reduces foreground depth error by 3.2 \times relative to sequential ChronoGS, showing that accurate rendering does not necessarily imply accurate geometry. Code and benchmark are available at this https URL.

[CV-118] UrbanTwin Sim2Real LiDAR Challenge 2026: Site-Calibrated Synthetic Data and Multi-Detector Fusion for Roadside 3D Detection

链接: https://arxiv.org/abs/2611.20431
作者: Arnav Kumar,Brijesh Soni
类目: Computer Vision and Pattern Recognition (cs.CV); Signal Processing (eess.SP)
备注:

点击查看摘要

Abstract:This report summarizes the work submitted by team arnav007 (submission ID 1054222) for the UrbanTwin Sim2Real LiDAR Challenge (V2X-Real track) at the IEEE BigData Cup 2026. The primary objective is to train 3D detectors for vehicles, pedestrians, and trucks using only synthetic roadside LiDAR data, and to generate synthetic scans that are realistic enough to be indistinguishable from real ones. A CARLA based pipeline is developed and calibrated to the fixed sensor site, accurately modeling the range image, including no-return beams, based on unlabelled real development scans. A combined inference detector, using multiple 3D detectors (VoxelNeXt, PV-RCNN, and SECOND) trained exclusively on synthetic data, is also proposed. Each scan is aligned to the synthetic ground plane and processed through the detectors at multiple ground heights. The resulting bounding boxes are fused using bird’s-eye view (BEV) clustering, followed by class-specific post-processing to refine heading, box size, and filter low-confidence detections. The final submission achieved a combined score of 0.8947. The ablation study shows that per-scan ground leveling more than doubled the real-scan mAP, while fusing boxes across different heights and models further improved performance.

[CV-119] InterRSGaussian: Interactive Multi-view Instance Segmentation for 3D Remote Sensing Scenes

链接: https://arxiv.org/abs/2611.20425
作者: Zipeng Qi,Zhengxia Zou,Zhenwei Shi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Remote sensing image segmentation plays a critical role in various applications. Interactive instance segmentation, a specialized task within image segmentation, aims to generate precise segmentation results for target objects based on user input prompts. However, previous studies on interactive instance segmentation primarily focus on single-view images, limiting their ability to comprehensively understand complex 3D remote sensing scenes. In this study, we extend interactive instance segmentation from a single-view to a multi-view framework, which can also be easily converted to a 3D segmentation format, thus achieving more accurate and comprehensive 3D remote sensing scene perception across multiple views. First, we construct a multi-view interactive instance segmentation dataset for remote sensing, comprising five distinct scenes, each captured from about one hundred different views. Second, we propose InterRSGaussian, a novel instance segmentation method based on 3D Gaussians. InterRSGaussian enables users to provide point clicks on target objects from any arbitrary view and generates view-consistent instance segmentation results across all input views without requiring manual annotations. The resulting multi-view segmentation can further be projected onto 3D point clouds for 3D segmentation. To improve training efficiency and segmentation performance, we introduce a prior distillation network for initialization. Overall, this study extends traditional 2D remote sensing scene perception into a 3D multi-view framework, enabling a more comprehensive understanding of target objects in complex environments with user interactivity.

[CV-120] raining-Free Anonymization of Driving Data with Promptable Segmentation and 3D Human Mesh Recovery

链接: https://arxiv.org/abs/2611.20412
作者: Max A. Buettner,Mark Azer,Fabian B. Flohr
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 6 figures, 3 tables

点击查看摘要

Abstract:Camera recordings for autonomous driving contain faces and license plates that should be anonymized before labeling, sharing, or training. Dedicated detectors can miss small plates and faces behind vehicle glass. We present FM-Anonymize, a training-free anonymization pipeline that combines promptable segmentation, 3D human mesh recovery, and generative replacement. SAM 3 locates people, faces, and vehicles and searches zoomed vehicle crops for plates, occupants, and windows. SAM 3D Body assesses face visibility from head orientation and projects head regions for blurring. Optional diffusion inpainting replaces eligible blurred heads with synthetic faces, guided by estimated pose and keypoints. On PP4AV, the blur pipeline protects 97.8% of faces and 95.5% of plates, the highest privacy recall of all compared methods at their default settings. On our recordings, where the localization thresholds were set, it protects 97.1% of faces and 93.4% of plates, compared with 93.7%/73.0% for UrbanAnonymizer, 42.7%/40.7% for EgoBlur, and 96.6% of faces for deface. Synthetic replacement raises full-body keypoint average precision over blurring from 54.1% to 79.6% on our recordings and from 39.9% to 68.3% on PP4AV. Code and configurations are available on this https URL

[CV-121] DominoBench: Evaluating Multi-Step Physical Reasoning from Still Images

链接: https://arxiv.org/abs/2611.20385
作者: Emilio Ferrara
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: DominoBench dataset, evaluation harnesses, and benchmarks available at this https URL

点击查看摘要

Abstract:Can a vision-language model look at a picture of a chain reaction and tell what will happen? Recent multimodal models describe scenes fluently and solve textbook physics problems, which suggests they should also predict how a scene of ramps, ledges, pendulums, seesaws and dominoes plays out once a ball is pushed. We tested this claim on 30 models from 10 families and reached a clear conclusion: unaided, no model predicts the outcome reliably. We built DominoBench, a benchmark of still images of 2D rigid-body scenes whose outcome a physics simulator computes. It holds 854 items in 427 minimal pairs, in which a change of one to three grid steps in one quantity decides whether one person survives, and every item comes with a four-rung hint ladder ending in the exact calculation that decides the outcome. Unaided, frontier models from Anthropic, Google and OpenAI are right about the person whose fate flips in at most 72% of items and answer both twins of a pair correctly in at most 34%. Pointing at the deciding step helps little, and exact geometry given as text barely helps, yet once the deciding calculation and the quantities it needs, including speeds that a still image cannot show, are written out, the same models reach 82-92% on the deciding person, while small open-weight models stay near chance. Physical prediction from a still image therefore fails at the quantitative step: the strongest models use the deciding quantity when they are given it, but do not produce it on their own. For models that plan or act in the physical world, a fluent account of a scene is thus no evidence of a correct prediction: such models must find and compute the deciding quantity unprompted, and DominoBench tests whether they do.

[CV-122] Signs of Wear: Evaluating Intelligent Speed Assistance Under Synthetic Adverse Conditions

链接: https://arxiv.org/abs/2611.20382
作者: Johannes Fabian Hahn,Mao Shan,Julie Stephany Berrio Pere
类目: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
备注:

点击查看摘要

Abstract:Intelligent Speed Assistance (ISA) systems rely on cameras to recognize speed-limit signs, yet existing benchmarks underrepresent the weathered, dirty, and overgrown signs commonly encountered on rural roads. We use generative image editing to introduce different forms of wear into real street-level images of speed-limit signs. The resulting ISA-Wear dataset contains paired original and edited images. Guided by text instructions, a multimodal image-editing model adds snow, rust, vegetation, faded paint, or mud to the speed-limit sign in a street-level frame at three severity levels. Where appropriate, the surrounding scene is adapted as well, and the viewpoint is kept. Every edited frame has an unedited counterpart, and we only evaluate frames whose original is recognized correctly, so any failure can be attributed to the wear. Since the model alters the sign’s geometry in 21% of the edits, all edits are filtered by a geometric check and a manual review. Severity is measured as the LPIPS distance between the original and the edited sign plate. On 716 worn frames, we find that in detect-then-read pipelines, wear breaks detection before it breaks reading. The detector finds only 62% of the worn signs, even though it finds every one of them in the original frames. Missed detections account for 78–87% of failed reads, a pattern that a zero-shot detector shares. Given the annotated sign box, as in the crop-based benchmarks, a classifier still reads 78% of the worn signs, but only 56% when it has to rely on the detector

[CV-123] VisDrone-ReID: A Unified Multi-Category Object Re-Identification Benchmark

链接: https://arxiv.org/abs/2611.20372
作者: Asmau Lawal Bilbis,Nathanael L. Baisa,Amudhavel Jayavel
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Object re-identification (Re-ID) benchmarks remain predominantly confined to single-category evaluations, limiting the development of unified systems capable of operating across diverse object types. This paper presents VisDrone-ReID, a unified multi-category Re-ID benchmark derived from the VisDrone2019-MOT dataset through a principled conversion pipeline. The pipeline resolves structural incompatibilities between multi-object tracking annotations and Re-ID evaluation requirements, addressing sequence-scoped identity collision, temporal redundancy, and the absence of a principled identity-level train/test split. The resulting benchmark comprises 9,516 identities and 150,016 images across ten object categories - pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor, sourced from 80 unmanned aerial vehicle (UAV) sequences captured across urban environments. A leave-one-out (LOO) evaluation protocol is adopted to provide a comprehensive assessment. Baseline experiments across four representative Re-ID models spanning CNN and transformer architectures establish the first systematic performance reference on this benchmark. Results reveal a consistent Rank-1 to mAP gap across all models, suggesting that VisDrone-ReID presents a genuine retrieval challenge beyond nearest-neighbour matching. Per-category analysis identifies pedestrian as the most challenging category and reveals that vision-language pretraining provides a measurable advantage on semantically distinctive categories. VisDrone-ReID provides a consistent platform for studying multi-category object re-identification under the scale variation, viewpoint challenges, and category diversity characteristic of real-world surveillance scenarios.

[CV-124] Structure-Guided Multi-Granularity Spatiotemporal Transformer for Efficient Monocular 3D Human Pose Estimation

链接: https://arxiv.org/abs/2611.20361
作者: Sheng Xu,Aoduo Li,Rongli He
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Existing multi-granularity methods for monocular 3D human pose estimation emphasize spatial joint features while underusing skeletal structure, including global connectivity and temporal continuity. Moreover, modeling entire frame sequences incurs high computational costs. We propose a Structure-Aware Multi-Granularity Spatiotemporal Transformer that integrates structural priors across spatial and temporal representations for accurate and efficient pose estimation. A three-level feature extractor captures human topology at the joint, bone, and limb levels. Gated Adaptive Spatial Attention (GASA) embeds structural constraints into spatial attention, while Multi-Granularity Feature Enhancement (MGFE) uses hierarchical pooling and fusion to strengthen local details and global semantics. Temporal Segment Fusion (TSF) aligns temporal contexts across granularities to capture both short- and long-range motion patterns. Hierarchical Cross-Attention (HCA) promotes spatiotemporal interaction between granularities, and Keyframe Attention (KFA) selects informative motion segments to reduce redundancy and computational cost. Experiments demonstrate improved accuracy and efficiency over conventional Transformer-based approaches, highlighting the effectiveness of structure-aware multi-granularity modeling for monocular 3D human pose estimation.

[CV-125] BlockAlign: Preference Optimization for Trajectory-Distilled Video Generation

链接: https://arxiv.org/abs/2611.20351
作者: Jingyuan Zhu,Biaolong Chen,Yangyang Qin,Shengyang Shen,Yaohua Wang,Weiming Bai,Miao Lu,Aixi Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Directly aligning trajectory-distilled video generators slashes rollout costs and optimizes the deployed few-step model without teacher fine-tuning or re-distillation. However, conventional flow-matching preference objectives assume single-head local velocity supervision, mismatching the finite-interval, block-average transition interface of trajectory-distilled generators such as Parallel Decoding Distillation (PDD) and Parametric Trajectory Distillation (PTD). Naively supervising only the primary initial-subinterval velocity sub-head rather than the fused block-average velocity at block starts (DirectAlign) induces a severe training–inference mismatch that degrades quality below the unaligned baseline. We introduce BlockAlign, an interface-consistent preference optimization framework that constructs states at valid block boundaries, queries the fused block-level velocity predictor, and supervises span-normalized transition residuals-seamlessly supporting DPO-, NFT-, and rollout-style objectives while preserving the native few-step sampler. We further build a unified post-training pipeline integrating offline pairs, reward-free semi-online continuations, and online rollouts with hierarchical preference selection and objective-aware block scheduling. Experiments on multi-reference video generation across MiniMax-H3 Ref2VA and an in-house model RefVid-V4 show that BlockAlign consistently outperforms unaligned distilled models and DirectAlign in structural plausibility, reference consistency, and visual quality. Notably, our approach unlocks PDD’s 5x few-step rollout speedup for 3x faster end-to-end online training, achieving statistically significant human preference gains over PDD baselines (54.0% vs. 33.3% wins/losses, or 61.8% non-tie win rate on RefVid-V4; 51.3% vs. 34.7%, or 59.7% non-tie win rate on MiniMax-H3) and up to an 82.7% win rate against DirectAlign.

[CV-126] Words Speak Louder? Exploring the Language Shortcut of LLM -Driven Incomplete Multimodal Learning

链接: https://arxiv.org/abs/2611.20349
作者: Wenhao Li,Chong Xiao,Xiaokun Bai,Rundong He,Zhuang Qi,Xusheng Liang,Zhibin Wu,Qiangchang Wang
类目: Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)
备注: 19 pages, 7 figures, 14 tables. Project page: this https URL

点击查看摘要

Abstract:Is a face or a voice still heard once the words are known? Incomplete multimodal learning increasingly uses large language models (LLMs) as fusion engines: modality features enter as token blocks and missing modalities become placeholders. We find that LLM-driven fusion can override non-verbal evidence it can read. Contrasting its full-input prediction with counterfactual views, the same model shown one modality or no content, we identify language shortcuts. We build MR-Bench: MR-Bench-A types every test sample of four sentiment benchmarks; MR-Bench-H contains 900 human-validated clips (620 language-shortcut, LS; 280 non-verbal-shortcut, NS) from seven sources, and every model is scored on all of them. Zero-shot MLLMs solve at most 32.7% of the LS clips; adding the transcript to the muted video lowers vision-language accuracy by 46.1-53.1 points. Yet inverting a text-only model’s answers solves 94.4%, so LS accuracy alone cannot establish use of non-verbal evidence. Inside the fusion LLM, fused states align increasingly with the text view late in the decoder. Counterfactual-View Reliability learning (CoVeR) trains views as unimodal judges and withholds fused supervision where they disagree. It achieves state-of-the-art results on five benchmarks without extra inference cost in accuracy mode. On MR-Bench-H, it reaches 61.7% LS and 93.6% NS accuracy; robust mode raises LS accuracy to 63.0% while lowering NS accuracy to 84.8%. Among LS clips solved without text, adding text causes errors on 30.1% with CoVeR versus 32.2% with LLM-Fusion. Project page: this https URL.

[CV-127] Unreason able Confidence from Simulator Misspecification: Directional Robustness on the SPEED Sim-to-Real Benchmark CEC NEURIPS2026

链接: https://arxiv.org/abs/2611.20321
作者: Nikolay Dolgov
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 22 pages, 5 figures, 20 tables. Extended version of a paper at the Sim2Science workshop, NeurIPS 2026. Code, per-image predictions and pre-registration record: this https URL

点击查看摘要

Abstract:Vision-based navigation systems for orbit must be trained almost entirely on synthetic imagery, because real labelled images of a given spacecraft in flight rarely exist. Benchmarks like SPEED+ are built for that situation: the training set comes from a simulator, and the test sets are real images of a satellite mockup under two kinds of lighting. That simulator approximates how light reflects off mylar and aluminium, and like every scientific simulator it is misspecified. Its error is radiometric, since it misses the specular highlights, hard shadows and sensor saturation common in real images. We probe the resulting synthetic-to-real gap, the central risk for such systems, with a deliberately simple monocular pose estimator. In a 2x2 study of backbone adaptation and photometric augmentation, trained on synthetic images and tested on both real hardware-in-the-loop domains, we find three things. (1) The model does not know it is failing: regions conformally calibrated to 90% on synthetic data cover only 9-15% of rotation errors on sunlamp and 21-38% on lightbox, so a wrong pose carries the confidence of a correct one. (2) Transfer is asymmetric: scaling data and features improved synthetic rotation error from 1.65 to 1.03 rad but moved the harshly lit sunlamp domain by only 0.05 rad, under twice its seed s.d., whereas adapting the backbone’s last stage carried 63-83% of its extra synthetic gain to both real domains. (3) Once the backbone adapts, augmentation helps only in the direction it models: our photometric stack improved the diffusely lit lightbox domain from 1.17 to 0.99 rad on every seed but moved sunlamp no more than seed-to-seed variation, while with the backbone frozen it helped sunlamp more. Every post-review experiment was pre-registered, and the code, per-image predictions and pre-registration record are at this https URL.

[CV-128] PoCoAttn: Position-Content Collaborative Sparse Attention for Fast and High-Fidelity Video Generation

链接: https://arxiv.org/abs/2611.20307
作者: Jianwen Wu,Meiling Wang,Junkang Zhou,Yipeng Shen,Haoran You
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 21 pages, 15 figures, 13 tables. Project page: this https URL

点击查看摘要

Abstract:Video diffusion transformers rely on 3D full attention over tens of thousands of spatiotemporal tokens, and its quadratic cost dominates inference latency. Sparse attention reduces this cost, but existing methods face a dilemma: static methods apply predefined masks and cannot capture input-dependent attention, whereas dynamic methods reconstruct the sparse mask repeatedly, paying online to rediscover positional structure that is largely stable across inputs. To this end, we propose PoCoAttn, a training-free sparse attention framework that separates input-stable from input-dependent attention and resolves the former offline and the latter online. PoCoAttn features: (1) A relative-offset analysis of 3D query-key pairs, which shows that video attention decomposes into an input-stable positional prior that carries most of the attention mass and a small content-dependent residual. (2) A position-content collaborative sparse attention mechanism, which retrieves positional keys from offline-calibrated tables, selects content keys online with a head-adaptive ratio between the two sources, and compensates for the discarded keys to approximate full attention. (3) A step-adaptive budget allocation algorithm, which models how early-step sparse attention errors are amplified along the denoising trajectory and derives the optimal sparsity budget per step. Extensive experiments across four video diffusion models and three tasks show that PoCoAttn consistently achieves a better speed-fidelity Pareto frontier than state-of-the-art (SOTA) methods, improving PSNR by 1.9 to 5.0 dB over baselines and delivering up to 4.1 \times attention and 2.0 \times denoising speedups over dense FlashAttention-3 attention.

[CV-129] CiGS: Conduction in Gaussian Splats for Interior Temperature Field from Thermal Images

链接: https://arxiv.org/abs/2611.20283
作者: Liu Zhenyuan,Bharath Seshadri,Bailey Miller,George Kopanas,Ionnais Gkioulekas,Bernd Bickel
类目: Computer Vision and Pattern Recognition (cs.CV); Graphics (cs.GR)
备注: 18 pages

点击查看摘要

Abstract:We present a method for inferring the interior temperature distribution of a solid object from registered multi-view visible-spectrum and thermal images, even when the thermal views cover only part of its surface. We formulate steady-state heat conduction as a Laplace boundary value problem and solve it with a direct boundary integral equation discretized over a point cloud: a Gaussian reconstruction of the visible-spectrum images supplies the boundary geometry, and the thermal images supply the boundary data. We do not build surface or volume meshes at any stage in our method. Instead, we sample collocation points from rendered depth without the pixels at depth discontinuities, and use a density subtraction that keeps the interior evaluation accurate up to the surface. The whole chain is differentiable, and we use it to recover the pose and depth of heating pipes, insulation defects, and the shape of cracks from thermal images. We validate our method through experiments on boundary value problems with Dirichlet-only and mixed Dirichlet and Robin boundaries, using synthetic and real captured data on scenes with challenging topology. These experiments show that our method matches analytical solutions and has better accuracy than FEM solvers.

[CV-130] Visual Marking as a Native Reasoning Capability for Multimodal LLM s

链接: https://arxiv.org/abs/2611.20278
作者: Jiazhen Liu,Yanghao Wang,Chaolei Tan,Long Chen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Visual reasoning often requires maintaining intermediate spatial states that are difficult to express in text. Existing approaches either rely on external drawing tools or image generators, or encode these states in latent tokens without explicit spatial correspondence to the input image. We introduce SelfMark, which enables MLLMs to construct and use visual marks directly in their native visual feature space. By treating the visual token grid as a canvas, SelfMark establishes explicit spatial correspondence between marks and images. This is achieved through a locate-mark-read procedure, where the model identifies relevant locations, constructs task-dependent mark features, and reads the resulting visual features to guide its reasoning. We jointly train a single model on spatial planning, dense counting, high-resolution VQA, and line tracing. Across six benchmarks, SelfMark outperforms state-of-the-art MLLM-based methods specialized for individual tasks. These improvements also persist with a stronger backbone, demonstrating the effectiveness of visual marking across different model capabilities. Overall, SelfMark provides a unified approach to visual reasoning by integrating spatially grounded marks directly into the visual feature space of MLLMs.

[CV-131] Self-supervised learning for zebrafish embryo phenotype evaluation: Systematic comparison of algorithms label-efficient classification and transfer learning

链接: https://arxiv.org/abs/2611.20269
作者: Thomas Lautenschlager,Gaëlle Hayot,Quentin Brassart,Angelo Jovin Yamachui Sitcheu,Linus Schilpp,Tijana Geroski,Tetyana Goldanova,Stephan Allgeier,Thibaut Soubrié,Milica Nikolic,Thomas Dickmeis,Ralf Mikut,Nils Friederich
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Zebrafish embryos are a prominent model organism for image-based phenotyping and especially suited for large-scale imaging studies, for which machine learning (ML) models can automate evaluation. However, training ML models usually requires large labeled datasets, and differences in imaging conditions often impede reuse across laboratories. Self-supervised learning (SSL) addresses both constraints by learning transferable representations from unlabeled images. Yet central aspects like the choice of algorithm remain underexplored for domain-specific applications. To close this gap, we benchmark popular SSL algorithms on the public EmbryoNet dataset, explore label-efficient learning, and test transfer to an unseen setting. Using representations from our strongest model, a ResNet50 network trained with SwAV, a simple linear classifier reaches 86.4% accuracy and a non-linear classifier 87.1%, rivaling the fully supervised EmbryoNet model (87.1%). The SwAV model decisively outperforms supervised training from scratch when labels are scarce (70.2% vs. 45.7% accuracy with 5% of labels). On a three-class problem (normal, dead, falsely cropped) with a small, previously unseen zebrafish toxicity dataset, recorded in a different setting, our model reaches 81.3% accuracy without retraining, while EmbryoNet achieves 75.9%. Overall, self-supervised models for zebrafish embryos seem to mostly facilitate label-efficiency and transfer to new settings rather than improving supervised results.

[CV-132] EvoMem3D: Adaptive-Depth Temporal Representation Memory for Multi-Camera 3D Detection

链接: https://arxiv.org/abs/2611.20241
作者: Moshe Kimhi,Idan Kligvasser,Yotam Intrator,Regev Cohen,Ehud Rivlin
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Modern multi-camera 3D detectors increasingly rely on large Vision Transformer (ViT) image encoders, making repeated full-depth encoding a dominant inference cost even though adjacent driving frames are highly redundant. We present EvoMem3D, an adaptive-depth temporal framework that caches the most recent full-depth detector-interface appearance representation as a read-only anchor and combines it with policy-selected partial features from the current frame. A lightweight pre-encoding controller chooses among semantic encoder exits before image processing, allocating shallow computation to stable frames and deeper computation when the visual state changes. To reuse appearance without carrying stale geometry, we decouple temporal appearance from geometric alignment: current-frame calibration defines the decoder geometry at every timestamp, while a motion-conditioned positional adapter accounts for anchor-to-current ego motion. On nuScenes, EvoMem3D achieves 61.50 NDS and 52.76 mAP while reducing end-to-end latency from 364.4 to 168.7 ms (x2.16) relative to standard dense DETR3D. With StreamPETR, it reaches 61.36 NDS and 52.55 mAP at a x1.92 speedup. Controlled studies isolate the effects of current evidence, temporal reuse, memory correction, and adaptive routing, while our audit confirms that the policy learns to distinguish semantic driving regimes.

[CV-133] Ocular Recognition: Interoperability Multi-Resolution and Spectrum Challenges

链接: https://arxiv.org/abs/2611.20219
作者: Fabian Stockhardt,Juan E. Tapia,Christoph Busch
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Paper submitted to IEEE journal in revision process. Abstract reduces to 1920 characteres

点击查看摘要

Abstract:Ocular recognition is typically developed under controlled acquisition conditions, limiting its use in semi-controlled and on-the-move scenarios. A key obstacle is the lack of representative data spanning real-world capture conditions. At a border checkpoint, for example, images may range from low-resolution visible-light periocular crops (RGB) to high-resolution near-infrared iris images (NIR). Using this heterogeneous evidence raises three questions: whether resolution-specific processing improves recognition, how to compare templates from different processing routes, and how to allocate training data across resolution specialists. We investigate resolution-specific processing and parameter sharing using three AdaFace-based architecture families: fixed-resolution specialists; Independent Backbones with Shared Heads (IBSH); and models with resolution-specific front-ends for 40x40, 80x80, and 160x160 inputs, followed by a shared identity trunk and embedding projection. We evaluate each architecture with route-specific training subsets and with a full-pool strategy, in which every specialist uses the complete selected training set. Results reveal a trade-off among same-resolution specialization, cross-resolution interoperability, and common-route re-extraction. Route-specific training produces distinct performance profiles across routes, whereas full-pool training makes route performance more uniform and primarily benefits the 160x160 route. Thus, adding training data does not improve all resolution routes equally. Comments: Paper submitted to IEEE journal in revision process. Abstract reduces to 1920 characteres Subjects: Computer Vision and Pattern Recognition (cs.CV) Cite as: arXiv:2611.20219 [cs.CV] (or arXiv:2611.20219v1 [cs.CV] for this version) https://doi.org/10.48550/arXiv.2611.20219 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[CV-134] AstroPIE-53: Towards Understanding Failure Modes of Human Pose Estimation in Space

链接: https://arxiv.org/abs/2611.20207
作者: Stephanie Käs,Marten Flaig,Lukas Woyke,Osmanalp Ömeroğlu,Felix Krueckel,Onur Alp Çulha,Bastian Leibe
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 4 pages + references, 4 figures. Published in the Proceedings of SPAICE 2026 and i-SAIRAS 2026

点击查看摘要

Abstract:Reliable monitoring of astronauts is essential for safety, situational awareness, and autonomous assistance in space missions. While modern computer vision enables accurate human detection and pose estimation on Earth, its robustness in space environments remains largely unquantified due to limited data availability. We present AstroPIE-53, a pilot dataset of 53 annotated astronaut videos collected from publicly available mission footage. To respect copyright restrictions, we release video identifiers, reconstruction scripts, and dense annotations rather than raw videos. The dataset includes bounding boxes, challenge classes, and human-annotated pose failure tags. We benchmark state-of-the-art person detection and pose estimation models under realistic zero-shot deployment conditions and perform the first systematic failure analysis for astronaut monitoring. Our results reveal substantial performance degradation caused by occlusions, unusual body orientations, low contrast, and microgravity motion patterns.

[CV-135] Radar Panoptic Transformer: Semantic-Guided Mask Queries for Dynamic Scene Understanding

链接: https://arxiv.org/abs/2611.20203
作者: Mayank Mayank,Bharanidhar Duraisamy,Rohit Mohan,Abhinav Valada
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 9 pages, 4 figures

点击查看摘要

Abstract:Automotive radar provides long-range sensing and direct Doppler velocity measurements, but its sparse and non-uniformly sampled point clouds pose significant challenges for panoptic segmentation. Dynamic objects often yield only a few detections amid a dominant static background, providing limiting the geometric cues for separating object instances. We propose \emphRadar Panoptic Transformer (RPT), an end-to-end point-based network that incorporates radar-adaptive feature extraction with semantic-guided mask prediction. Its backbone aggregates pose-aligned temporal context through grouped vector attention and learns per-point neighborhood radii conditioned on range, azimuth, and radar cross section (RCS). Point-wise semantic predictions initialize queries at likely foreground detections and guide masked cross-attention through query-point class agreement, producing instance masks without geometric clustering. An auxiliary RCS-consistency loss matches instance-level statistics between predicted and annotated instances. RPT achieves state-of-the-art performance on RadarScenes with 71.6% mIoU and 82.2% PQ. We further evaluate RPT on MAN TruckScenes, where it outperforms the evaluated baselines in a long-range, full-surround radar setting with sensing ranges up to 230,\mathrmm . RPT achieves real-time end-to-end inference at 31 ,Hz on RadarScenes, exceeding the sensor frame rate.

[CV-136] GlyphAtlas: Disentangling Stroke Detail from Layout for Scene Text Synthesis

链接: https://arxiv.org/abs/2611.20192
作者: Luozheng Qin,Changhai Zhou,Yiyi Zhou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Scene text synthesis requires stroke-accurate characters rendered in a specified layout, as well as visual coherence with the surrounding scene. To achieve this, existing methods commonly pre-render the target text at its intended locations to form a glyph image that jointly specifies layout and glyph structure. However, this representation ties the resolution of each glyph to its target region, limiting fine-grained stroke guidance for small, densely packed, or structurally complex text. To tackle this, we propose GlyphAtlas, a scene text synthesis framework that retains the glyph image as a spatial anchor and supplies fine-grained stroke references through a dedicated glyph atlas. The atlas renders and encodes each character or word in a separate cell, decoupling reference resolution from target-region size. We further introduce atlas-anchored attention, which lets each text region retrieve stroke detail only from its corresponding atlas cells to refine the coarse glyph structure provided by the glyph image. This coarse-to-fine conditioning simplifies stroke rendering while preserving explicit layout correspondence. We optimize GlyphAtlas with a mask-weighted flow matching objective and a timestep-weighted text perceptual loss for writing accuracy and background consistency. Random resolution augmentation of atlas cells supports adjustable atlas budgets at inference. Extensive experiments demonstrate that GlyphAtlas achieves stroke-accurate synthesis of small, dense, and structurally complex text, zero-shot transfer to languages unseen during task-specific training, and leading performance in scenes with many text regions. project page: this https URL

[CV-137] GeoReVis: Learning Selective Revision and Evidence Consolidation for Remote Sensing Segmentation

链接: https://arxiv.org/abs/2611.20183
作者: Ke An,Liang Chen,Yupei Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 31 pages

点击查看摘要

Abstract:Recent methods try to inspect and revise masks to mitigate referential ambiguity and localization errors caused by complex spatial relations in language-guided remote sensing segmentation. However, these solutions often lead to Revision Degradation: revising missing regions may introduce errors into previously correct ones. Even a globally better mask can omit correct regions found in inferior masks. We term this Evidence Fragmentation: correct target evidence is scattered across imperfect masks. All masks generated by the model at this time are suboptimal. We propose GeoReVis, which unifies revision guided by execution feedback, Candidate Selection, and Cross-Mask Evidence Consolidation. It learns fine-grained revision from real actions that improve segmentation when executed and learns to select a better mask through quality comparisons between actual candidates, reducing the risk of degradation during revision. The selected whole mask can still discard useful regions. GeoReVis therefore retains all candidates across rounds, partitions their spatial disagreements into composable regions, and jointly reconstructs regions conditioned on the full instruction and context. This makes complementary and fragmentary evidence reusable regardless of its source candidate’s overall quality, yielding improved segmentation. We further build a data pipeline to generate fine-grained multi-round segmentation trajectories and derive supervision for revision actions, candidate selection, and region composition from the model’s own outputs, extending training beyond whole-mask IoU supervision. We also propose ReVis-Eval, a benchmark protocol evaluating multi-round revision and cross-round evidence consolidation accuracy. Experiments on ReVis-Eval and 5 benchmarks for remote sensing reasoning and referring segmentation demonstrate the effectiveness of GeoReVis.

[CV-138] Aerial Person Detection under Edge Constraints: Held-Out Evaluation and Edge-Ground Verification

链接: https://arxiv.org/abs/2611.20127
作者: Raghu Venkat,Tricha Anjali
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注: 14 pages, 1 figure, 8 tables. Code, truth registers and result files: this https URL

点击查看摘要

Abstract:A search drone sees a missing person as a few pixels, carries a small computer, and shares a narrow radio link. We study person detection under those limits; the primary evaluations were declared before they ran, and later analyses are marked exploratory. In an exploratory comparison, holding out recording blocks inferred from thermal frame identifiers cut recall at a fixed false-alarm target from 32.7% to 12.3%. With nothing tuned on test blocks, a 46,000-parameter detector gated by a contrast cue finds 10.4% of people against 5.7% for the same network applied everywhere. A standard detector, YOLO11n, trained under the same folds finds 54.5% at a similar false-alarm rate for about eleven times the arithmetic, and on our 4K field footage at 117 m has a walking person among its top 16 detections in up to 92% of frames against 74% for our board. Run alone, YOLO11n’s thresholds set at one location overshot their false-alarm target 1.7 to 3.4 times at new locations, and its best configuration at altitude cost 4 to 24 standard passes per frame. Re-scoring its 16 best crops (154 KB per frame) with an 842,000-parameter ground model, thresholds frozen on a validation location, raised recall from 44.9% to 53.7% against the same 16 crops with YOLO’s own scores; in a post hoc paired analysis over 15 test locations the gain was +8.8 points [+2.4, +12.5]. Pooled false detections fell from 0.86 to 0.64 per frame, but that paired change was not distinguishable from zero. Coverage of a still person in earth tones was very low for both detectors in the clips we flew.

[CV-139] Sequential Semantic-Non-Semantic Watermarking: Robustness Against UnMarker Attacks

链接: https://arxiv.org/abs/2611.20115
作者: Md Ifraham Iqbal
类目: Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM); Image and Video Processing (eess.IV)
备注: 13 pages, 16 figures. Preprint based on this http URL . thesis research

点击查看摘要

Abstract:Advancements in artificial intelligence (AI) have accelerated digital content generation and intensified concerns around fraud, misuse, counterfeiting, and content provenance. Defensive image watermarking provides a practical mechanism for identifying and tracing digital content, but recent watermark-removal attacks challenge its robustness. This study investigates Sequential Semantic Non-Semantic Watermarking (SSNSW), which combines HiDDeN and StegaStamp as complementary non-semantic and semantic watermarking schemes. Using 50 images from MS-COCO and CelebA-HQ, both embedding orders are evaluated, followed by UnMarker’s non-semantic and semantic attacks applied individually and sequentially. Results show that both watermarks remain highly decodable after dual embedding, with only limited embedding-order effects. Under individual attacks, one watermark generally remains strongly decodable: StegaStamp under the non-semantic attack and HiDDeN under the semantic attack. Sequential attacks are therefore required to substantially reduce both watermark signals, while attack order has a limited and configuration-dependent effect. The semantic attack also shows a broader effect across the two watermark types than the non-semantic attack. LPIPS remains relatively low in several settings, although qualitative inspection reveals noticeable distortions, particularly in selected facial images, indicating that LPIPS should not be used alone to assess visual quality. Overall, the findings show that combining complementary watermark types can increase robustness against single-stage removal attacks and motivate broader multi-layer watermarking, stronger perceptual evaluation, and further study of sequential watermark-removal strategies. Comments: 13 pages, 16 figures. Preprint based on this http URL. thesis research Subjects: Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM); Image and Video Processing (eess.IV) Cite as: arXiv:2611.20115 [cs.CR] (or arXiv:2611.20115v1 [cs.CR] for this version) https://doi.org/10.48550/arXiv.2611.20115 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[CV-140] rading Parameters for Accuracy in MambaVision for Chest Radiograph Classification

链接: https://arxiv.org/abs/2611.20103
作者: Qaiser Farooq,Tehseen Zia
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 4 figures, 9 tables. Submitted to an IEEE journal

点击查看摘要

Abstract:MambaVision, a hybrid state-space and attention backbone, places most of its parameters in its widest, fourth stage. We quantified what that stage contributes to multi-label chest radiograph classification. The fourth stage was replaced with lighter single-scan state-space blocks adapted from Spatial-Mamba, with multi-scale dilated depthwise convolutions applied to the scan output; the first three stages were retained. Tiny, Small, and Base models were evaluated across nine dataset-scale combinations drawn from two source cohorts: the National Institutes of Health ChestX-ray14 archive (full ChestMNIST and a high-resolution subset) and Indiana University Open-i. At the Tiny scale, replacement reduced parameter count by 35.7% and increased graphics processing unit throughput by 19.3%; the savings came from the stage’s dense projections, not from removing quadratic attention. Without distillation, the mean area under the receiver operating characteristic curve on ChestMNIST was 0.004 lower than the baseline (95% bootstrap confidence interval, 0.002 to 0.006), and 0.003 to 0.006 lower across all nine combinations. Knowledge distillation from an unreplaced teacher narrowed the gap to the undistilled baseline to 0.001-0.002 and met the equivalence criterion on ChestMNIST, but distilling the baseline with the same recipe improved it by 0.001-0.003, so a gap of 0.002-0.005 persisted under symmetric training. At the Base scale, the unreplaced Small model was both smaller and more accurate. Fourth-stage replacement is a useful trade-off only for Tiny and Small models. These measurements let developers of chest radiograph screening tools weigh model size and speed against a quantified loss of diagnostic discrimination.

[CV-141] Geo2Gcode: From 3D Geometry to CNC Programs through Structured Machining Primitives

链接: https://arxiv.org/abs/2611.20092
作者: Sangmin Hong,Dohee Cho,Mohsen Yavartanoo,Kyoung Mu Lee
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Turning a 3D shape into a Computer Numerical Control (CNC) program is still largely manual. For every part, a machinist chooses setups, tools, step-downs, and toolpath patterns. We present Geo2Gcode, which converts a target mesh directly into standard G-code through structured machining primitives. Each primitive couples an indexed machining direction, a closed per-view reconstruction of the target, and explicit tool-radius, layer-height, and clearing-pattern parameters, and compiles deterministically into layerwise toolpaths via signed-distance queries. Most of the fidelity comes from this compiler. A lightweight selector trained with a voxel stock simulator fills in primitive parameters without per-part manual settings, approaching the finest fixed setting at about a quarter of its toolpath length. In simulation on ABC parts, Geo2Gcode reaches roughing fidelity comparable to an expert-tuned commercial CAM setup with a moderately longer toolpath, and clearly improves on a default CAM configuration. Generated programs run in a commercial G-code simulator, and one part is machined on a physical CNC mill. These results demonstrate that roughing programs can be generated directly from geometry without per-part CAM tuning.

[CV-142] Latent World Modeling for Point-Cloud Video Pretraining

链接: https://arxiv.org/abs/2611.20090
作者: Zhuoyue Zhang,Jian Liu,Jihua Zhu,Ajmal Saeed Mian
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Masked modeling is the dominant self-supervised paradigm for point-cloud videos. However, it leaks the information it is meant to predict, whether it reconstructs masked points as in MAE-style methods or predicts their latent features. Because point-cloud tokens are anchored at points sampled from each frame, querying a target at its own position reveals where its geometry has moved, in reconstruction and latent prediction alike. For latent predictors, random spatio-temporal masks introduce a further shortcut, as visible frames on both sides of each target allow interpolation rather than forecasting. The pretext task thus reduces to completion. Once the target positions are withheld, a masked latent predictor performs only slightly better than copying the last frame. To remove this leakage, we propose PointWM, a latent world model that turns completion into forecasting by predicting the features and geometry of future frames from strictly earlier context. Each target is queried only with its last-seen position and timestamp, while a latent action learned without annotations is meant to capture the part of the future that the past leaves undetermined. On HOI4D action segmentation with a P4Transformer backbone, PointWM improves F1@50 over training from scratch by 24.9 offline and 29.1 under strictly causal inference. These results show the promise of latent world models as a leakage-free pretraining paradigm for point-cloud this http URL code is publicly available at this https URL.

[CV-143] See the Lives of Stars in a Handful of Light

链接: https://arxiv.org/abs/2611.20089
作者: Shuhong Liu,Gengjia Chang,Quanfeng Xu,Ziteng Cui,Qianyu Zhou,Xiaoyu Dong,Lin Gu,Tatsuya Harada
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Galaxies evolve over hundreds of millions of years, and no telescope has ever watched one move. Cosmological simulations make this evolution visible as a sequence of multi-view images, inviting the question of whether a galaxy can be recovered in four dimensions from its images. This setting defeats existing dynamic scene methods, since millions of stars span only a few pixels each, faint populations blend into diffuse light, and free deformations reproduce the images without obeying gravity. We present Cosmos4D, which rests on two components. Gravitational 4D Gaussian Splatting represents each star as a self-luminous Gaussian whose motion is integrated through a gravitational potential and a low-dimensional force field shared by all stars. Luminosity-stratified dynamics give bright stars individual orbits supervised across image scales and carry faint populations with a velocity field estimated from optical flow. We further build a 4D dataset of eight galaxies from four cosmological simulations, on which Cosmos4D outperforms dynamic Gaussian baselines in both photometric quality and physical accuracy. With these advancements, we glimpse eight hundred million years of a galaxy’s life in ninety-six frames.

[CV-144] SGIP: Learning Scene-Grounded Interaction Priors for Scene-Aware Human Motion Prediction NEURIPS2026

链接: https://arxiv.org/abs/2611.20076
作者: Donghyeon Soon,Jiil Yoon,Hyeondong Kim,Daehee Park
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to NeurIPS 2026. Project page: this https URL

点击查看摘要

Abstract:Human motion prediction (HMP) aims to forecast future movement from an observed motion sequence, and recent methods have incorporated 3D scene geometry and object semantics to improve long-term plausibility. However, the destination of future motion remains ambiguous in cluttered scenes: the same observed motion can lead to different interaction targets or navigation modes. Existing methods either handle this ambiguity implicitly or rely on auxiliary signals like gaze that are hard to obtain in practice. To address this, we propose SGIP, which explicitly predicts where in the scene future motion will be grounded across both objects and open navigation regions, and uses it to condition future trajectory and full-body pose prediction. This scene-grounded interaction prior is learned via intent distillation: at training time only, a teacher model scores candidate scene regions by combining an LLM prompted with a textual description of the person’s intent with geometric motion-scene features. A student model learns to reproduce these scores from only the observed motion and 3D scene, making intent and LLM unnecessary at test time. Across multiple benchmark datasets, SGIP outperforms prior scene-aware baselines on both trajectory and pose accuracy in a realistic deployment setting where neither gaze nor intent is available at inference, demonstrating the effectiveness of the proposed framework. The project page is available at this https URL.

[CV-145] Prompt Re-Specification and Target-Mask Availability in Mask-Conditioned Visual Reinforcement Learning for Moving-Platform Landing

链接: https://arxiv.org/abs/2611.20075
作者: Ho Tim Tsui,Jongyun Kim
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 11 pages, 2 figures, 12 tables

点击查看摘要

Abstract:Visual reinforcement learning (RL) policies that land a drone on a moving platform can fail when the landing pad looks different from the one used in training. Using binary target masks as the visual input removes most pixel-level appearance information while preserving the spatial representation of the target, and lets a text prompt specify the target. In simulation, we test whether a frozen (fixed-weight) mask-based DrQ-v2 policy can handle a new pad when only the prompt of its fine-tuned SAM 3 detector changes. A banana-themed pad lowered landing success from 75.6% to 12.0%, and SAM 3 rarely detected it under the original prompt. A revised prompt, chosen after testing several candidates, restored success to 68.4% with no weight updates. In a separate experiment with a policy trained on ground-truth masks, filling in masks that were missing while the pad was visible helped landing far more than the tested mask-centroid and area corrections. Tracking the pad over time supplied masks more often and also kept producing them after the pad was confirmed absent from the rendered image, but it did not establish a reliable improvement in landing success. Whether a usable mask reaches the policy at all thus appears to be an important limiting factor for this approach.

[CV-146] MARLIN: Multiscale And Directional Sparse Representation Learning for Imbalanced SAR Ship Classification

链接: https://arxiv.org/abs/2611.20065
作者: Benyamin Hosseiny,Kamirul Kamirul,Odysseas Pappas,Alin Achim
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Despite recent progress in SAR maritime monitoring, ship classification remains challenging due to limited labeled data, class imbalance, large intra-class variability, fine-grained inter-class similarity, and variations in target scale, orientation, and backscattering structure. To address these challenges, this paper proposes MARLIN, a class-frequency-aware framework with directional sparse and hierarchical multi-scale feature extraction for ship type classification in SAR images. The proposed architecture first constructs two complementary representations of each input image, where each views are processed by parallel encoder branches. Each branch begins with a directional sparse convolutional stem designed to capture elongated and orientation-dependent ship scattering patterns, followed by customized hierarchical multiscale feature extraction blocks. To improve learning under imbalanced class distributions, we introduce a class-frequency aware objective. The proposed objective combines margin-based classification with class-frequency-weighted prototype compactness and class-pair frequency weighted margin separation. This formulation shapes the embedding space by encouraging samples from the same class to form compact clusters while enforcing stronger separation for class pairs involving underrepresented ship categories. Extensive experiments on three complementary datasets demonstrate that in addition to being lightweight and parameter efficient, MARLIN improves both overall accuracy and class-balanced metrics, confirming its effectiveness for robust SAR ship-type classification under limited and imbalanced training conditions. The codes are available at this http URL.

[CV-147] Machine Learning for Near Real-Time Identification of Wildfire Burned Areas

链接: https://arxiv.org/abs/2611.20036
作者: Khoi Nguyen,Adam W. Bargteil,Milton Halem
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
备注:

点击查看摘要

Abstract:In this article, we present a near real-time method that estimates burned probability anywhere in the world using satellite images. Our approach takes advantage of the state-of-the-art pix2pix model to translate a preprocessed index calculated from Landsat 8/9 images into an approximate of USGS Burn Probability indicating per pixel burn probability. Since this is a machine learning-based approach, it can be automated with a minimum amount of human intervention with significantly less time delays and geographic restrictions compared to other contemporary methods. We include results from our experiments and compare our approximation with the USGS Burn Probability using RMSE score.

[CV-148] acuSimNet: Multi-View Self-Occlusion-Awared Visibility Learning for Cranio-Cervical Acupuncture Points

链接: https://arxiv.org/abs/2611.20025
作者: Jiatao Ma,Qilei Sun
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 19 pages, 6 figures, 3 tables

点击查看摘要

Abstract:The localization of acupuncture points (acupoints) in Traditional Chinese Medicine (TCM) presents unique challenges since they are defined by abstract principles rather than distinct anatomical landmarks. Existing approaches are typically constrained to single view or rely on indirect calculations of relative positions with respect to other landmarks, thereby overlooking human anatomical variations. Furthermore, acupoint visibility assessment, which determins whether points are occluded by human itself, hair, clothing, or other objects, has received limited attention in practical applications. Acupoint localization itself does not require 3D reconstruction, but inferring their occlusion relationships with anatomical surfaces does, which adds computational cost and limits real-time inference. In this work, we introduce acuSimNet, an efficient hierarchical multi-task learning architecture for multi-view, self-occlusion-aware visibility prediction of acupoints, achieving 99.97% accuracy on the validation set. Our approach also addresses the challenges of high-dimensional classification (174 acupoints in cervicocranial region), negative convergence issues for visible acupoints, and inter-task scheduling optimization, resulting in substantially accelerated convergence. We improved the training efficiency from the exisiting methods of 3000 epochs to achieve 99% validation accuracy, to our optimized framework achieving 90% accuracy in 39 epochs and 99% accuracy in 86 epochs. This architecture overcomes the limitations of existing methods, could enable practical applications in acupoints detection and visualization, advancing the automation of TCM.

[CV-149] FastGRN: Few-Step Generative Refinement Networks

链接: https://arxiv.org/abs/2611.20021
作者: Haoyang Hu,Masataka Seo,Liang Li,Yen-wei Chen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Generative Refinement Networks (GRN) generate an image by repeatedly re-predicting and refining every token of a discrete latent. They reach high fidelity, but need dozens of Markov steps. We therefore train few-step GRNs on the distribution of their own samples. We use its own uniform-state corruption to perform distribution matching on discrete tokens, which lets a discrete generator learn by backpropagation, and we pair noise with samples in the manner of rectified flow. We find that the steps of a few-step GRN do not carry the same amount of information or the same repair load. We therefore exploit its discrete refinement during training and build few-step GRNs in three parts. Oversampling samples the student with more steps than it was trained for and refines its judge on these samples. Non-uniform sampling trains only the last step and hands it most of the information at sampling time. Refinement confidence keeping holds on to the tokens the student is already sure of. We train a fast class-conditional generator on ImageNet 256 and, with the same method, a 2B text-to-image model. On ImageNet 256, the 3-step FastGRN beats the FID of GRN. On a 2B binary-token text-to-image GRN, a 4-step student also nearly matches the GenEval of GRN-T2I 2B.

[CV-150] ReactBench-AD: Do Driving VLAs and World-Action Models React to Imminent Hazards in Their Path?

链接: https://arxiv.org/abs/2611.20011
作者: Nikos Theodoridis,Reenu Mohandas,Sushil Sharma,Ciarán Hogan,Ganesh Sistu,Anthony Scanlan,Tim Brophy,Ciarán Eising
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Vision-Language-Action models (VLAs) and World-Action Models (WAMs) have emerged in autonomous driving with the promise of strong generalization to long-tail scenarios. Although these models achieve state-of-the-art results on autonomous driving benchmarks, most existing benchmarks are dominated by nominal scenarios and therefore cannot test this promise. To test one such long-tail capability that is critical for safety, we introduce ReactBench-AD, a benchmark that measures how driving models react to objects that appear in the ego vehicle’s path. Built on real NAVSIM and PhysicalAI-Autonomous-Vehicles scenes, ReactBench-AD composites a crossing pedestrian, a standing pedestrian, or a stationary car consistently across all camera views and history frames. Each edited scene thus forms an exact, pixel-level counterfactual pair with its original, which allows us to isolate the causal effect of the inserted object on the predicted trajectory. Evaluating four VLAs and two WAMs, we find that they react weakly, often barely at all. Finally, we test how easily one of the evaluated VLAs, ReCogDrive, can be taught to brake hard for imminent hazards: a short fine-tuning stage on edited scenes substantially improves its reactions, with the learned behavior generalizing to unseen objects and fully real scenes, at minimal cost to nominal driving. These results call into question the out-of-the-box long-tail capabilities of these models, but suggest that they can be a strong foundation for acquiring them with little additional training, highlighting the value of real or synthetic data that demonstrates the long-tail behaviors nominal datasets rarely contain. Code and data will be made publicly available.

[CV-151] PlanDiv: Test-Time Scaling with Latent Plans for Vision-Language-Action Models

链接: https://arxiv.org/abs/2611.20010
作者: Hanbo Zhang,Lianlei Shan,Liuhaichen Yang,Zezhi Tang
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 10 pages, 2 figures

点击查看摘要

Abstract:Test-time scaling improves vision-language-action (VLA) policies by drawing several action candidates and executing the best one. We find that its returns diminish quickly when the candidates are resampled from one decoder: they differ by noise rather than by plan, and for each of three resampling sources, going from four to eight candidates adds only 0.06-0.07 to the average number of completed tasks on CALVIN. We introduce PlanDiv, an action head that makes the candidates differ by plan: a lightweight planner proposes K latent plan hypotheses, a shared masked-diffusion executor decodes one action chunk per plan from a single backbone pass, and the policy executes the chunk that its own likelihood ranks highest. The plans are learned from demonstrations alone, without plan labels or rewards. With the backbone, data and training recipe held fixed, eight plan-conditioned candidates raise five-task chain success on CALVIN ABC \rightarrow D from 56.9% to 69.9%, while eight candidates from autoregressive, plan-free or same-plan resampling reach at most 64.9%. The same-plan baseline is PlanDiv’s own checkpoint with the plan held fixed, so this gap comes from the plans, not from drawing more samples. On LIBERO-Long, four plans remove 40% of single-candidate failures, against at most 21% for matched resampling. On a real Franka arm the gain appears where a task admits more than one solution: PlanDiv is on par with a single-output OpenVLA-OFT+ baseline on a one-solution task and succeeds in 51 of 54 trials against 40 of 54 on two tasks with two solutions.

[CV-152] Glance then Select: Evidence-Tempered Visual Token Selection for Extremely Compressed Multimodal LLM s

链接: https://arxiv.org/abs/2611.20006
作者: Yuxuan Wang,Guangxiang Zhao,Yuxuan Tian,Yahui Han,Xinyu Tan,Feiyu Wang,Zongwei Lv,Xiangzheng Zhang,Lin Sun,Tong Yang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Multimodal large language models (MLLMs) encode a high-resolution image into thousands of visual tokens, which dominate the cost of the prefill stage. Visual token reduction removes most of these tokens, but two problems remain at extreme budgets. The vision encoder still processes every patch of the image, and the few surviving tokens often miss the small regions that hold the answer. In this paper, we propose Glance-ETD, a two-stage token reduction framework guided by a lightweight question-conditioned scorer. In the first stage, a glance encoder views the image at low resolution and re-encodes only the question-relevant regions at native resolution. In the second stage, Evidence-Tempered Diversity (ETD) selects the final tokens by combining diversity with the evidence predicted by the scorer, so that the kept tokens cover the scene and retain the answer. A single-rollout self-distillation recipe trains the MLLM with the glance encoder. With only 5% of the visual tokens, Glance-ETD retains 91.6% of the full-token performance of Qwen3.5-4B across fourteen benchmarks and outperforms the strongest prior method, LT-OPD, by 17.3 points and the same training recipe with DivPrune by 6.0. It also speeds up the time to first token on 4K images by 3.1x and generalizes to three other MLLMs and to high-resolution document benchmarks, with the largest gains on text-rich tasks.

[CV-153] What Matters for Cross-Modal Alignment

链接: https://arxiv.org/abs/2611.20000
作者: Yipeng Sun,Jiajun Wang,Siyuan Mei,Yan Xia,Siming Bayer,Vincent Christlein,Andreas Maier
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:The Platonic Representation Hypothesis posits that models trained on different modalities converge toward a shared representation. Related work aligns frozen image and text encoders on natural images without paired data. We test zero-pair alignment on five datasets across four modalities: natural images, environmental audio, chest radiographs, and 3D computed tomography (CT). It succeeds only on COCO and fails on the other four. We analyze the failures and find that alignment must first locate the subspace where the modalities co-vary and then orient the mapping within it. Consistent with previous studies, medical imaging is the hardest case: its shared subspace carries the smallest fraction of the variance. When this fraction is small, alignment needs image-text pairs, and random-matrix theory predicts how many pairs are needed. These findings motivate LightAlign: a few pairs let a ridge regression between two frozen encoders locate and orient the mapping simultaneously, and a classifier fitted only on report text then predicts findings from the mapped images. On 3D CT, LightAlign surpasses models trained end to end on image-report pairs without finding-level labels: with only 200 pairs, LightAlign outperforms CT-CLIP on CT-RATE (0.754 vs. 0.733 AUROC), and with only 100 CT-RATE pairs, it does so on RAD-ChestCT, an out-of-distribution dataset from a different hospital. On CheXpert, it reaches 0.862 mean AUROC with only about 5% as many pairs as models trained on MIMIC-CXR.

[CV-154] KneeFS-Net: A Lightweight Deep Learning Model for Knee Osteoarthritis Severity Assessment with Cross-Domain External Validation

链接: https://arxiv.org/abs/2611.19988
作者: Shafqat Alam,Galib Anjum Talukder Mahi,Yeasir Rabbani Tanvir,Md. Ibrahim,Md. Shahed Hossain,Shashank Ghimire,Mahmud Hasan Mostofa Kamal,ASM Shahidul Hossain,Taufiq Hasan
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 23 pages, 6 figures, 4 tables

点击查看摘要

Abstract:Knee osteoarthritis (OA) grading from radiographs using the Kellgren-Lawrence (KL) scale requires expertise that may be scarce. We present KneeFS-Net, a lightweight ordinal grader (0.8 million parameters; 0.33 GFLOPs) combining spatial, Fourier and state-space features with landmark-guided attention. On the Osteoarthritis Initiative (OAI) test set, KneeFS-Net achieved a quadratic weighted kappa (QWK) of 0.863, mean absolute error (MAE) of 0.325 and 97.2% within-one-grade accuracy, comparable to models with 9-35-fold more parameters. We evaluated cross-population generalization on two Bangladeshi cohorts, a single-center hospital cohort and a multi-center teleradiology cohort. Without further training or adaptation, the OAI-trained model achieved QWK 0.803, MAE 0.339 and 98.2% within-one-grade accuracy on its hospital cohort, declining to QWK 0.656 and MAE 0.522 on multi-center teleradiology images. Self-supervised pretraining on 6,092 unlabeled radiographs followed by joint fine-tuning achieved QWK 0.823, MAE 0.329 and 99.7% within-one-grade accuracy in cross-validation. These findings demonstrate accurate, efficient KL grading and potential adaptation to new populations with limited annotation.

[CV-155] Morphology-Decoupled Segmentation for Multiple Tunnel Defects Under Complex Environments

链接: https://arxiv.org/abs/2611.19981
作者: Mohan Ren,Hang Li,Yifei Liu,Shaowei Wang,James O’Kane,Xuanzhi Wang,Shiyu Zhao,Yumeng Zhang,Songtao Ji,Jurij Karlovsek
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 16 pages, 7 figures

点击查看摘要

Abstract:Automated tunnel-lining inspection must segment different defect geometries: cracks are thin, elongated, and connectivity-sensitive, whereas leakage, lining falling off, and segment damage are spatially extensive and context-dependent. Existing approaches generally apply a common decoder across defect classes or introduce boundary, topology, and directional mechanisms as isolated enhancements. This paper introduces Morphology-Decoupled Segmentation (MDS), in which defect morphology becomes the organising principle for feature routing, decoding, and interaction. A supervised routing module partitions multiscale encoder features into complementary linear and areal streams. The linear stream employs Deformable Strip Attention with image-conditioned direction hypotheses to preserve crack continuity without assuming fixed horizontal or vertical orientations, while the areal stream combines Pyramid Pooling with feature-pyramid fusion to capture diffuse boundaries and regional context. The representations are coupled through gated cross-morphology attention, allowing complementary information exchange while retaining morphology-specific features. A resolution-adaptive tiling and probability-blending pipeline preserves fine defects across variable-resolution imagery. MDS achieves 81.17% mIoU on the internal validation set and improves foreground image-level recognition over a backbone-matched SegDINO model, primarily for areal defects. Controlled ablations demonstrate the complementary contributions of routing supervision and cross-stream interaction while revealing a trade-off between areal-class performance and crack specialisation. The submitted model achieved 67.55% mIoU on an organiser-controlled hidden test, and the team placed second. Selective dynamic INT8 quantisation reduced model storage by 2.83 times with negligible accuracy loss.

[CV-156] CreaTrace: Video Creative Plagiarism Detection and Evidence Tracing

链接: https://arxiv.org/abs/2611.19976
作者: Chang Liu,Sheng Li,Zhenxing Qian,Xinpeng Zhang
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Video creative plagiarism refers to the unauthorized reproduction of creative ideas across videos despite substantial differences in visual appearance, posing a new challenge to video copyright protection. However, existing video copy detection methods primarily focus on visual content similarity, struggling to capture abstract creative that are entangled with specific visual realizations. To address this issue, we introduce a novel task, named Video Creative Plagiarism Detection (VCPD), and propose an iterative creative evidence tracing framework, termed CreaTrace, to verify creative similarity and accurately localize the corresponding evidence segments. Specifically, we design a Multimodal Semantic Event Segmentation strategy to divide similar long videos into coherent narrative events, facilitating event-level dataset construction and retrieving similar segments. Based on these events, CreaTrace employs Coarse Creative Matching to identify candidate event pairs through narrative semantics and camera language. Then CreaTrace presents Creative Decoupling to disentangle abstract creative patterns from specific visual realizations. Meanwhile, Creative Swap Contrastive Learning is devised to enhance creative consistency through cross-realization feature swapping. Benefiting from the above designs, our CreaTrace enables accurate creative similarity detection and temporal localization, while providing interpretable evidence report. Extensive experiments demonstrate the effectiveness and robustness of CreaTrace across diverse creative plagiarism scenarios.

[CV-157] Multimodal AI in Healthcare: Review of Vision-Language Foundation Models for Real-World Medical Applications

链接: https://arxiv.org/abs/2611.19958
作者: Taha Razzaq,Murtaza Taj,Asim Iqbal
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 20 Pages, 3 figures, 3 Tables

点击查看摘要

Abstract:The emergence of foundation models has marked a transformative shift in AI, enabling robust generalization across diverse downstream tasks through putative zero-shot learning. Large Language Models and Vision Language Models have demonstrated strong capabilities in tasks such as image interpretation, report generation, and question answering by effectively learning from multimodal data-images paired with associated text often with minimal supervision. In the healthcare domain, this ability to align visual and textual information reduces the reliance on extensive manual annotations, as models can leverage existing clinical reports and imaging data to learn meaningful representations. This integration holds promise for improving diagnostic support, treatment planning, and overall patient care, even in data constrained settings. In this review, we provide a definitive taxonomy of the medical VLM landscape, tracing the evolution from early Contrastive Alignment and Generative MLLMs to the cutting edge frontiers of Dense Pixel Grounding, Sparse Mixture-of-Experts (MoE), and Reasoning Incentivized (RL) architectures. We critically examine the “medical bottleneck” identifying the persistent challenges of data scarcity, the “hallucination” risks in generative diagnostics, the computational strain of 3D volumetric processing, and the lack of standardized, clinically grounded evaluation metrics.

[CV-158] FinVision: Multimodal AI for Visualizing Financial Narratives

链接: https://arxiv.org/abs/2611.19956
作者: Yixuan Tang,Hui Wang,Yi Yang,Allen H. Huang
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Financial narratives contain rich information about firms’ performance and business conditions. Conventional text analysis often summarizes this information into numerical representations, which may not fully preserve the dynamic structure described in the narrative. In this study, we develop FinVision, a domain-specific multimodal generative model that translates financial narratives into visual representations of contemporaneous market dynamics. Using 87,940 paired observations of Management Discussion and Analysis (MDA) disclosures and stock price charts, FinVision learns the average relationship between a firm’s narrative account of a fiscal year and its stock-price dynamics over the same period. Given an MDA alone, the trained model generates a text-implied stock chart, extending financial text analysis beyond numerical representations toward text-to-visual generation. We validate FinVision in two ways. First, held-out tests show that the model generates stock-chart representations that closely resemble observed charts and outperform direct LLM prompting, general-purpose text-to-image models, and the same architecture without financial-domain adaptation. Second, we examine the economic validity of the generated representations by comparing text-implied and observed stock charts. The resulting discrepancy is systematically associated with weaker investor responses around subsequent 10-K filings and with worse subsequent market and operating outcomes. Overall, our findings show how domain-adapted multimodal generative AI can create a new visual representation of financial text and provide an additional modality for studying business communication.

[CV-159] wo-Stage Prompt Learning: A Unified Framework for Person ReID without Clothing Hypothesis

链接: https://arxiv.org/abs/2611.19939
作者: Yuxuan Xiao,Shanshan Zhang,Shengcai Liao,Jian Yang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Person re-identification (ReID) in real-world scenarios involves three clothes conditions: the same clothes (SC), changed clothes (CC), and the unknown (GEN) scenario where clothes changes are not specified. Existing methods typically train separate SC- or CC-specific models, which limits generalization across all conditions. Meanwhile, pre-trained visual-language models (VLMs) such as CLIP have shown strong generalization ability in various vision tasks, motivating recent ReID methods to incorporate textual prompts. However, to distinguish SC and CC, prior approaches rely on fixed predefined instructions, which lack learnable context embeddings in prompts and cannot effectively handle the GEN scenario. To overcome these limitations, we propose Two-Stage Prompt Learning (TSPL), a unified framework that integrates clothes-identity decoupling and scenario-aware task prompts within VLMs. In Stage 1, appearance-decomposed prompt learning disentangles clothes and identity information under frozen encoders, producing stable embeddings where clothes cues can be selectively preserved or suppressed. In Stage 2, task-aware prompt learning trains the image encoder with learnable context embeddings guided by Stage-1 representations, enabling robust adaptation across SC, CC, and GEN conditions. Extensive experiments on five benchmarks show that TSPL achieves state-of-the-art performance under all scenarios, improving mAP by up to 9.2% on LTCC (SC) and by 6.4% on VC-Clothes (GEN), respectively. The code will be released at this https URL.

[CV-160] SAGE-VQA: A Training-Free Self-Evolving Agent ic Framework with Grounded Evidence for Knowledge-Based Visual Question Answering

链接: https://arxiv.org/abs/2611.19933
作者: Ruirui Shen,Jianqi Gao,Xiangfeng Luo,Yonggang Zhang,Muran Yu,Barney Haoyun Miao,Hang Yu,Yike Guo
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Knowledge-based visual question answering (KB-VQA) aims to combine visual information with external knowledge to answer user questions. Existing methods face two main bottlenecks: visual perception errors can propagate into hallucinations, and generalization across scenarios remains limited. To address these issues, we propose SAGE-VQA (Self-evolving Agent framework with Grounded Evidence), a training-free agentic framework that combines evidence verification with experience evolution. Our key insight is to explore an Evidence-Driven Adaptive Reasoning (EDAR) strategy. It accurately acquires and verifies question-relevant fine-grained visual information. It also extracts and refines reusable experience to support robust reasoning across tasks. Specifically, we design a Co-verified Visual Grounding module that verifies question-relevant visual information and provides grounded evidence for subsequent reasoning. We further introduce a Self-evolving Hierarchical Memory (SHM) mechanism. SHM organizes experiences into declarative and procedural memories based on their nature. These memories can be reused in subsequent reasoning and continually refined through feedback, without updating model parameters. Experimental results show that our method outperforms competing methods based on lightweight open-source vision-language models (VLMs) as well as those based on closed-source VLMs. Moreover, our method can serve as a plug-and-play reasoning framework for existing VLMs without modifying their architectures.

[CV-161] Does synthetic pre-training pay off on a microcontroller? Accuracy latency and energy across two MCUs

链接: https://arxiv.org/abs/2611.19931
作者: Achintya Yadav
类目: Machine Learning (cs.LG); Hardware Architecture (cs.AR); Computer Vision and Pattern Recognition (cs.CV)
备注: 43 pages, 5 figures, 14 tables; supplementary tables, figures and notes as ancillary files. Analysis plan registered on OSF (doi: https://doi.org/10.17605/OSF.IO/BPQGA ). Under review

点击查看摘要

Abstract:Rendered images are cheap to label, but a microcontroller classifier must survive 8-bit integer (INT8) quantisation and compute correctly on the chip. Five models, from MobileNetV2 at 224 px to a 61K-parameter convolutional neural network (CNN), were fine-tuned on a six-class scene dataset with and without a synthetic stage, against thresholds fixed in advance, and deployed on an ESP32-S3 and an RP2040. With full real data the stage gave no detectable gain. For the main deployable model the INT8 difference was +0.312 percentage points (pp), its interval below the +1.6 pp gain recomputed from the source study; quantisation did not detectably change it, nor did model size. With 10% of the real data the stage paid off for the CNN, by +3.438 pp on one seed and +2.715 pp over three; its cause is open. On the ESP32-S3 the main model took 152.7 ms and 16.3 mJ above idle per inference, and adding the stage changed neither beyond preset bounds. The model’s INT8 outputs matched TensorFlow Lite’s reference kernels exactly, but only after a memory overrun in a vendor kernel was patched. On the RP2040 the same model took 1.80 s and 27.3 mJ above idle, about twelve times slower, and the largest 96 px model exceeded the board’s flash-capacity limit. Neither the RP2040 nor the host interpreter used for scoring reproduced the reference arithmetic. A synthetic stage is worth adding only when real data are scarce, and on-device INT8 outputs should be checked against reference kernels, not inferred from host scores.

[CV-162] Lao Siji: Adapting openpilot to Closed-Loop End-to-End Driving in Neural Reconstruction

链接: https://arxiv.org/abs/2611.19922
作者: Chengzhi Gao,Yongzhi Liu,Jinchang Xu,Weichao Zhuang,Guodong Yin
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: 9 pages

点击查看摘要

Abstract:We present Lao Siji, an end-to-end driving policy that ranks third in the 2026 NVIDIA AlpaSim E2E Driving Challenge, ahead of most entries built on foundation models. Instead of training from scratch on a general-purpose vision backbone, we adapt openpilot, an open-source L2 production model. We freeze its vision encoder and fine-tune the policy layers with an ego-state adapter and a route memory. On closed-loop rollouts in neural-reconstructed scenes, we supervise the policy with a privileged map teacher for lateral self-correction and a body-extent loss for road and obstacle clearance, while distilling the base model’s remaining outputs to retain car following. Ablations show that self-correction samples carry the gain, while the clearance loss substitutes for them where no teacher exists. Longitudinal safety instead rests on the base model’s lead-vehicle estimate behind a rule-based speed layer. Longer training, imitating human logs, and human-log speed targets yield either no measurable gain or adverse effects. Without further training, Lao Siji also places among the leaders on HUGSIM and Odyssey. The final training takes only one GPU-hour and shifts the weights by 3.9% from the shipped model. Code will be released after review.

[CV-163] HEmIF-Bench: Benchmarking HE-to-mIF Models BeyondVirtual-StainFidelity

链接: https://arxiv.org/abs/2611.19905
作者: Fuxian Zhao,Hyokeun Park
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 11 pages, 3 figures, 2 tables

点击查看摘要

Abstract:Hematoxylin and eosin (HE)-to-multiplex immunofluorescence (mIF) generators predict spatial protein expression from routine histology, yet pixel-level marker reconstruction does not establish downstream utility. We present HEmIF-Bench, evaluating five frozen generators across eight datasets spanning molecular, cellular, tissue, and patient scales. HEmIF-Bench decouples internal encoder features from generated outputs, benchmarking both against the pathology foundation reference UNI2. Mirroring the clinical workflow where pathologists rely on mIF to define cell phenotypes, delineate tertiary lymphoid structures (TLSs) and germinal centers (GCs), and stratify patient risk, we evaluate external CD3 recovery, single-cell classification, dense nuclear segmentation, TLS/GC segmentation, and patient survival prediction. Across the benchmark, foundation-model-based MIPHEI-ViT leads among the generators, while non-foundation models show task-specific strengths. Comparing the two routes reveals that internal encoder features generally outperform generated outputs, and larger output marker panels do not guarantee stronger downstream utility across datasets. Moreover, neither internal encoder features nor generated outputs consistently surpass frozen UNI2. These findings show that pixel-level stain fidelity does not equate to downstream utility, and that generators trained at current paired-data scales have yet to surpass large-scale pure-HE foundation models. With the expansion of paired translation cohorts and the advancement of foundation-based generative architectures, HEmIF-Bench establishes a definitive benchmark for tracking when HE-to-mIF generation crosses the threshold to deliver genuine value beyond routine histology for tumor microenvironment analysis and precision oncology.

[CV-164] Remember Replay Adapt: Learning from Past Experience in Online Vision-and-Language Navigation

链接: https://arxiv.org/abs/2611.19901
作者: Chenhao Ding,Jizhou Han,Lixin Xiu,Muyao Wang,Songlin Dong,Hideki Nakayama
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Existing feedback-based methods for online vision-and-language navigation (VLN) primarily learn from current episode outcomes, leaving previously acquired feedback underutilized. Inspired by human episodic memory, we propose R2A, an online VLN adaptation framework that continually reuses successful and failed navigation experiences for policy learning. Specifically, Success-Failure Memory (SFM) retains successful and failed navigation episodes with acquired feedback and retrieves task-relevant experiences based on incoming instructions, enabling historical feedback to guide subsequent navigation. Prior-Anchored Adaptation (PAA) learns from these experiences by reinforcing successful trajectories and suppressing failed ones relative to a frozen pretrained policy, while regularizing behavioral drift. Together, these modules enable continual feedback reuse and stable policy adaptation without source-domain data or expert supervision. Extensive experiments across three benchmarks, REVERIE, R2R, and R2R-CE, and five navigation backbones demonstrate consistent performance gains. On the REVERIE validation unseen split, R2A improves DUET’s SR from 46.98% to 69.84% and SPL from 33.73% to 50.98%.

[CV-165] HIVE-ECIA: Identity-Aware Historical Evidence Acquisition for Bandwidth-Efficient Multi-Object Tracking

链接: https://arxiv.org/abs/2611.19897
作者: Hai Nguyen,Tarik Taleb,Qize Guo
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Panoramic multi-object tracking (MOT) requires reliable identity association across a 360° field of view, yet bandwidth-constrained operation forces a trade-off between low-resolution (LR) identity ambiguity and costly high-resolution (HR) transmission. We propose HIVE-ECIA, an identity-aware framework that selectively retrieves historical HR evidence to recover fragmented identities without modifying detection boxes or the underlying tracker. An ECIA Base-Guided query policy determines when additional HR evidence is warranted; HIVE ranks historical LR observations to select an informative retrieval timestamp; and ECIA Full integrates the retrieved HR crop through learned gated fusion and conservative identity safeguards. We further introduce DualPano-MOT, a real-world dual-fisheye panoramic evaluation dataset featuring controlled pedestrian interactions, panorama-seam challenges, and occlusions, with model-assisted, manually verified tracking annotations. On held-out QuadTrack sequences and zero-shot DualPano-MOT, HIVE-ECIA improves identity association across four MOT trackers without tracker-specific retraining. Across these trackers, identity switches decrease by up to 21.8% and 23.4%, while the largest IDF1 gains reach 3.81 and 6.74 percentage points, respectively. Under an offline JPEG-based acquisition model, our framework reduces encoded data volume by 89.2% and 79.3% relative to continuous Full-HR acquisition.

[CV-166] Illusion of Explainability: A Sanity Check for Spectral Band Attribution in Satellite Time-Series Crop Classification WACV2027

链接: https://arxiv.org/abs/2611.19889
作者: Sumit Tahiliani,Avi Mathur,Sandarbh Agarwal,Rakshith Sathish,Ashwathi Nambiar
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: Accepted at the 4th Computer Vision for Earth Observation (CV4EO) Workshop at WACV 2027

点击查看摘要

Abstract:Attribution methods are commonly validated through convergence across methods and consistency with domain knowledge, yet neither criterion establishes whether identified features are necessary for a task. This limitation is particularly relevant in multispectral remote sensing, where overlapping spectral information makes individual bands potentially substitutable. We investigate this gap in crop-type classification using GradientSHAP, Shapley sampling, and Occlusion on a U-TAE segmentation model trained on Sentinel-2 time series from the PASTIS benchmark. All three methods identify vegetation-sensitive spectral regions consistent with domain knowledge, while Shapley and Occlusion exhibit strong ordering sensitivity in deletion and insertion tests. However, these properties do not reliably predict retraining performance. Aggregate attribution consensus masks substantial per-class disagreement, with pairwise mean Spearman correlations ranging from 0.09 to 0.79. Models retrained from scratch using the five lowest-ranked bands perform comparably to or better than those using the five highest-ranked bands. Furthermore, attribution rankings from an independently trained model with identical configuration and seed identify a different top-five subset that yields better retraining performance. Analysis of all 252 five-band subsets reveals consistently low effective dimensionality (1.35 to 2.62), suggesting substantial spectral redundancy. These findings demonstrate that plausible and apparently faithful attribution rankings do not necessarily identify task-relevant bands for improved retrained performance, highlighting the limitations of band-level feature attribution in spectrally redundant input spaces.

[CV-167] PanDWM: Panoramic 4D Driving World Models

链接: https://arxiv.org/abs/2611.19857
作者: Jiajun Sun,Zeyu Zhang,Zeyu Ren,Guangfeng Jiang,Yang Zhao,Hao Tang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Code: this https URL , Website: this https URL

点击查看摘要

Abstract:Driving world models usually generate videos from six pinhole cameras with separate optical centres and small overlaps, and must learn to keep these views consistent. We present PanDWM, which stitches them into one 360^\circ panoramic band. Camera boundaries then fall inside one image, so keeping the views consistent means keeping the image continuous, and each pixel still records its camera ray. Conditioned on layouts, a causal student generates the band window by window in four steps; it is distilled from a causal teacher and trained on its own rollouts. A feed-forward decoder lifts each frame along these rays into 3D Gaussians, and static Gaussians from neighbouring frames complete them into 4D scenes. On nuScenes, scored in one stitched 360^\circ view, PanDWM has lower FID and FVD than the multi-camera baselines at 8 - 46\times their speed, and keeps its 50-step teacher’s FID while lowering FVD by 22-25%.

[CV-168] Passable for Whom? Few-Shot Personal Calibration of Sidewalk Accessibility Perception

链接: https://arxiv.org/abs/2611.19839
作者: Wesley Maia,Angel Martinez-Sanchez,Maitrayee Keskar,Ross Greer
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Whether a sidewalk is passable depends on who is trying to pass: people who use the same mobility aid disagree about the same image, so one label per scene, or per aid, is wrong for many of them. We introduce few-shot personal accessibility perception: from K answers of a person never seen before, predict their judgment of sidewalk scenes never seen before. PassCal separates a shared description of the scene from the person’s interpretation of it. The description comes from the internal state of a frozen vision-language model (VLM) and from barrier type, severity and context properties predicted by a detector trained on 14.9k public audits; the interpretation, a calibration of the person’s response style and a weighting of the properties, is fit from their answers. On 8,883 judgments in which each of 135 verified mobility-aid users rated the same 50 sidewalk scenes, with people and scenes both held out, eight answers lower the Brier score from 0.550 to 0.479. The same VLM prompted with those answers reaches 0.562, and the same calibration without the scene 0.521. Conformal risk control turns predictions into passable, blocked or “depends on you” statements with a bounded expected rate of wrong statements over people; at 5%, eight answers raise the share of decisive scenes from 27% to 42%. Audit-supervised properties improve all sixteen vision and VLM encoders we test and outperform attributes from a sidewalk navigation VLM.

[CV-169] AGPF: Dynamic Source Selection for Feature Pyramid Fusion

链接: https://arxiv.org/abs/2611.19836
作者: Xu Yang,Gady Agam
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Feature pyramids provide dense predictors with features at several spatial resolutions. However, it remains largely unexplored which pyramid levels each output location should aggregate information from. We investigate this question by allowing each output location to attend to features from every pyramid level and measuring how attention is distributed across scales. Surprisingly, locations on the finest output grid allocate approximately 75% of their attention to features two levels coarser, rather than to the current or adjacent levels. Even on the same output grid, the balance shifts toward coarser features as object size increases. These observations motivate Asymmetric Global Pyramid Fusion (AGPF), a neck that preserves the original output grids while allowing each location to dynamically select its contextual sources across scales. Each location interacts with a multiscale memory whose features are aligned in image coordinates, enabling flexible cross-scale aggregation without altering prediction resolutions. Compared with FPN, AGPF improves all 19 object detection and all seven instance segmentation configurations evaluated on COCO. A single fusion block improves Cascade Mask R-CNN by 1.7 box AP and 2.3 mask AP, and the gains consistently transfer to semantic and panoptic segmentation. Subjects: Computer Vision and Pattern Recognition (cs.CV) Cite as: arXiv:2611.19836 [cs.CV] (or arXiv:2611.19836v1 [cs.CV] for this version) https://doi.org/10.48550/arXiv.2611.19836 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[CV-170] Beyond Vehicle-Centric Data: Scaling Autonomous Driving through Roadside-to-Vehicle Generation

链接: https://arxiv.org/abs/2611.19830
作者: Yichen Zhang,Zihan Wang,Hao Chen,Runyi Huang,Ruixue Zong,Jing Wang,Yunong Mao,Lei He,Keqiang Li
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Training autonomous driving models requires large and diverse datasets, collected primarily by vehicles. Scaling them entails investment in fleet operation, data processing and annotation. Road infrastructure records traffic and multi-vehicle interactions, yet the viewpoint gap between roadside and onboard sensors limits its direct use for training. Here we introduce AnyR2V, a framework that transforms multimodal roadside observations into seven-camera surround-view videos for different vehicle perspectives. It aggregates sequential observations into a shared scene representation of geometry and road-user motion, then renders appearance, depth and layout conditions along target-vehicle trajectories. Supervised by spatiotemporally aligned roadside-vehicle pairs, AnyR2V generates surround-view videos guided by scene structure and road-user motion. Existing roadside 3D bounding boxes are transformed into target-vehicle coordinates and paired with generated images, allowing one traffic recording and its annotations to yield training data from multiple vehicle perspectives. We construct R2V360, a dataset of 89 real traffic sequences at an urban intersection, with aligned multimodal roadside data and onboard surround-view videos. AnyR2V outperforms six scene reconstruction methods in image fidelity and in depth and semantic agreement between generated and real onboard images. We further evaluate generated videos along other vehicles’ trajectories by adding them to real training data. Adding generated frames equal to 50% of the original real training frame count improves mean average precision (mAP) by 3.3-3.4 AP-points across three 3D object detectors on the real onboard test set, relative to real-data-only baselines. These results demonstrate that roadside traffic records can provide training data from multiple vehicle perspectives, offering a new data source for autonomous driving.

[CV-171] PRISM: Primitive-Integrated Separable Modulation with Priority Visual Memory Queue for Compositional Zero-Shot Learning

链接: https://arxiv.org/abs/2611.19828
作者: Jianhua Wang,Zhihang Zheng,Yang Liu,Jiale Du
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Compositional Zero-Shot Learning (CZSL) recognizes unseen attribute–object pairs by recombining primitives from seen compositions. Existing methods stop sharing evidence at deployment: fixed prototypes cannot exploit unlabeled test streams, while composition-wise refinement cannot propagate corrections across pairs sharing a primitive. We introduce PRISM, an online CZSL framework that extends primitive sharing to test-time adaptation. Its Hierarchical Knowledge Accumulation Module (HiKAM) factorizes prototype migration into composition-, attribute-, and object-level residual banks. A Priority Visual Memory Queue (PVMQ), initialized by Prototype Queue Initialization (PQI), grounds text updates with bounded visual evidence. The same adaptive core supports exhaustive Full PRISM and recall-calibrated Adaptive PRISM-Lite, which selects candidate budgets from held-out ranks. Comprehensive closed- and open-world experiments on UT-Zappos, MIT-States, and C-GQA demonstrate the state of the art, with Full PRISM achieving the best open-world AUC and harmonic mean on all three benchmarks and the best closed-world results on UT-Zappos and C-GQA. Fine-grained long-tail and non-stationary primitive-block evaluations reveal gains on under-supported concepts and stable adaptation under prolonged skew. Adaptive PRISM-Lite further delivers substantial efficiency gains in the full open-world space with only a small accuracy cost. PRISM thus combines primitive-shared online adaptation, multimodal grounding, and candidate-aware inference.

[CV-172] When Does Reusing Layers Help Multimodal Embeddings?

链接: https://arxiv.org/abs/2611.19827
作者: Zhizhang Hu,Yue Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:We ask whether a compact universal multimodal embedding models can produce better search embeddings by running the same layers again. Looped transformers add computation with nearly unchanged model size, making them an appealing alternative to a larger encoder. Their success in language-model reasoning and recent use for retrieval motivate the idea, but higher retrieval scores alone do not reveal what an extra pass contributes. We study eleven loop designs and training variants on three checkpoints of one vision-language architecture. Our protocol distinguishes whether a model uses its extra pass, whether its image states respond to the query, and whether those updates improve retrieval. We prove that an image placed before the query remains query-independent through repeated layers when information only travels forward. We then open a query-to-image path in three ways, including order-alternating recurrence, which places the query first on the next pass. Each path produces query-sensitive image states. For bidirectional attention and order alternation, freezing later visual updates does not lower accuracy. On all 78 datasets of the Massive Multimodal Embedding Benchmark version 2 (MMEB-V2), a standard added loop loses 0.46 points against its trained control; an intermediate generative stage changes the effect to +0.36. A model pretrained with recurrence gains +4.23 points over its own loop-removed control and benefits from query-conditioned image updates when answering visual questions or locating objects from text.

[CV-173] he Flip Test: Video World Models Fail to Learn the World State

链接: https://arxiv.org/abs/2611.19822
作者: Mi Luo,Lijun Yu,Adams Wei Yu,Ang Cao,Baochen Sun,Ming-Hsuan Yang,Nan Hua,Yuan Hao,Alex Dimakis,Ya Xu,Ed H. Chi,Xuezhi Wang,Yinxiao Liu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Robotics (cs.RO)
备注: Project Page: this https URL

点击查看摘要

Abstract:Video world models promise to predict the future of a scene from its present. Do current models keep this promise? We probe this question counterfactually. When a small change to the present flips the future, such as a box resting a few millimeters before or past a table edge, a model that tracks the state of the world must flip its prediction too. We build FlipTest, a three-level hierarchy of increasingly strict tests over 1,680 initial frames from 10 simulated and 15 real-world tasks. At the base, visual minimal pairs, two nearly identical frames whose true futures differ, test whether predictions change across the boundary; matched controls go further and test whether they change only where the future does; and dense sweeps, the strictest level, locate exactly where they change. All 11 video world models we evaluate fall short. The best averages 18.3% pair accuracy over real and simulated scenes, below the 25% of random guessing. Models flip as often on steps that leave the future unchanged as on steps that change it, and their response stays flat across the sweeps. On the same frames, a vision-language model is selective and traces a near S-shaped curve near the boundary, and a coding agent that measures the image comes close to the true boundary. This implies that the deciding information is in the frame, and video models just miss it. We therefore urge the community to rethink world modeling as pure video prediction. Next-frame prediction must do more than draw plausible pixels; it must also identify the causal states that control future dynamics.

[CV-174] Pros4D: Modeling Evolving 4D Worlds in Latent Space for Dynamic Spatial Reasoning

链接: https://arxiv.org/abs/2611.19803
作者: Rongxin Gao,Dongxuan Liu,Chu Li,Chengyang Ji,Jialing Huang,Runyu Chen,Kairun Wen,Chenxin Li,Jingrui Ye,Haosong Zhen,Xinghao Ding,Yue Huang,Yuzhi Huang,Xiaodong Tu,Jingyan Jiang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Project page: this https URL

点击查看摘要

Abstract:Dynamic spatial reasoning from monocular video enables vision-language models (VLMs) to understand how 3D environments evolve over time, supporting physical world understanding and embodied intelligence. Existing efforts primarily address observed-dynamics understanding, whereas agents in dynamic environments also require future-dynamics prediction beyond the observation boundary. Moreover, external geometric processing and textual descriptions can increase inference complexity or lose information when representing continuous dynamics. These limitations motivate a shared internal state that preserves observed dynamics and remains predictive beyond the input video. Accordingly, we present Pros4D, a framework for modeling evolving 4D worlds in latent space that unifies retrospective understanding of observed-dynamics with prospective reasoning for future-dynamics prediction. Specifically, Pros4D-Engine derives geometry-grounded supervision and constructs paired examples for both tasks, while Pros4D-Bench evaluates them across eight spatial task categories spanning dynamic cameras, objects, and scenes. At the model level, Pros4D compresses observed dynamics into latent observed states and autoregressively rolls them forward into latent future states. 4D Latent Alignment (4DLA) grounds these states through complementary visual-semantic and geometric supervision, while joint language supervision connects them to task reasoning. Experiments across benchmarks show consistent improvements over strong baselines in both capabilities.

[CV-175] Fine-Grained RLAIF for Visual Safety: Learning to Zoom into Hazards for Multimodal Safety Alignment

链接: https://arxiv.org/abs/2611.19786
作者: Chunxue Xu,Yiwei Wang,Yujun Cai,Songze Li
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Recent multimodal safety alignment heavily relies on a text-centric paradigm, assuming risks are explicit in language or globally prominent in images. However, real-world hazards are often visually subtle or peripheral, causing existing models to fail due to their reliance on coarse, global visual representations. In this work, we study fine-grained visual safety alignment and introduce a challenging 1,200-sample benchmark across four domains, where hazards necessitate localized inspection. To address this, we propose Fine-Grained RLAIF (FGR), a reinforcement learning framework that casts image zoom-in as a learnable perception policy. FGR employs a decomposed reward design to jointly align three core safety behaviors: (i) where to look via a visual evidence reward, (ii) when to act via a tool-formatting constraint, and (iii) how to reason via visually grounded safety explanations. FGR achieves accuracy ranging from 86.71% to 93.00%, consistently outperforming strong multimodal and agent-based baselines.

[CV-176] A Threshold-Free Visual-Semantic Matching Metric for Evaluating Consistency in Subject-Driven Image Generation

链接: https://arxiv.org/abs/2611.19780
作者: Bowei Shan,Heng Yin
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 16 pages, 5 figures

点击查看摘要

Abstract:In subject-driven text-to-image generation, quantitatively evaluating localized visual consistency across varying poses and environments remains a major challenge. Existing metrics either rely on global image embeddings (CLIP, DINO) and Vision-Language Models (VLMs) that pool spatial representations and miss fine-grained textural artifacts, or depend on scalar threshold-gated dense metrics such as Visual-Semantic Matching (VSM) that suffer from point-alignment dropouts, severe threshold sensitivity, and score instability. To overcome these limitations, we propose TF-VSM, a fully continuous, threshold-free visual-consistency metric paired with an extended feature-disentanglement retraining schedule for dual-branch ResNet feature-aggregation networks. Rather than enforcing rigid binary cutoffs (T_s,T_v) , TF-VSM weights every object point by a non-linear semantic match confidence w_j=ReLU(\hats_j)^2 and continuously integrates visual similarity over all consistency levels r\in[0,1] , yielding a closed-form confidence-weighted score without any hand-tuned hyperparameters. On a controlled 100-pair benchmark with ground-truth inpainting oracles, TF-VSM attains Pearson 0.638 and Spearman 0.674 oracle correlation, improving over the pretrained VSM baseline ( 0.454/0.589 ) by +0.171/+0.116 after retraining, and drastically outperforming CLIP ( -0.052/-0.004 ), DINO ( 0.087/0.121 ), and VLMs ( 0.096/0.003 ). Six diagnostic ablations confirm that smooth confidence weighting is the right mechanism, that hard semantic gating is brittle (dropping to 0.533 Pearson at T_s=0.8 ), and that the threshold-free metric is essentially invariant to the training-time loss balance \alpha that the thresholded metric still depends on.

[CV-177] When Backdoors Can Be Switched Off: Suppression-Aware Backdoor Attacks on Vision-Language-Action Models

链接: https://arxiv.org/abs/2611.19774
作者: Nur Afsa Syeda,Mohamed Elmahallawy
类目: Computer Vision and Pattern Recognition (cs.CV); Cryptography and Security (cs.CR)
备注:

点击查看摘要

Abstract:Vision-Language-Action (VLA) models map visual observations, robot states, and natural-language instructions to robotic actions, but their multimodal interfaces introduce new security risks. Existing VLA backdoor attacks primarily study trigger-based activation and stealthy action manipulation, leaving conditional suppression of malicious behavior underexplored. We investigate suppression-aware backdoors, which combine conjunctive trigger activation with instruction-conditioned suppression. We instantiate this design through three variants: Cross-Modal Trigger, combining robot-state and spatially separated visual triggers; Self-Suppressing Trigger, requiring a state trigger and a [DISRUPT] instruction token for activation, with [RESTORE] token suppressing the attack; and Multi-Trigger, additionally requiring a visual trigger while retaining suppression even when physical triggers persist. We evaluate these attacks using SmolVLA on two SO-100 manipulation datasets against four baselines: State Backdoor, BadVLA, TabVLA, and SilentDrift. Self-Suppressing and Multi-Trigger reduce Isolation Forest detection from 96–98% to 1–18% and action-space l divergence by 10–14X, achieving 63–71% attack success rates with perfect clean-task performance. Experiments with OpenVLA on LIBERO further show that the attacks transfer to a 7B-parameter policy, evading STRIP and action-distribution-shift detection. Our findings establish conditional suppression as an important dimension for evaluating backdoor threats in VLA policies.

[CV-178] SemLite-GS: Semantic-Selective Parallel Gaussian Splatting for Lightweight Building Reconstruction from UAV Oblique Imagery

链接: https://arxiv.org/abs/2611.19773
作者: Tengfei Wang,Xin Wang,Zihan Li,Quanjian Ji,Yifei Yu,Dong Lin,Zongqian Zhan
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Lightweight 3D building models support urban spatial analysis, but their production often requires dense scene reconstruction, building extraction, and mesh simplification. Although 3D Gaussian Splatting (3DGS) is more efficient than conventional multi-view stereo, existing large-scale methods lack semantic selection and parallel instance-level training. We present SemLite-GS, a building-centric framework that reconstructs lightweight building meshes from large-scale UAV oblique imagery. First, an orthographic reference map establishes globally consistent building identities, which are propagated to oblique views to form camera-instance pairs and regions of interest. Second, semantic-selective parallel PGSR training jointly processes multiple pairs within relevant image regions while updating only target Gaussians; contextual Gaussians remain fixed to preserve visibility. Third, architectural planes extracted from rendered depth, opacity, and normal maps guide signed distance field optimization, producing meshes with adaptive polygon allocation and well-defined structural boundaries. Experiments on multiple large-scale UAV datasets demonstrate that our SemLite-GS substantially reduces building reconstruction time compared with representative full-scene 3DGS methods. With comparable polygon budgets, SemLite-GS achieves superior preservation of dominant planar structures and sharp architectural transitions compared with conventional dense reconstruction followed by mesh decimation. The source code is available at this https URL.

[CV-179] JEPA4R: Embodied JEPA World Models with 4D Point Planning

链接: https://arxiv.org/abs/2611.19771
作者: Jiarui Dong,Zhiheng Liang,Yanjun Li,Zeyu Zhang,Heejeong Nam,Haoyu Chen,Hao Tang
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: Code: this https URL , Website: this https URL

点击查看摘要

Abstract:Joint-embedding predictive architecture (JEPA) world models predict future embeddings instead of pixels and plan in latent space. Two challenges remain: 2D images leave the metric geometry that manipulation acts on implicit and entangled with appearance, and what these models imagine stays hidden, as they are mostly judged in simulation by short-horizon latent planning. We present JEPA4R, a JEPA world model that observes single-depth-camera point clouds, predicts and plans in their latent space, and acts through a policy head on its frozen latents; we call this 4D point planning. A decoder on the frozen world model shows what it imagines as 4D point clouds, with a mean point error within 1.1 cm over 200 steps. Under the policy-learning protocols of OGBench and CortexBench, JEPA4R leads pixel JEPA models on OGBench Cube-Single, Cube-Double, Cube-Triple and Scene, and ranks first on MetaWorld and Adroit Relocate with 98.4% and 88.0%, above the runner-up, R3M, at 96.0% and 65.3%. Trained only in a digital twin, JEPA4R and its policy heads run zero-shot on the real arm with both a gripper and a dexterous hand.

[CV-180] When Does an AUROC Comparison Survive a Change in Population Mix?

链接: https://arxiv.org/abs/2611.19762
作者: Mahtab Movahhedrad,Hamidreza Hasani Balyani,Seyed Pouyan Mousavi Davoudi
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注: 30 pages, 5 figures. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:Does an AUROC comparison survive a change in the proportions of tasks, populations, or image corruptions? A pooled-AUROC ranking of failure detectors or uncertainty scores describes one particular mix of groups. We study two fixed scores for a common binary target under mixtures of fixed group-conditional laws. Using the cross-group AUC decomposition, we prove that every rational symmetric quadratic comparison geometry is realizable by finite scalar-score distributions. Universal dominance is thus coNP-hard, and even exhaustive checks of every mixture on fewer than K groups can miss a K -group reversal. A grid bound gives a dimension-uniform approximation certificate with sharp 1/m order. Common score marginals across groups make the comparison numerator linear, with an error envelope for imperfect alignment. These results yield exact finite-bank audits, explicit reversal mixtures and distances, and simultaneous population statements. For failure detection with a frozen ResNet-50 on ImageNet-C, 15 of 18 confidence-score comparisons win in every corruption family on a discovery split, and all 15 are finite-bank certified over all family mixtures on a held-out split; over all 15 corruption types, of 14 discovery winners the corrected grid certifies 10, leaves 3 unresolved, and finds one confirmation reversal of at least 0.05 AUROC. An audit of all 11,019 released Stable LM 2 12B outputs and 35 uncertainty methods finds 66 confirmation-set comparisons that win within all four tasks yet reverse under mixing. Discovery-fitted marginal rank alignment removes all 66; an alignment-error envelope certifies 83 comparisons. Label checks also change several selected reversals. The findings distinguish robustness of frozen scores from robustness after calibration; they do not establish human-error detection, robustness to unseen corruptions, or generalization beyond the evaluated banks.

[CV-181] Spatiotemporal Reconstruction of Auroral Precipitations with Neural Fields WACV2027

链接: https://arxiv.org/abs/2611.19757
作者: Nazar Misyats,Mizuki Fukizawa,Yoshimasa Tanaka,Satoshi Ikehata
类目: Computer Vision and Pattern Recognition (cs.CV); Space Physics (physics.space-ph)
备注: Accepted to the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2027). Camera-ready version with supplementary material. 17 pages, 15 figures

点击查看摘要

Abstract:Auroras provide remote optical observations of energetic electrons precipitating into the upper atmosphere, and whose energy distribution encodes key information about large scale magnetospheric processes. Recovering the underlying electron precipitation field from ground-based camera images of auroras remains a challenging inverse problem. Existing auroral tomography techniques are often limited to simplified static settings, discrete spatial and energy grids, or strong assumptions on the electron energy spectrum. We propose to model the dynamic auroral electron precipitation field via a temporally conditioned neural network. The state of the precipitation field at each frame is represented by a per-frame latent code driven by learned discrete dynamics. A shared decoder maps horizontal positions and latent states to electron energy spectra. The predicted flux is converted to multi-wavelength volume emission rates and rendered into camera images accounting for delayed emissions. We evaluate the method on synthetic dynamic auroral sequences and real multi-view auroral videos, showing that the proposed representation produces temporally coherent precipitation reconstructions and image projections consistent with the observations.

[CV-182] Drift-Conditioned Structural Replay for Continual Cardiac Ultrasound Segmentation

链接: https://arxiv.org/abs/2611.19756
作者: Zahid Ullah,Jaeyoon Seo,Jihie Kim
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Conventional experience replay typically applies fixed preservation constraints to historical exemplars, overlooking variations in how learned representations, predictions, and anatomical structures change during sequential training. To address this limitation, we propose Drift-Conditioned Structural Replay Loss (DCSRL), an exemplar-adaptive supervision mechanism that regulates anatomical preservation according to source-referenced knowledge drift. DCSRL operates within a reliability-guided replay framework that selects informative historical exemplars using anatomical content and boundary characteristics while adaptively regulating replay during training. For each retained exemplar, DCSRL constructs a frozen source-task reference signature and subsequently measures changes in feature representations, class probabilities, boundary responses, and predicted anatomical structures. These complementary drift measurements are aggregated and normalized to generate bounded, exemplar-specific coefficients that dynamically adjust boundary and structural replay losses without modifying the underlying memory-selection policy. We evaluate DCSRL under sequential learning across CAMUS, CardiacNet, and EchoNet-Dynamic, comparing its performance with regularization-, distillation-, and replay-based continual-learning approaches. Experimental findings indicate that DCSRL improves overall final segmentation performance and anatomical boundary preservation, particularly under heterogeneous imaging conditions, although trade-offs remain in knowledge transfer across tasks. These findings suggest that source-referenced multidimensional drift conditioning offers a promising direction for adaptive anatomical knowledge preservation in continual cardiac ultrasound segmentation.

[CV-183] Input-Space versus Feature-Level Conditioning for Robust Facial Expression Recognition

链接: https://arxiv.org/abs/2611.19754
作者: Weihang Yuan
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 2 figures, 4 tables

点击查看摘要

Abstract:Facial expression recognition (FER) degrades sharply when illumination, exposure or resolution departs from the training distribution. Two remedies dominate practice: enriching the input signal, or conditioning intermediate features on an estimated image quality. The two are rarely compared under matched backbones, training budgets and condition annotations, and their cost is seldom reported. We present a controlled comparison on RAF-DB. Input-space conditioning uses four deterministic, label-free views (logarithmic stretching, inverse-logarithmic stretching, down-up resampling and identity), each processed by a shared backbone with a branch-specific BatchNorm and linear head, with the branch logits averaged. Feature-space conditioning is implemented as FiLM and Conditional BatchNorm under an identical protocol. Across six backbone architectures the input-space variant improves top-1 accuracy on every one (mean +0.94 percentage points; bootstrap interval [0.51, 1.63]; sign test p = 0.016), with a seed-controlled estimate of +0.42 percentage points on ResNet-18, whereas feature-level conditioning yields +0.10 and +0.03 percentage points. We attribute the gap to the information content of the heuristic condition label rather than to the conditioning mechanism. The gain costs 4.00x FLOPs, 4.18x latency and 3.79x training time, i.e. 0.26 percentage points per additional GFLOP, and we state an explicit rule for when the trade is worth making.

[CV-184] SEAT: Sign-Type Evidence-Audited Transport for Multi-Label Tongue Diagnosis

链接: https://arxiv.org/abs/2611.19748
作者: Liuyi Xu,Qing Wang,Zikai Liang,An-Yang Lu,Yining Qian
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 3 figures, 4 tables

点击查看摘要

Abstract:Automated tongue-sign recognition can improve the consistency of non-invasive examinations that otherwise rely on subjective visual judgment. A single tongue image often exhibits multiple coexisting signs characterized by heterogeneous evidence, including localized morphology, global chromaticity, and inter-label dependence. However, existing methods typically rely on shared representations and uniform multi-label heads, making it difficult to verify whether predictions stem from clinically relevant evidence rather than spurious visual correlations. We propose SEAT, an evidence-audited framework that matches distinct evidence types with dedicated prediction and evaluation mechanisms. Built on frozen DINOv2 patch features, SEAT couples mask-guided sign-specific attention for localized morphology, a semi-relaxed optimal transport pathway over LAB/HSV color prototypes to enhance pale-tongue recognition, and a co-occurrence-aware transport regularizer to shape the batch-level predicted label marginal. These components are evaluated through audits matched to their evidence types. On the TongueDx benchmark, SEAT achieves a mean macro-F1 of 72.59 with a standard deviation of 1.04 across six random seeds, with a six-model ensemble reaching 74.15. The ensemble outperforms a backbone-matched baseline and a larger, domain-specialized vision-language model by 8.06 and 2.99 percentage points, respectively. Our results demonstrate that aligning dedicated prediction pathways and audits with sign-specific characteristics improves multi-label accuracy while offering more targeted evidential grounding than uniform post-hoc spatial explanations. Code is available at this https URL.

[CV-185] How Far Is Too Far? Object Recognition Declines Monotonically with Semantic Distance NEURIPS2026

链接: https://arxiv.org/abs/2611.19747
作者: Domenic Bersch,Hai Van Tran,Gemma Roig
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: Accepted at NeurIPS 2026 (Spotlight)

点击查看摘要

Abstract:Vision models are sensitive to object-background context, yet it remains unclear how recognition changes as the semantic distance between objects and backgrounds increases. We introduce ImageNet-OOC1k, a benchmark enabling graded control of object-background relationships along a continuous consensus visual-semantic dissimilarity axis. This formulation reveals a highly consistent monotonic decline in recognition accuracy across 35 models spanning convolutional neural networks, Vision Transformers, and vision-language models. Internal analyses of Vision Transformers reveal patterns consistent with aggregation failure, where object evidence is present at the patch level but not consolidated into the final prediction. Using this characterization as a diagnostic probe, a simple spatial re-ranking strategy recovers a subset of these errors and improves in-sample accuracy by up to 5.7 percentage points. These results establish context sensitivity as a continuous and predictable property of modern vision systems, with Vision Transformer analyses suggesting that aggregation limitations may contribute to some failure types even when object evidence remains available.

[CV-186] Event-based object detection under adverse lighting and motion: lessons for construction sites

链接: https://arxiv.org/abs/2611.19746
作者: Zhuoran Feng,Xiaoshan Zhou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 15 pages, 12 figures, 9 tables

点击查看摘要

Abstract:Vision-based monitoring on construction sites fails at night, at dawn and dusk, in tunnels and under floodlights, when conventional RGB cameras under- or over-expose. Event cameras record per-pixel brightness changes with microsecond latency and over 120 dB of dynamic range, but their benefit for site-relevant object detection has not been quantified on real data. We compare three recent event-based detectors (SMamba, SATE and EvRT-DETR) with an RGB detector trained on the same data, using the pixel-aligned PEOD dataset. On low-light frames (70% of the test set) the best event detector reaches 27.2 AP against 21.3 AP for RGB, and a training-free late fusion of both sensors is the best option in every lighting condition (33.3 AP). An object-level analysis of 53,091 objects shows that brightness has little effect on event detectors, whereas fast motion matters: above 20 pixels per frame RGB finds 11-19% of medium-size objects and EvRT-DETR 39-60%. Activation patching locates the damage caused by losing events at the 1/16-resolution features, with small objects depending on the finest scale, and shows that event detectors fail when few signal events meet added noise. A model-agnostic pruning study keeps about 95% of accuracy with 60% of the input but halves small-object recall. On rendered construction sites, used zero-shot, dimming the floodlights drops RGB to near zero while the event detector keeps most of its AP50. We close with camera-placement guidance for construction sites.

[CV-187] IVS: Mapping the Sensitivity Landscape of Visual Token Transformations in MLLM s

链接: https://arxiv.org/abs/2611.19745
作者: Yikang Qiao,Jianye Zeng,Duan Huang,Ziying Song
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 9 pages, 12 figures, 3 tables

点击查看摘要

Abstract:Recent efforts to improve the efficiency of multimodal large language models (MLLMs) have explored visual token pruning, quantization, and feature compression. However, the properties that determine whether transformed visual tokens remain usable by a frozen decoder are not well characterized. We introduce \textbfTIVS (Token Identity Versus Statistics), a training-free intervention framework that separately studies sequence permutation, token retention, numerical precision, and statistical replacement under controlled conditions. On 192 POPE questions, three token permutations change TinyLLaVA accuracy by only -0.17 percentage points on average, while changing 4.69% of individual predictions. Diagonal and rank-64 Gaussian replacements decrease accuracy by 25.0 and 28.1 points on the same subset. Further experiments with exact moment constraints, quantization error, token retention, and additional visual tasks reveal how strongly the effects depend on the original feature values and the evaluation setting. These findings provide an empirical basis for assessing visual token transformations in frozen MLLMs and designing more reliable multimodal interfaces.

[CV-188] From Cloud to Edge: A Fully Local Pipeline for Pothole Risk Assessment and Driver Warning Generation

链接: https://arxiv.org/abs/2611.19737
作者: Abdalmalek Aburaddaha,Samir A. Rawashdeh
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 23 pages, 8 figures, 7 tables. And submitted to MDPI Sensors

点击查看摘要

Abstract:Vision-language models can turn road images into clear, geometry-aware pothole warnings, but current systems depend on commercial cloud services. Our earlier five-stage pipeline needed two cloud calls per scene, taking about 20 s and USD 0.105, and could not run offline. We rebuild every stage to run on a single local GPU and evaluate it with the same data, prompts, and protocol. Segmentation keeps the original Segment Anything weights but runs them in half precision with graph compilation, which makes it 2.7x faster without changing any downstream risk class; five lightweight variants could not match its masks on small road defects. Severity is graded by six ConvNeXt students distilled from 3092 cloud labels. They reach Cohen’s kappa = 0.268 against human consensus at 28 ms per pothole, close to the cloud model’s 0.321, within the camera setup they were trained for. A local 8-billion-parameter vision-language model writes the warnings under a grammar that admits only in-lane maneuvers. Three blind, independent raters scored these warnings safer than the cloud warnings (4.18 vs. 3.65 of 5; p = 0.001; 50 scenes) and equally clear, whereas an automated LLM judge ranked them the other way. The complete pipeline runs at 1.27 s per scene, 16x faster than the cloud system, offline and at no per-scene cost.

[CV-189] Cache or Denoise Less? Residual Reuse in a Flow-Matching World Action Model

链接: https://arxiv.org/abs/2611.19709
作者: Abhinav Gupta,Nissan Pow
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: Honorable Mention at the CoRL 2026 Workshop on Efficient Foundation Models for Real-Time Embodied AI (EEAI). Code and data available at this https URL

点击查看摘要

Abstract:Flow-matching robot policies generate each action chunk through repeated network evaluations, one per denoising step. Cross-cycle caches such as C ^3 ache reuse computation from earlier replanning calls, but whether they beat simply taking fewer steps is unclear. We compare the two in Fast-WAM, a 6B-parameter world action model, on LIBERO-Goal. Two uncached steps keep the success of ten: on all 500 initial states they are non-inferior within two percentage points, for one checkpoint and noise seed. Under PyTorch’s default compile mode, they also cut a replanning call on an RTX A6000 from 202 to 92 ms. This leaves little room for caching. A cache that keeps two fresh steps cannot be faster than two uncached steps: every call still encodes the images, runs the video expert and computes those two steps. Caching the first of two steps, refreshed every fourth call, saves at most 11–14% of a call by our cost model. Cached policies are also fragile. Their losses tracked the share of the flow path the cached steps cover: under Fast-WAM’s current time shift, C ^3 ache’s eight cached steps cover 80% of the path instead of 44%, and the cached policy succeeds on 63% of states against 95% uncached. Resampling the noise on every call cost caches covering 60% or more of the path 14–40 points. These results argue for fewer-step baselines and for testing reuse under the deployed sampler and noise settings. We release our evaluation harness, logs and a checklist.

[CV-190] Beyond the Training Cameras: Coverag e-Controlled Probes of Viewing Direction in Frozen Vision Models

链接: https://arxiv.org/abs/2611.19642
作者: Peize Ding
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Vision models are trained on 2D images of a 3D world. Do their hidden features encode the direction from which the camera views a scene, and if so, in what form? We study the relative viewing direction \alpha of two objects in rendered scenes where both object centres stay at fixed pixels while the camera moves, so that perspective, shading and shadows change but the projected layout does not. In eight pretrained models, from diffusion models to self-supervised and image-text encoders, \alpha is encoded in a strikingly simple form: a single two-dimensional plane in the hidden space, shared across scenes, in which the difference between the two objects’ features traces (\cos\alpha, \sin\alpha) . The plane keeps its orientation when lighting, colours, floors and object shapes change. This encoding is more than a summary of a few visible measurements: readouts fitted on one set of camera paths stay accurate at camera positions outside their range in seven of the eight models, whereas decoders of the measured object and shadow geometry, equally accurate within that range, fail there. In the diffusion transformer PixArt- \alpha , the plane forms abruptly at a single block and carries direction information to later blocks, yet editing it changes the objects rather than the generated viewpoint. We also examine how this encoding carries over to scanned real objects.

[CV-191] Floorplan-Context-Aware Cross-Environment Human Activity Recognition from mmWave Radar Point Clouds ICASSP2027

链接: https://arxiv.org/abs/2611.19640
作者: Zixiong Han,Xilai Wang,Chenzhe Zhao,Amr Mourad,Miodrag Bolic
类目: Computer Vision and Pattern Recognition (cs.CV); Signal Processing (eess.SP)
备注: Submitted to ICASSP 2027

点击查看摘要

Abstract:Human activity recognition (HAR) models based on mmWave radar point clouds and trained in controlled open spaces often experience performance degradation when deployed in furnished rooms. We propose a Floorplan-Context-Aware (FCA) HAR enhancement method that augments the input with subject-furniture relative position information to improve cross-environment performance. Because existing mmWave point-cloud datasets lack registered floorplans, we introduce a pseudo-landmark augmentation strategy for training the FCA model. In cross-environment evaluation on the mm-ADL-Home dataset, FCA achieves 77.70% accuracy and 65.49% macro-F1, improving upon the point-cloud-only direct-transfer baseline by 10.73 and 10.05 percentage points, respectively, and exceeding the strongest evaluated adaptation baseline by 7.37 and 6.27 percentage points. At deployment, FCA requires only a one-time registration of furniture landmarks, with no target-domain activity annotations or model weight updates.

[CV-192] Where slide-level foundation-model classifiers fail on atypical breast lesions

链接: https://arxiv.org/abs/2611.19630
作者: Nerea Hernandez,Luis J. Herrera,Cesar Ramirez,Olga Valenzuela,Ignacio Rojas,Francisco M. Ortuño
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Whether a ductal breast lesion is called atypical hyperplasia or low-grade carcinoma in situ increasingly determines surgery. However, slide-level classifiers built on pathology foundation models (PFMs) perform worst on these lesions. Using four PFMs, cross-validation on BRACS and preregistered hypotheses, we studied where the classification of atypia fails. Because slides are labelled with their most severe lesion, only 39% of the slides with annotated atypical ductal hyperplasia (ADH) were labelled ADH. Attention ranked the lesion high but rarely first, and showing the model only the lesion helped little, whereas classifiers trained per decision with lesion-level labels improved all four atypical boundaries. ADH regions inserted into benign slides raised the probability of malignancy more than that of atypia. In a hospital with pathologist-labelled single-subtype slides and in HISTAI, benign slides were scored closer to lesions and triage specificity fell from 0.48-0.53 to 0.00-0.07 and 0.22.

[CV-193] oward Planning Representations That Separate Task Structure from Visual Content

链接: https://arxiv.org/abs/2611.19621
作者: Aarav Srivastava,Vishwas Sathish,Rajesh P. N. Rao
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Effective planning requires representations that preserve task-relevant geometry and object identity while ignoring visual variation that does not alter the task. Existing evaluations of visual generalization often change several factors simultaneously, making the source of a failure difficult to identify. We introduce a controlled evaluation of object-goal navigation (ObjectNav), the task of reaching an instance of a named object category, in procedurally generated ProcTHOR houses. Training and evaluation houses share floor plans, object locations, goals, and start states, while visual factors are changed individually or in combination. Across model-free reinforcement learning (RL), model-based RL, and agents using frozen pretrained visual representations, our results show that appearance changes impair navigation, but their effects differ sharply by factor. Surface material changes account for most of the average change in success, whereas clutter and lighting have little effect. The magnitude of a global image statistics shift does not reliably predict performance: changing only the goal object’s appearance can substantially impair an end-to-end agent despite a small image-level change. We found that a model based agent that plans over frozen features from the pretrained DINOv2 image encoder achieves the highest success (0.83) under the full shift, but remains sensitive to surface appearance. Our experiments reveal an important requirement for planning representations: they must separate invariant spatial task structure from appearance.

[CV-194] Learning Spatially Aware Audio-Visual Features from Videos with Mono Audio WACV2027

链接: https://arxiv.org/abs/2611.19593
作者: Mengyu Yang,You Zhang,Vishnu Raj,Olha Townsend,James Hays,Andrea Fanelli,Gautam Bhattacharya
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: WACV 2027

点击查看摘要

Abstract:Comprehensive audio-visual perception requires both semantic and spatial reasoning. Existing spatial models suffer from poor generalization due to data scarcity, while semantic-only models fail to capture spatial characteristics completely. We introduce a model capable of jointly reasoning over binaural audio and vision on semantic and spatial tasks with improved generalization. To overcome the shortage in real-world data, we develop an automatic pipeline that localizes the sound source in a video, estimates its relative 3D trajectory, and renders binaural audio. We then use this to spatialize VGGSound, allowing us to train at a much larger scale. We demonstrate improved performance and generalization on spatial reasoning tasks while still remaining competitive on semantic tasks compared to semantic-only models.

[CV-195] EndoComparer: Comparative Surgical Video Question Answering with Pseudo-Siamese Visual Adaptation and Perceive-then-Compare Reasoning

链接: https://arxiv.org/abs/2611.19578
作者: Shi Li,Lalith Sharan,Joël L. Lavanchy,Kun Yuan,Lorenzo Arboit,Nabani Banik,Elif Yazici,Maria Vannucci,Vinkle Srivastav,Pietro Mascagni,Nicolas Padoy
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 33 pages, 14 figures, 8 tables

点击查看摘要

Abstract:Purpose: Surgical training benefits from explicit comparison between cases: in patient condition, in technique and in skill proficiency. Surgical vision-language models (VLMs) offer a promising and flexible interface for this purpose. Yet current VLMs typically see one video at a time, and they often fall short when asked to compare videos. Methods: We propose EndoComparer, a VLM for comparative surgical video question answering with two components. (a) A Pseudo-Siamese Visual LoRA (PSVL) gives each video slot its own low-rank adapter over a shared frozen vision encoder, separating the two videos in the visual pathway at the cost of a single encoder pass. (b) A Perceive-then-Compare chain of thought (PtC-CoT) trains the model to first state what each video shows, then derive the answer from these statements. We further curate EndoDiffVQA, a benchmark of nearly 11K multiple-choice questions on video-pair over cholecystectomy and gastric bypass procedures, covering anatomy, tool, action and skill. Results: EndoComparer outperforms strong generalist and surgical VLMs on detecting and specifying difference between videos. Ablations show that two proposed components contribute complimentarily, and the model remains more robust under a video-swap control that exposes confusion between the two videos. Conclusion: Providing a dedicated visual reasoning path through PSVL and PtC-CoT makes cross-case comparison of surgical videos more robust for VLMs. EndoDiffVQA and EndoComparer provide a benchmark and a strong baseline for this task. Project page: this https URL

[CV-196] Consistency-Aware Map Generation at Multiple Zoom Levels Using Aerial Image

链接: https://arxiv.org/abs/2611.19576
作者: Linwei Chen,Zheng Fang,Ying Fu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 14 pages, 9 figures

点击查看摘要

Abstract:The multi-level tiled map service is widely used and serves as a kind of digital infrastructure. These map tiles are usually rendered from vector data, whose update needs to walk or drive with professional equipment to check every points of interest. This leads to inconvenience and expensive cost in timely updating maps. Compared with vector data, aerial images are much easier and cheaper to obtain. In this paper, we propose a novel multi-level map generation framework that can automatically generate accurate and consistent maps with multiple zoom levels from aerial images. It consists of a level-aware map generator and a consistency-aware map generator. The level-aware map generator is able to generate accurate initial maps with realistic details for each zoom level. The consistency-aware map generator regards initial maps at each zoom level as a sequence and builds the connection between them, so as to guarantee content consistency between maps at different zoom levels. Furthermore, we collect a large-scale high-quality dataset called Multi-Level Map (MLM) for map generation at multiple zoom levels. Experiments on our MLM dataset show that our method outperforms previous state-of-the-art map generation methods on both comprehensive quantitative metrics and perceptual quality.

[CV-197] Spectral Steering: Local Inverse Geometry for Frozen Diffusion Generators

链接: https://arxiv.org/abs/2611.19572
作者: Angelo Moroncelli,Alessandro Leanza,Loris Roveda
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注: Under review

点击查看摘要

Abstract:A frozen diffusion generator may need to change several output properties while preserving others. We study when guidance faithfully realizes such a structured request and whether local inverse correction improves its direction. Our account links this benefit to the interaction between the request and the objective-specific remaining-sampler Jacobian: conditioning alone is insufficient. Spectral steering combines a request-specific cosine-gain predictor with damped inverse correction and fixed-time relinearization for finite edits. Scalar nulls and matched-movement controls test the predicted mechanism. Across 8,923 instances in a retrospectively identified local regime, predicted and realized gains correlate at Spearman 0.79, versus -0.11 for condition number. Experiments span images, audio, robotics, and control and planning, including complex character motion generation. On 50 paired inpainting cases, four corrections improve unobserved-pixel PSNR over Adam by 3.25 dB at equal total derivative traversals. Together, the analysis and controls connect request-dependent geometry to directional gains and establish where these gains translate into downstream performance.

[CV-198] Sharp First-Hit Identification from Finite Depth Samples

链接: https://arxiv.org/abs/2611.19563
作者: Hamidreza Hasani Balyani,Mahtab Movahhedrad,Seyed Pouyan Mousavi Davoudi
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注: 36 pages, 9 figures. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:Which depths in a new camera view are determined by finitely many source-depth measurements? Pointwise uncertainty bounds do not answer this question: unobserved geometry between samples can change which surface is encountered first. We study a bounded Lipschitz class of continuous source inverse-depth fields with interval observations and known camera pose. Building on classical Lipschitz envelopes, we characterize the exact set of attainable first-hit depths and construct a globally admissible scene for every value in that set. The set can be disconnected. We also give a necessary and sufficient test for several proposed depths to be first hits of one shared scene. An epipolar reduction uses scalar cone inequalities without enumerating surfaces. In a rectified model with source spacing h , interval half-width \varepsilon , baseline b , and slope bound L , we prove a sharp worst-case inverse-depth diameter of \max\Lh+2\varepsilon,2\varepsilon/(1-bL)\ when bL1 , subject to sufficient depth and domain range for attainment. Exact samples still give the sharp bound Lh at bL=1 , while above this threshold ambiguity need not vanish as sampling becomes arbitrarily dense. Any positive measurement uncertainty can destroy uniform stability at the critical threshold. We provide constructive examples, an exact rational validation oracle, and a frozen ten-scene evaluation using reference, stereo, and learned-depth observations. The experiments separate solver correctness under the scene assumptions from failure of those assumptions: fixed learned-depth bands often yield narrow but incorrect sets. The results identify what a specified scene class permits; they do not infer a continuum regularity or physical-scene guarantee from sampled predictions.

[CV-199] Read Then Shade? How Visual Evidence Changes Incentive-Driven Misreporting in Vision-Language Advisors

链接: https://arxiv.org/abs/2611.19561
作者: Hamidreza Hasani Balyani
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: 45 pages, 5 figures. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:Vision-language models (VLMs) increasingly read a scene and report it to someone who acts on the report, sometimes under objectives that differ from that person’s. We ask whether a stated incentive moves what a VLM reports, and whether this changes when the value must be read from pixels rather than given as text. In seeded synthetic scenes, a model reads the position of a target on a labelled axis and reports one integer to a fixed receiver; it is paid, in simulation, for accuracy or for a report a few units to one side, and every scene is also presented with the true value as text. Across a pilot and three follow-up studies with hash-frozen protocols, four small open VLMs from three families, and 15,170 held-out trials, a stated payoff moved reports in both modalities, and exact compliance at three units was significantly lower with the image than with text for three of four models; Qwen2.5-VL 7B showed no reliable difference. Adding the true value as text to the image prompt removed the gap in every model. For InternVL3 the gap vanished on scenes it read correctly when it wrote its reading first; for Gemma 3 it persisted. For Qwen2.5-VL 3B, which read every scene correctly, the gap held in four of five payoff paraphrases, nearly vanished when the model wrote its reading before its report or when the prompt stated that the client wants the truth, vanished with an unpaid goal, and persisted when the model reported before stating its reading; how often it shaded at all, even with text, depended strongly on the wording. When told that the client wants the true value, three of four models still shaded every text report. Visual input did not make these advisors reliably more truthful; it made compliance with a stated incentive depend on whether the read value is made explicit and on output order, a dependence that text-only evaluations and mean-shift metrics do not reveal.

[CV-200] Lesion-Aware Explainable AI for Diabetic Retinopathy Grading from Fundus Images

链接: https://arxiv.org/abs/2611.19550
作者: Truong Thinh Nguyen,Hai Phong Nguyen,Truong Thanh Hung Nguyen,Giang Son Tran,Hung Cao
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 12 pages, 6 figures, 9 tables. Truong Thinh Nguyen and Hai Phong Nguyen contributed equally. Corresponding author: Truong Thinh Nguyen

点击查看摘要

Abstract:Post-hoc explanations of deep-learning models for diabetic retinopathy (DR) grading are difficult to trust because different attribution methods highlight different retinal regions for the same image, and a visually plausible heatmap does not show that the highlighted pixels drive the prediction. We propose the Consensus-Calibrated Explanation Map (CCEM), which fuses Grad-CAM++, Ada-SISE, and SmoothGrad Integrated Gradients (SG-IG) using image-specific weights derived from annotation-free reliability cues: deletion sensitivity, inter-method agreement, compactness, retinal containment, and peak structure. CCEM explains a frozen ordinal grader, an EfficientNet-B7 with CBAM and a scalar regression head, pre-trained on EyePACS 2015 in Stage 1 and fine-tuned on APTOS 2019 in Stage 2 with quadratic weighted kappa (QWK)-optimized thresholds; it reaches a QWK of 0.920 on the hidden Kaggle test set and 0.559 on the external MAPLES-DR cohort. Explanations are evaluated on MAPLES-DR, 198 MESSIDOR images with expert pixel-level lesion annotations that are never used for training, threshold fitting, or fusion. On the 193 images with annotated lesions, CCEM obtains the highest Soft Pointing Game score (69.61%; 95% CI 63.7-75.3), defined as the ratio of the strongest lesion-supported saliency to the global map maximum rather than a lesion detection rate, together with the lowest deletion AUC (0.570) and the best deletion-insertion balance (0.115); each gain is significant against the strongest single explainer on paired tests. Annotation-free, per-image reliability weighting therefore turns three disagreeing explainers into one map whose peak evidence is more lesion-supported and more faithful to the model than that of any single explainer.

[CV-201] PathLoop: Self-Evolving Multi-Agent Reasoning for Evidence-Grounded Whole-Slide Image Analysis

链接: https://arxiv.org/abs/2611.19527
作者: Jincheng Li,Zhengyi Liu,Zelin Liu,Yihui Zhan,Yifei Sun,Xiangyu Li,Minye Shao,Lingyao Kong,Tao Hou,Xinheng Lyu,Lanfeng Zhong,Ziquan Zhu,Rongzhao Zhang,Rui Yan,S. Kevin Zhou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Analyzing whole-slide images (WSIs) requires combining evidence across tissue regions and spatial scales. Multi-agent systems offer a promising approach to this task. However, existing systems make limited use of molecular evidence, often select visually prominent but less representative regions while overlooking global context, and struggle to reuse experience from previous cases. To address these challenges, we present PathLoop, a self-evolving multi-agent framework that integrates omics data for evidence-grounded WSI analysis. Specifically, PathLoop coordinates visual and omics interpreters across slide, patch, and pixel levels, using molecular signals to guide region selection and combining local tissue details with global slide features. A judgment agent then checks the consistency of the collected evidence against Pathology-Wiki, a knowledge base built from 48 entries across 24 organ systems, and triggers re-examination when the evidence is inconsistent. To reuse past experience, a four-level memory stores reasoning trajectories, practical tips, skills, and tools to guide future cases, forming a loop of evidence collection, review, and memory update. We evaluate PathLoop on four TCGA cohorts with biomarker prediction, survival prediction, and subtype classification. The results demonstrate the effectiveness of PathLoop across these tasks and support its use for evidence-grounded WSI analysis.

[CV-202] DBHiCo: Dual-Branch Hierarchical Contrastive Learning for MRI Image Classification

链接: https://arxiv.org/abs/2611.19511
作者: Chengsheng Zhanga,Wei Li,Chenghao Sun,Jian Gong,Xinmei Tian
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 14 pages

点击查看摘要

Abstract:Automated Magnetic Resonance Imaging (MRI) image classification provides a crucial non-invasive approach for neurological diagnosis, yet it is fundamentally challenged by inherent acquisition noise and high inter-subject structural similarity, which obscure subtle pathological variations, thereby limiting the performance of deep neural networks. While contrastive learning (CL) offers a promising mitigation strategy, existing paradigms rely on rigid binary partitions that ignore negative sample heterogeneity, and they fail to capture the cross-view structural consistency vital for fine-grained anatomical alignment. To address these barriers, we propose Dual-Branch Hierarchical Contrastive Learning (DBHiCo), a novel framework explicitly designed to collaboratively enforce structural consistency and semantic discrimination. Specifically, DBHiCo leverages a natural anatomical multi-view pair (original and skull-stripped MRI scans), processing them through a dual-branch architecture to seamlessly integrate holistic structural robustness with localized intracranial details. Crucially, we introduce a hierarchical contrastive loss that discards conventional binary constraints in favor of a margin-ranked similarity hierarchy. This mechanism explicitly orders positive and negative pairs based on their structural and semantic difficulty, thereby amplifying the fine-grained semantic discriminability of the learned representations. Extensive experiments across three diverse neurological datasets demonstrate that DBHiCo achieves state-of-the-art performance, consistently outperforming existing MRI multi-view representation learning and advanced contrastive fine-tuning methods.

[CV-203] Vegetation Health Classification for Low-Data and Edge Hardware Deployment

链接: https://arxiv.org/abs/2611.19495
作者: Yifei Cheng,Haoyu Li
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Unmanned aerial vehicles (UAVs) equipped with cameras are popular, viable tools for vegetation and environmental health monitoring. However, existing image classification methods are incompatible with UAV implementation for independent practitioners in remote, resource-constrained environments due to large dataset requirements, GPU hardware, inflexible retraining, and low interpretability. This paper presents a lightweight, interpretable vegetation classification algorithm operating with as few as five reference images per class, without GPUs or heavy machine learning frameworks. It uses a 2D hue-saturation histogram feature representation and Earth Mover’s Distance or Quadratic Form Distance, paired with a log-likelihood probability score, to classify images into three vegetation health categories. We evaluated six systems: three custom pipelines and three established models across low-data scenarios ( n = 5, 10, 25 ) and a full-data regime of 2,172 drone images from 28 South American locations. Our custom pipelines achieve competitive Top-1 accuracy (81–85%) under low-data constraints, maintaining sub-1ms inference times that enable real-time classification and subsequently tasks such as seeding and watering. The algorithms also support incremental retraining without full model reconstruction. We validate edge feasibility by deploying the QFD variant on an ESP32 microcontroller, achieving 82.8% accuracy with 1.01% heap memory utilization (45kb), demonstrating that effective vegetation classification is achievable on low-cost, low-power embedded hardware without cellular connectivity.

[CV-204] Beyond Paired Inputs: Learning to Leverag e RGB-T Cross-Modal Relations in Multimodal Large Language Models

链接: https://arxiv.org/abs/2611.19493
作者: Tianyi Zhao,Qihao Xu,Jiawen Xi,Duanni Meng,Bo Zhao,Maoxun Yuan,Xingxing Wei,Bo Li
类目: Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)
备注: 9 pages, 4 figures

点击查看摘要

Abstract:RGB and thermal infrared (TIR) sensors provide complementary observations, yet their distinct sensing characteristics make effective cross-modal coordination challenging. Although existing multimodal large language models (MLLMs) can process paired RGB-T inputs, they often fail to translate cross-modal relations into effective task-specific decisions, leaving the potential of these relations to improve task performance and robustness largely untapped. Motivated by this observation, we introduce RelMLLM, an RGB-T MLLM that explicitly learns to exploit cross-modal relations in task-specific decision making. It represents modality reliability, the cross-modal relation, and the dominant modality as a verifiable decision state. We derive task-grounded supervision for this state by comparing RGB-only, TIR-only, and paired RGB-T task outcomes. A hierarchical relation-guided reward supervises modality-specific evidence, relation-state prediction, and the final task output, while encouraging consistency across evidence, relation, and decision. We evaluate RelMLLM on natural RGB-T grounding, RGB-T visual question answering, and held-out sensor degradation. RelMLLM improves grounding accuracy and degradation robustness over its Qwen3-VL baselines while preserving performance on general RGB-T reasoning benchmarks. These results demonstrate that effectively exploiting cross-modal relations leads to more accurate and robust RGB-T decisions. Code and model weights are publicly available.

[CV-205] PALIN: Pixel-Aligned LiDAR-NIR Multimodal 3D Object Detection with Benchmark

链接: https://arxiv.org/abs/2611.19485
作者: Yanan Zhang,Amit Vishnia,Chengbo Ai,Eleni Christofa,Runan Cheng,Siyuan Meng,Yuanchang Xie,Yu-Min(Thomas)Yang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 17 pages, including references and appendices

点击查看摘要

Abstract:3D object detection is critical for intelligent transportation systems. LiDAR has been widely adopted for accurate geometric measurements; however, purely geometric representations remain ambiguous under sparse observations and occlusions, while cross-sensor multimodal fusion suffers from calibration and synchronization errors. To overcome these challenges, we leverage an Ouster LiDAR to construct PALIN (Pixel-Aligned LiDAR-NIR), a large-scale intrinsically pixel-aligned multimodal dataset jointly capturing geometry and NIR intensity, comprising approximately 8,000 annotated frames from static urban intersections and diverse mobile suburban scenes. Building upon this dataset, we propose CMambaFusion, a novel fusion framework consisting of a four-directional Cross-Mamba module for bidirectional cross-modal interaction and a Gated Concat-Mamba module for joint feature refinement. Extensive experiments demonstrate that PALIN enables effective multimodal learning, and CMambaFusion achieves superior accuracy-efficiency trade-offs over both LiDAR-only and representative fusion baselines.

[CV-206] StyleJudge: A Human-Validated In-Context-Learning Vision-Language Evaluator for Artistic Style Transfer

链接: https://arxiv.org/abs/2611.19475
作者: Lokesh,Pramukto Mandal,Somanshu Juneja,Jashnoor Singh,Amit Kumar,Pranav Adlinge,Richa Thakur,Kunal Swami,Aditya Nigam
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted in ICVGIP 2026

点击查看摘要

Abstract:Automatic evaluation of artistic style transfer remains an open problem: existing metrics score an image pair in isolation, without seeing an example of the intended style transformation, and correlate poorly with human judgment. We introduce StyleJudge, a compact vision-language evaluator conditioned on a small set of reference source-to-stylized pairs before scoring a query pair, mirroring how a human rater calibrates to a style before judging a candidate. On a human-annotated benchmark of 878 rated pairs across 12 artistic styles (728 from 10 training/validation styles, 150 from 2 unseen styles), our efficient 0.8B-parameter StyleJudge reaches a Pearson correlation of 0.291 (Spearman 0.325) with human ratings on seen styles and 0.381 on unseen ones. On seen styles, it achieves the highest positive correlation among tested metrics: Gram-matrix style loss and CLIP similarity correlate weakly ( r=0.041 , r=0.168 ), while LPIPS (scored higher when closer to the source) correlates negatively ( r=-0.404 ). We achieve this by distilling evaluation knowledge from a 72B-parameter vision-language teacher used only during training, which reaches r=0.327 , confirming that most of its human alignment survives compact distillation. For training, we build a dataset of source-to-stylized pairs across the 12 styles via LoRA-adapted diffusion, densified with graded intermediate examples from a high-resolution image-morphing pipeline so the evaluator recognizes partial and complete stylization. On two held-out validation styles, StyleJudge reaches 28.2% exact-match agreement with its teacher’s structured scores versus 1.1% for the untuned base model, with 67.8% shorter reasoning traces, and scores a pair in 39 s (median) using 2.5 GB of GPU memory on one NVIDIA T4. We release the human-rating benchmark, annotation guideline, SDXL-adapted FreeMorph code, and model weights.

[CV-207] InvarOrd: Self-Supervised Ordinal Knowledge Distillation with Rank-Invariant Alignment for Label-Efficient Cross-Species Hyperspectral Fish Freshness Assessment

链接: https://arxiv.org/abs/2611.19472
作者: Kazi Nabiul Alam,Omenuko Joseph Nwebonyi,Pooneh Bagheri Zadeh,Akbar Sheikh-Akbari
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: Accepted at IEEE TAFE

点击查看摘要

Abstract:Hyperspectral models of fish freshness rarely transfer between species, and labelling a new species is slow and destructive. Freshness is also ordinal, and standard domain alignment can erase that order. We present InvarOrd, which adapts to a new species from 30 labelled cubes (5 per rank) in three stages: self-supervised spectral pretraining on unlabelled source cubes, cross-species alignment by a shrinkage covariance discrepancy computed only within matched ranks, and ordinal distillation into a compact student. The alignment term is provably invariant to the rank means, so it gives no incentive to collapse the ordinal axis. On paired salmon and cod data (800 cubes each, 16 days), a source-trained teacher reaches 1.6 to 1.9 days mean absolute error (MAE) on the other species, with quadratic weighted kappa (QWK) below zero. The 423K-parameter student, 2.54 times smaller than its teacher, attains the lowest MAE (1.476 days) and highest within 1-day accuracy (56.1%) among 14 configurations, matching the best of six ImageNet-pretrained backbones (ConvNeXt-T, 1.503 days) at 66 times fewer parameters, 1.44ms per cube on a GPU and 5s re-adaptation. InvarOrd beats target-only training by 0.29 days (p = 0.031). Among alignment objectives, rank-invariant alignment retains the most ordinal structure (embedding rank correlation 0.183, QWK 0.112), whereas global discrepancy lowers the rank correlation to 0.049 and QWK to zero at only 0.014 days higher MAE.

[CV-208] Semantics First Pixels Later: Mission-Driven Event Streams for Drone Video over Bandwidth-Starved Links

链接: https://arxiv.org/abs/2611.19463
作者: Cong Xu
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注: 20 pages, 3 figures; includes the supplementary material as an appendix; a demonstration video is attached as an ancillary file

点击查看摘要

Abstract:A surveillance drone on a bandwidth-starved link cannot send the video its operator would need, yet the operator’s questions are often not about pixels: did three people converge, when, and where? We study what should cross a link of 1 Mbps down to 1 kbps instead, and in what order. BPS, a behavior-primitive stream, sends semantics first and pixels later: it runs a detector and a tracker onboard and sends short records (25 bytes on average) of multi-frame events selected by a mission card of 16-125 bytes, ahead of positions and pictures. We measure it by what a ground-side language model answers from it, under a preregistered protocol with a held-out set, against H.265 read by the vision fine-tune of the same model, H.265 on which the ground runs our own perception, a stream of quantized tracks, and onboard captions. On the development set (84% of questions about convergence, mostly one location) the event stream answers 74.3% of questions at 0.13 kbps. A reader that receives nothing answers 57.4%, and the video reader 59.6% at 100 kbps and 64.7% on the source frames. When the ground instead runs our perception on the video, video matches the stream at 100 kbps, about 800 times the stream’s rate, and falls behind at 64 kbps and below. A track stream matches it at 96 times the bytes and loses 0.03 at 17 times. On held-out data the result is weaker: with events taken from annotated tracks, an oracle, the stream beats H.265 at 100 kbps (+0.13), but with onboard perception it does not (+0.01 [-0.04, +0.06]) and is no better than receiving nothing. Sent alone through emulated links the event records answer 0.739 at 1 Mbps and still 0.724 at 0.3 kbps; an every-frame track stream holds down to 1 kbps and video with ground-side perception down to about 100 kbps. On a Jetson Orin Nano the unchanged pipeline runs at 4.5 fps and misses most events.

[CV-209] Few-Shot Visual Concept Extraction for Steering Diffusion Transformers NEURIPS2026

链接: https://arxiv.org/abs/2611.19440
作者: Nabyl Quignon,Antitza Dantcheva
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to NeurIPS 2026 Main Track

点击查看摘要

Abstract:Image steering aims to provide fine-grained control over concepts in generated images, enabling targeted and continuous manipulation without broadly altering the generation process. Recent steering methods for diffusion models primarily focus on the text-conditioning space or rely on costly auxiliary training. Deviating from that, we here propose a novel steering approach that extracts visual concept directions directly from the image-side activations of Diffusion Transformers. Specifically, we showcase few-shot concept extraction in the activation space using a simple difference-of-means estimator over intermediate residuals. We observe (a) scene-level concepts such as lighting, weather, and style are distributed across image tokens, as well as (b) local concepts such as object colors, textures, materials, and facial attributes are concentrated in specific regions. Towards controlling both concepts, we introduce a new unified training-free framework that uses global difference-of-means for (a) distributed concepts and masked difference-of-means for (b) local concepts, preventing local signals from being diluted by global averaging. Notably, using only a handful of positive and negative prompt pairs, our method identifies and extracts reusable concept directions that can be injected at inference for continuous steering. Quantitatively, our method outperforms text-encoder steering and training-free activation-editing baselines, as well as is competitive with specialized editing models. Our results suggest that as concept directions can be applied at patch level, our method enables precise targeted and compositional steering, manipulating multiple concepts in the same image.

[CV-210] Prepare Lighter and Faster for Open-Vocabulary Queries: A Query-Wise 3D Segmenter for Gaussian Splatting WACV2027

链接: https://arxiv.org/abs/2611.19429
作者: Nhat-Quynh Le-Pham,Khoi Nguyen,Binh-Son Hua
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Graphics (cs.GR)
备注: Accepted at WACV 2027

点击查看摘要

Abstract:Open-vocabulary 3D understanding aims to localize and segment arbitrary objects beyond predefined categories. However, existing approaches based on Gaussian Splatting often rely on expensive preprocessing and lengthy training, limiting their applicability when fast and lightweight solutions are desired. We propose a lightweight, training-free framework for query-driven 3D instance segmentation on Gaussian Splatting. Instead of conventional mask-wise CLIP embeddings, we use dense pixel-level CLIP features to preserve fine-grained semantic information and enable flexible text-driven localization without relying on predefined instance proposals. Given a text query, we identify a high-quality target region in a carefully selected pivot view and propagate it across views using a video segmentor, providing consistent and comprehensive coverage of the target instance. We then lift the resulting 2D tracklets to 3D Gaussians through a novel visibility voting strategy, yielding accurate 3D segmentation masks with clean object boundaries. Extensive experiments demonstrate state-of-the-art performance on LERF-OVS, ScanNet, and ScanNet++, competitive results on 3D-OVS, and significant speedups over existing open-vocabulary 3D segmentation methods. These results highlight the effectiveness of our framework for efficient open-vocabulary 3D understanding without per-scene training.

[CV-211] Memorization Lives in a Few Attention Heads: Mitigating Memorization in Diffusion Models via Projection-Based Suppression

链接: https://arxiv.org/abs/2611.19427
作者: Puwei Lian,Yujun Cai,Songze Li,Yuezu Lv,Guanghui Wen
类目: Computer Vision and Pattern Recognition (cs.CV); Cryptography and Security (cs.CR)
备注:

点击查看摘要

Abstract:Diffusion models can memorize and reproduce training samples during inference, raising serious privacy and copyright concerns. Existing inference-time mitigation methods often suppress conditional information globally, which may reduce memorization but can also harm semantic alignment and image quality. In this paper, we revisit memorization from the perspective of cross-attention heads. Through systematic analysis and targeted ablation, we find that memorization is mainly driven by a small subset of shallow-layer attention heads with abnormally strong semantic control, rather than by the entire cross-attention mechanism. Motivated by this finding, we propose Projection-based Top-k Attention Head Suppression, a targeted inference-time mitigation method. Our method first identifies the top-k attention heads most related to memorization using a semantic-control score, and then removes their condition-specific components through directional projection between conditional and unconditional representations. This design selectively weakens memorization-related features while preserving useful semantic information. Extensive experiments on different diffusion models and datasets show that our method effectively reduces memorization and achieves a better trade-off between memorization mitigation and image quality than existing approaches.

[CV-212] SplatFlow3D: Learning Scene Flow via 3D Motion Feature Splatting

链接: https://arxiv.org/abs/2611.19409
作者: Bo Wang,Dekang Zhu,Jian Li,Zhipeng Meng,Bo Wang,Dewen Hu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Considering the motion continuity and stability of 3D objects over short inter-frame intervals, introducing multi-frame settings in scene flow estimation can potentially address occlusion. In this paper, we propose a two-frame backbone, SimpleFlow3D, which integrates dense information from the image plane and sparse information from the depth dimension in a decoupling-followed-by-coupling manner by hierarchically constructing the 3D motion feature. SimpleFlow3D achieves SOTA results with minimal parameters. We also propose a weight-sharing multi-frame framework to splat the previous 3D motion feature to the current estimation process. The framework unifies the first and non-first frame inference by constructing pseudo inputs. After combining the proposed framework and backbone, the newly born multi-frame method SplatFlow3D further advances the state of the art. It achieves lower errors than the two-frame backbone, with fewer parameters than current methods. On both KITTI and Spring, SplatFlow3D ranked 1st over all published and submitted methods. Experiments show that SplatFlow3D improves occlusion handling, non-rigid motion estimation, and generalization to real-world scenes. Code is available at this https URL.

[CV-213] Detecting Small Space Objects for Space Situational Awareness with Collaborative Edge AI

链接: https://arxiv.org/abs/2611.19394
作者: Xingyu Qu,Peng Hu
类目: Computer Vision and Pattern Recognition (cs.CV); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:With the rapid growth of space activities, near-Earth space is becoming increasingly congested, leading to a higher risk of spacecraft collisions and space debris generation. Recent close-proximity encounters and collision incidents necessitate efficient and cost-effective onboard systems for space object detection (SOD) and collision avoidance. While deep learning (DL) enabled Edge AI using low-cost vision sensors on single or multiple satellites has shown promises for SOD, significant challenges remain, including limited datasets, small SOD, and the lack of comprehensive performance assessments. This paper presents a collaborative Edge AI framework for multiview SOD and systematically benchmarks DL models and fusion strategies. Results show that higher image resolution significantly improves small SOD. For multiple views, image-level fusion reduces the combined communication and inference latency by 30.13% and 32.80% on average compared with feature-level and attention-based fusion across the evaluated inter-satellite link (ISL) bandwidth. For larger view sets, feature-level and attention-based fusion achieve better detection accuracy at the cost of analysis in the sensing-communication pipeline, revealing that multiview collaboration introduces minimal latency overhead under ISL constraints when reception capacity is not exceeded. Our work provides valuable design insights for future collaborative satellite network design, supporting the emerging space situational awareness and space traffic management requirements.

[CV-214] AcousticBEV: Acoustic BEV Object Localization for Intelligent Vehicles

链接: https://arxiv.org/abs/2611.19381
作者: Ziang Liu,Shiming Wang,Joseph Adeola,Julian F.P. Kooij
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注: Ziang Liu and Shiming Wang have equal contributions. Shiming Wang is the project lead

点击查看摘要

Abstract:Modern autonomous driving and driver assistance systems rely predominantly on cameras, radars and LiDARs for environmental perception. However, these modalities face different environmental limitations, including poor illumination for cameras and adverse weather for LiDAR. Acoustic sensing, particularly using microphone arrays, could provide complementary perception capabilities for autonomous driving, but has remained largely underexplored. In this work, we design a horizontally arranged microphone array that functions as the “ears” of a vehicle, enabling localization of nearby moving vehicles in a bird’s-eye view (BEV) representation that captures both their relative direction and distance. Using this array, we collect a real-world driving dataset, AcousticScenes, which contains 2,117 static and 2,863 dynamic samples with synchronized spatial acoustic signals, LiDAR point clouds and 3D bounding box annotations. Building on this dataset, we propose AcousticBEV, a learning-based method for acoustic BEV object localization, which estimates the center locations of moving traffic participants in BEV from the microphone-array signals, conditioned on the ego vehicle’s own speed. Experimental results demonstrate that AcousticBEV outperforms both classical signal-processing baselines and other learning-based methods, achieving F1 scores of 47.7 \pm 1.2 % on “Easy” scenarios and 45.1 \pm 1.4 % on “Difficult” cases. Our findings demonstrate that acoustic signals from a low-cost, vehicle-mounted microphone array can localize nearby moving vehicles in both azimuth and distance.

[CV-215] RETR: A Structure-Preserving RGB-Event Transformer for Robust 3D Lane Detection NIPS2026

链接: https://arxiv.org/abs/2611.19358
作者: Jingtao Dong,Hao Zhuang,Hao Yang,Liyuan Pan,Wei Liang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted by NIPS2026

点击查看摘要

Abstract:Robust 3D lane detection requires accurate metric road geometry recovery from visual inputs, yet conventional RGB-based approaches suffer from fragile lane evidence under adverse illumination, motion blur, and long-range perspective compression. Event cameras offer complementary high-temporal-resolution, high dynamic-range structural cues that are robust to these challenges, but their sparse, motion-dependent responses cannot be directly fused into consistent 3D geometry. In this paper, we present the first attempt to introduce event cameras into 3D lane detection. To enable systematic research, we first build two multimodal benchmarks with metric 3D lane annotations: DSEC-3DLD (real-world sequences) and Ev-OpenLane (large-scale simulated sequences). We further propose RETR, a structure preserving transformer that converts complementary RGB and event observations into coherent 3D lane geometry. RETR first aligns modality consistent evidence through Reciprocal Context Flow Fusion, then preserves thin and uncertain lane structures with Uncertainty Aware Structural Consolidation, and finally decodes ordered lane hypotheses using a Geometry State Decoder with proposal conditioned initialization, reference conditioned query evolution, and diversified geometric inquiry. Extensive experiments show that RETR achieves state-of-the-art performance on both benchmarks, especially showing strong improvements in challenging lighting and complex road geometry scenarios. Code and dataset will be available at this https URL.

[CV-216] HongguoTrending: A Longitudinal Multimodal Archive of Trending Videos and Topics on a Leading Micro-Drama Platform

链接: https://arxiv.org/abs/2611.19351
作者: Rongxin Ouyang
类目: Computers and Society (cs.CY); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Micro-dramas (duanju) are reshaping digital production and consumption, yet a lack of longitudinal data connecting series to their platform visibility persists, limiting our understanding of the emerging medium. We introduce HongguoTrending, a longitudinal, multimodal dataset of hourly snapshots from seven trending lists of Hongguo, China’s largest micro-drama platform. Collected through a two-stage asynchronous pipeline, the archive links 63,200 ranking entries across 632 web list-hours and 157 client hours (889 distinct series) to metadata, cover images, sampled episode videos, and audience engagement metrics. Exploratory analyses show distinct turnover rates between web and mobile trending lists, production-driven mainstream values alongside kinship-centered synopses, genre-specific cover palettes, and strong associations between user sharing and next-day rank gains. Accessible via a RESTful API, this dataset supports empirical research on algorithmic gatekeeping, popularity dynamics, and multimodal digital culture.

[CV-217] Attention U-Net Cellular Automata for Multi-Class Land Use and Land Cover Change Simulation in Lucknow India: Evaluation Against Null and Classical Baselines

链接: https://arxiv.org/abs/2611.19348
作者: Shubham Phapale,Poonam Sharma
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 10 pages, 4 figures, 7 tables

点击查看摘要

Abstract:Convolutional networks are often used to learn the transition rules of land-use cellular automata (CA), yet their predictions are seldom compared with a null model that predicts no change. We built AttentionUNet-CA, which reads a 32 x 32 neighbourhood of land-cover states with four drivers (elevation, distance to major roads, distance to the central business district (CBD) and population density) and predicts the state of each cell ten years later. We tested it on Lucknow, India, with GLC_FCS30D land cover for 1990, 2000, 2010 and 2020, a 30 m series mapped consistently over time, and compared its built-up class with GHSL and ESA WorldCover. AttentionUNet-CA was trained on 1990-2010, selected on held-out spatial blocks and evaluated once on 2010-2020, with three seeds. Baselines were the null model, a random-forest CA (RF-CA) and a U-Net CA without attention; each network was trained with and without class weights. Only 8.1% of the study area changed in the test decade, so no model matched the overall agreement of the null model (91.9%). On change, measured by the figure of merit, the class-weighted networks scored highest (U-Net CA 22.0 +/- 1.2%, AttentionUNet-CA 21.0 +/- 1.4%), ahead of RF-CA (20.3%), but by margins comparable to seed-to-seed variation; the networks estimated the amount of change better than RF-CA. The attention gates did not improve accuracy. Population density and distance to the CBD were the most useful drivers. Projected to 2030, built-up land grows by 22-35%, mostly at the city edge.

[CV-218] A Think Block Is Conditioning Not a Plan

链接: https://arxiv.org/abs/2611.19341
作者: Jian Xu,Delu Zeng,Qibin Zhao
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Generative models increasingly write a chain of thought before drawing, and report large gains from doing so. We ask what that text contributes, by substitution rather than ablation: holding the generator, sampler and seeds fixed, we replace the think block with controlled alternatives and measure the image. Across a unified model with a think mode, a reinforcement-learned reasoning generator, and a planner driving a frozen diffusion model, the same picture emerges. Content-free prose of matched length is worth nothing or less ( -0.05 to -0.15 ). One sentence from an unrelated 8B model, naming the subject and a few identifying details, matches the generator’s own plan in three to six times fewer conditioning tokens. The value sits where the referent is in doubt: writing down a name is worth +0.33 when it is right and +0.05 when it is wrong, and where the generator already succeeds the whole effect is -0.07 \pm 0.07 . Cutting a plan at the text encoder’s window leaves 272 of 286 images byte-identical, and a plan rewritten to fit that window ties a bare identification ( +0.003 \pm 0.044 ), so a plan contains nothing the identification does not already supply. Field-level interventions find only colour executed reliably, and a control writing the same sentences into the user prompt shows that limit belongs to the generator rather than to the planning channel. Two judges and two annotators replicate the central contrasts. A think block earns its keep as a specification of what to draw, not as a plan that is carried out.

[CV-219] Unlocking Fine-Grained Cell Representations via Soft-Interior Rim Dense-Grid Pooling in Pathology Foundation Models

链接: https://arxiv.org/abs/2611.19325
作者: Wenwen Lou,Mingyi He,Xinyi Guo,Yining Li,Weiming Chen,Jiawen Li,Lianghui Zhu,Zhiwei Yang,Xitong Ling
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Fine-grained cell representations are essential for characterizing cellular heterogeneity and tissue organization in computational pathology. Pathology foundation models (PFMs) provide powerful pretrained features for transferable tissue analysis. HoVer-Net uses task-trained segmentation and typing, whereas CellViT++ averages native-grid tokens assigned to detected nuclei. However, the former requires cell-specific training; the latter ignores fractional nuclear coverage and can mix nuclear content with neighbouring tissue. Therefore, we propose SRDP, short for Soft-interior Rim Dense-grid Pooling, a training-free geometry-aware readout. We apply the frozen patch projection at half stride, interleave four offset lattices, map the cell mask to fractional coverage weights, and concatenate whole-cell pooling with two inner boundary pools. Across four datasets and six frozen encoders, SRDP improves mean balanced accuracy by 5.68 percentage points over native token averaging across 24 settings, including 11.22 points on average for CoNIC and Lizard. SRDP converts a coarse frozen PFM lattice into a mask-aligned, fine-grained cell representation without updating the encoder.

[CV-220] SAR-TopoGuard: Training-Free Speckle-Aware Topological Verification for Structure-Preserving SAR-to-Optical Translation

链接: https://arxiv.org/abs/2611.19316
作者: Satabarto Sarkar,Debojyoti Ghosh,Subhranil Biswas,Kaustav Mandal,D. Sobya
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 10 pages, 4 figures, 7 tables. Accepted manuscript of the article published in IEEE Geoscience and Remote Sensing Letters (2026). \c{opyright} 2026 IEEE. Code: this https URL

点击查看摘要

Abstract:SAR-to-optical translators often synthesize detail that the radar measurement does not support, and pixel metrics alone cannot reveal it. We propose SAR-TopoGuard, a training-free inference-time layer for frozen generators: a stochastic sampler draws K candidates (Monte Carlo dropout for GANs, seed variance for diffusion), a speckle-aware verifier ranks them by gradient-magnitude correlation with the SAR input and a hallucination-frequency index, and a SAR-guided frequency gate admits high-frequency detail only where SAR edges support it. With all five hyperparameters frozen and three training seeds, we evaluate on 500 scene-disjoint SEN1-2 test patches, 500 patches of a held-out summer scene, and 200 cross-dataset SEN12MS winter pairs. The layer improves Pix2Pix by +0.060, +0.012, and +0.077 dB PSNR and by +0.0034, +0.0057, and +0.0135 SSIM, respectively (Wilcoxon p 1e-6), and it improves CycleGAN and CUT on all three sets (+0.013 to +0.076 dB). It raises edge precision at the cost of edge recall (edge F1 -1.3% on the test set). On the latent diffusion model COP-GEN, it adds +0.33 dB (base) and +0.16 dB (Beta) on the winter set, the only set with dual-polarization SAR. An ablation shows that the gate produces the gain for MC-dropout GANs, whose candidates are nearly identical, whereas verifier selection contributes up to +0.26 dB for diffusion, whose samples differ. The layer has zero trainable parameters and adds 0.4 ms per patch beyond the K=5 generator passes.

[CV-221] VLN-Flash: Parallel Action Decoding with Adaptive Execution for Vision-and-Language Navigation

链接: https://arxiv.org/abs/2611.19287
作者: Ruizhe Xie,Honghua Chen,Binglin Wang,Tong Guo,Xu Tang,Puti Yan,Shaoxiong Cai
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:When a vision-and-language navigation policy generates an action sequence autoregressively, the robot waits for a chain of dependent predictions. A small discrete action vocabulary makes it possible to score all short-horizon plans in parallel. We introduce VLN-Flash, a training-free decoder that evaluates the complete action-prefix tree in a single forward pass. Branch-specific attention and matched token positions preserve the original policy’s conditional action distribution for each prefix in exact arithmetic. We use the score margin between the stepwise-greedy prefix and its strongest alternative at each depth to decide how many actions to execute. The robot then executes the plan with the highest joint score at the chosen depth before observing again. On matched inputs, VLN-Flash achieves a nearly 3\times speedup in action decoding over native autoregression on Uni-NaVid and StreamVLN. Compared with the native autoregressive decoders, it improves success rate and success weighted by path length (SPL) on R2R-CE and RxR-CE, with success gains up to 2.7 percentage points. On a Unitree GO2, VLN-Flash improves navigation success and cuts the mean wait between actions by about 160 ms after subtracting network transmission time.

[CV-222] Generate Align and Restore: MLLM -Empowered Diffusion for Generalizable Low-light Image Enhancement

链接: https://arxiv.org/abs/2611.19285
作者: Kaisheng Pang,Feng Li,Chen Zhang,Junjun Jiang,Haoang Li,Sam Kwong,Runmin Cong
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 14 pages, 13 figures

点击查看摘要

Abstract:Low-light image enhancement (LLIE) is essential to numerous vision applications operating under adverse illumination. Despite substantial advances, real-world nighttime enhancement remains challenging due to the scarcity of strictly aligned day-night pairs and complex degradations involving extreme darkness and heterogeneous artificial light sources, leading to fragile generalization. To overcome these challenges, we propose MrLIE (MLLM-empowered reference Low-light Image Enhancement), a novel reference-based unsupervised LLIE framework that synergizes the generative and reasoning capabilities of multimodal large language models (MLLMs) with conditional diffusion. MrLIE employs the MLLM as a high-fidelity reference synthesizer and diagnostic critic, iteratively generating, diagnosing, and correcting normal-light references directly from degraded nighttime observations. Based on this, our method governs the enhancement process through two complementary mechanisms, in which a confidence offset module rectifies spatial-geometric inconsistencies for reference reliability and a structure-anchored injection scheme harmonizes aligned reference cues, low-light structural constraints, and ControlNet additive anchoring during the denoising process. Extensive experiments on diverse benchmarks demonstrate that MrLIE achieves superior perceptual quality, natural appearance, and robust generalization.

[CV-223] LC-ISP Benchmark: Learning Advanced ISP with Live and Capture Streams

链接: https://arxiv.org/abs/2611.19284
作者: Kai Chen,Junyu Lou,Jinhua Zhang,Weiyi You,Shuhang Gu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Code: this https URL

点击查看摘要

Abstract:High-quality paired RAW-to-sRGB datasets are essential for data-driven ISP, yet existing datasets often rely on cross-device collection, causing misalignment and failing to faithfully reflect real capture conditions. We therefore introduce LC-ISP, a large-scale benchmark built from the intrinsic gap between a camera’s live preview and offline capture streams. LC-ISP contains 53.9k spatially aligned groups, each comprising live RAW/live sRGB and multi-frame fused RGB/capture sRGB, and covers day, night, indoor, outdoor, and HDR scenes. It also provides capture-stream intermediates for modular supervision. We further develop a compact two-stage baseline that uses these intermediates as auxiliary targets. Under a 2.5-GFLOP budget for 448 \times 448 RAW inputs, it outperforms the strongest single-frame baseline by over 2 dB PSNR while improving texture, denoising, and color reproduction. LC-ISP supports realistic benchmarking, the design of supervision strategies, and the development of lightweight ISP models.

[CV-224] dit.xpu : Taming DiT Sparse Inference on Mobile SoCs

链接: https://arxiv.org/abs/2611.19261
作者: Jiayu Chen,Jinye Hu,Maoliang Li,Anqian Cen,Hailong Zou,Zihao Zheng,Xinming Wei,Guojie Luo,Jie Zhang,Xuanzhe Liu,Xiang Chen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Diffusion Transformers (DiTs) enable on-device image generation and editing but incur high attention costs at high resolutions. Existing sparse attention reduces computation, yet dynamic sparsity introduces preprocessing and irregular data accesses that poorly fit mobile NPUs. We present this http URL, a sparse DiT inference framework for heterogeneous mobile SoCs. this http URL combines tile-aligned sparse attention with anchor-based error compensation by exploiting slowly varying dense-sparse output residuals across denoising steps. Heterogeneous XPU collaboration overlaps NPU dense attention with GPU sparse calibration, while mask-guided dataflow reduces data transfers and layout conversion by overlapping data movement with computation. Across five DiT models on two Snapdragon smartphones, this http URL achieves up to 1.56x end-to-end denoising speedup over the Hexagon-NPU backend and 11.5x attention-kernel speedup over the fastest alternative, while preserving generation quality close to dense attention. Code is available at this https URL.

[CV-225] Size-Matched Nulls Reveal What Block Ablation Measures in MambaVision

链接: https://arxiv.org/abs/2611.19248
作者: Tahmina Kousar,Tehseen Zia
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 15 pages, 5 figures

点击查看摘要

Abstract:Hybrid vision backbones interleave Mamba state-space mixers with self-attention, and a block’s role is usually probed by removing it and measuring the damage. We test what that measurement means in MambaVision. On all 50,000 ImageNet validation images and three model sizes, every block ablation is paired with a size-matched null, the same perturbation with its channels shuffled or their signs flipped. In top-1 accuracy and in true-class rank, real and null damage are nearly indistinguishable in absolute terms: pooled over the blocks of MambaVision-T, the content effect lies within [-0.044,+0.016] percentage points (90% interval). Two attention-only controls show that the test detects content, with real ablations worse than their nulls by up to 4.2 points in DeiT-Small and 1.0 point in Swin-T. Part of MambaVision’s near-equivalence reflects how little damage its blocks do (0.11–0.34 points on average, against 1.7 and 2.1). Scaled by damage, its content share is comparable to Swin-T’s in accuracy, lower in rank, and far below DeiT-Small’s: single-block damage in MambaVision is small and mostly positional. Ablation also hides what blocks carry together: removing all blocks of one type compounds damage to 3.7–8.4 times the sum of single effects, far more than size-matched nulls, including nulls that keep the alignment between the blocks’ perturbations; for adjacent pairs, aligned nulls reproduce most of it. Freezing input-dependent selection in all Mamba mixers shows that it accounts for 20–32% of the Mamba contribution, mainly through the B and C projections. We turn these results into a protocol for causal attribution in hybrid networks.

[CV-226] GameWM: Toward Next-Generation Video Games with Interaction-Centric World Models

链接: https://arxiv.org/abs/2611.19247
作者: Bing He,Runze Wang,Fei Zha,Zhongyuan Wu,Ke Xu,Tao Zhuang,Yuxuan Xie,Zhongyuan Liu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 35 pages, technical report

点击查看摘要

Abstract:For more than half a century, video games have pursued better graphics, more natural interaction, and more engaging gameplay. Generative video models (VMs) now bear on all three. Neural networks already draw most of the pixels in shipped games, and recent world models let a lightweight game engine keep the world state while a VM turns the engine’s coarse proxy video into final frames. These systems, however, treat the VM as a renderer that only restyles motion the engine has already authored, which leaves most of its generative capacity unused and limits the diversity a game can offer. We present GameWM, a design for next-generation games built on a clear division of labor. The GameWM-engine keeps what must be exact: rules, state, camera, and the layout of the world, which also spares the video model from having to remember it. The GameWM-VM generates what is expensive to author by hand: the look of the world and the detail of the player’s interactions. One game mechanic can then be realized in many visually and physically distinct ways, while its outcome stays exactly as the engine computed it. To realize this division of labor, we build GameWM as a complete system: a data engine centered on interactions rather than exploration, an interface through which authored character sheets, scene art, and skill descriptions enter the model, and a model family from 5B to 33B parameters with a real-time causal variant for play. Together they give every player a fresh version of the same game: each player’s own character, world, and skills turn the same engine mechanics into a new experience, while every event keeps the timing and outcome the GameWM-engine computed.

[CV-227] Beyond Alignment: Supervising the Generator for EEG-to-Image Reconstruction

链接: https://arxiv.org/abs/2611.19232
作者: Sangmin Jo,Wootaek Jeong,Heung-Il Suk
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Visual brain decoding translates neural activity into human-interpretable images. While recent EEG-to-image reconstruction methods rely on aligning EEG signals with pretrained visual embeddings to condition an image generator, we empirically demonstrate that stronger cross-modal alignment do not translate into comparable gains in reconstruction quality. This indicates that improving the input condition alone is insufficient; the generative process itself requires explicit supervision. Although Representation Alignment (REPA) offers a solution by guiding diffusion transformers with pretrained vision encoder features, we find that EEG-conditioned generators struggle to follow these fixed targets. To address this mismatch, we introduce generation-derived supervision (GDS): an image-conditioned teacher that translates static image-encoder features into the internal representation space of a diffusion transformer. Our EEG-conditioned student generator closely tracks these dynamic targets during denoising, discarding the teacher at inference. Evaluated on THINGS-EEG and THINGS-MEG, our approach consistently outperforms conditioning-centric baselines, achieving state-of-the-art reconstruction fidelity.

[CV-228] Incremental Learning of Vision-Language Models via Task Subspace Projection and Dynamic LoRA

链接: https://arxiv.org/abs/2611.19213
作者: Hao Fu,Qian Feng,Hanbin Zhao,Jiahua Dong,Xu Tan,Lina Wei,Chao Zhang,Hui Qian
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Pre-peer-review version of an article accepted for publication in the International Journal of Computer Vision (IJCV)

点击查看摘要

Abstract:Recent pre-trained vision-language models often encounter a Multi-Domain Task-Incremental Learning (MTIL) scenario in practice, where a set of classes of multi-modal tasks arrive incrementally. Due to privacy concerns and memory constraints, MTIL with pre-trained models encounters forgetting of knowledge from old tasks, degradation of zero-shot transfer capability, and underfitting of new-task knowledge. To overcome these challenges, previous MTIL methods attempt to learn a discriminative cross-task identification (CTI) module and an effective new-task adaptation (NTA) module. However, current CTI modules suffer from severe task confusion between seen and unseen tasks, and NTA modules cannot adaptively balance the performance and parameter cost while incorporating task-specific knowledge. To alleviate the above dilemmas, we propose an effective and efficient TSP-DLoRA method for MTIL, which consists of Task Subspace Projection (TSP) and Dynamic Low Rank Adapter (DLoRA) modules. Specifically, our TSP module includes a task identifier classifier based on task-specific subspaces and a feature projection strategy that can determine the identifier associated with samples from both seen and unseen tasks. Our DLoRA improves the knowledge adaptation from new tasks by dynamically assigning Low Rank Adapter (LoRA) across transformer layers based on the task distributions. Experimental evaluations across 11 datasets, using three performance metrics, demonstrate the effectiveness of our proposed method. Code is available at this https URL.

[CV-229] LogT: Logically Think with Images for Visual Search NEURIPS2026

链接: https://arxiv.org/abs/2611.19200
作者: Yanjun Fu,Quanzeng You,Jiadong Guo,Yujie Lu,Sai Charitha Akula,Haiping Wu,Miao Liu,Sanghamitra Dutta,Lu Yuan
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注: NeurIPS 2026

点击查看摘要

Abstract:Thinking with images empowers vision-language models (VLMs) to use an external Zoom-in tool for fine-grained visual search. Recent works employ supervised fine-tuning (SFT) and reinforcement learning (RL) with tool-specific designs to enable such capabilities. However, achieving logical reasoning alongside accurate and efficient tool use remains a significant challenge. To address this, we shift the focus away from increasing RL complexity and instead identify SFT data quality as a key bottleneck. We then propose LogT (short for Logically Think with Images), a fully automated SFT data engine that explicitly enforces logic-chain-guided reasoning. LogT generates an SFT dataset featuring challenging, visually grounded questions and structured tool-use traces with human-like logic. We validate our approach by fine-tuning two base models, Qwen2.5-VL-7B and InternVL3.5-8B, on the dataset and evaluating them across seven benchmark test sets. We show that LogT-SFT based on Qwen2.5-VL-7B outperforms the state-of-the-art trained with both SFT and RL, despite using 3.1 \times less data and 120\times fewer GPU hours. Furthermore, applying a simple RL recipe without tool-specific design on LogT-SFT yields a model that surpasses prior methods with complex RL setups by a substantial 9.59-point margin. In addition to strong performance gains, our model exhibits higher tool-use accuracy and efficiency, as well as more structured and logically consistent reasoning traces, directly validating the effectiveness of our approach.

[CV-230] Ref-MIL: Plug-and-Play Posterior Refinement for Multiple Instance Learning in QSM-Based Brain Disease Classification

链接: https://arxiv.org/abs/2611.19193
作者: Ang Gaoa,Rui Jina,Yin Liu,Pengfei Rong,Feng Liu,G. Bruce Pike,Hongfu Sun,Yang Gao
类目: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
备注: 5figure, 6tables;

点击查看摘要

Abstract:Quantitative susceptibility mapping (QSM) is a novel quantitative MRI contrast for studying tissue susceptibility changes in brain disorders, including Parkinson’s disease (PD) and multiple sclerosis (MS). However, small cohorts and the lack of local annotations make developing deep learning models for QSM-based diagnosis challenging. In this work, we propose PoRef-MIL, a weakly supervised multiple instance learning (MIL) framework combining iterative slice relabeling and network retraining with a plug-and-play posterior refinement module (PRM). During training, a 2D dual-branch diagnosis network and a subject-level classifier were alternately trained for five rounds. The classifier was used to update slice pseudo-labels for retraining the diagnosis network. The PRM was applied during inference to refine slice probabilities through bidirectional posterior inference and calculate entropy-derived confidence from the refined probabilities, which was then used in sliding-window scoring to obtain subject-level predictions. PoRef-MIL was evaluated on an in-house PD cohort of 287 participants and a public MS cohort of 150 participants. Compared with eight classification methods, PoRef-MIL achieved the highest mean areas under the receiver operating characteristic curve of 0.942 and 0.872, and accuracies of 0.926 and 0.833, for PD and MS classification, respectively. Ablation experiments were conducted to validate the effectiveness of iterative relabeling and PRM, and the plug-and-play feature of the proposed PRM was also evaluated using four state-of-the-art MIL models. Overall, PoRef-MIL offers a promising approach to QSM-based disease classification in small cohorts without manual slice annotations.

[CV-231] Calibrating the Uncertainty of Self-Supervised Monocular Depth without Labels

链接: https://arxiv.org/abs/2611.19188
作者: Jun Seok Kim,Rui Huang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 5 figures, 4 tables

点击查看摘要

Abstract:Self-supervised monocular depth networks can predict a per-pixel uncertainty, which is mainly evaluated by how well it ranks pixels according to their error. Such metrics are insensitive to the magnitude of the uncertainty, so an uncertainty half as large as the actual error obtains the same score. In this paper, we investigate whether the predicted uncertainty is as large as the depth error, both on average (scale) and across confident and uncertain, near and far pixels (shape). On KITTI, the scale is miscalibrated for every uncertainty estimate we evaluate, and the error of MonoProb is nearly twice its predicted uncertainty. We propose MonoCal, which corrects both the scale and the shape without ground-truth depth. Scale calibration multiplies the uncertainty by a single constant measured from stereo pairs, requiring no retraining, and this constant closely matches the one fitted with ground truth. Since a constant cannot change the shape, we further retrain the uncertainty to predict where the depth of the network disagrees with that of a second self-supervised network. Experiments on KITTI show that MonoCal achieves better calibration across uncertainty levels and depth ranges than rescaling with ground truth, while still using a single image at inference. Code will be made publicly available.

[CV-232] Do VLMs Really Look at the Road? A Diagnostic Evaluation for Autonomous Driving

链接: https://arxiv.org/abs/2611.19184
作者: Shankar Gangisetty,Muskan Singh,C V Jawahar
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Vision-language models (VLMs) are increasingly being evaluated for autonomous driving through question-answering benchmarks, yet high benchmark accuracy does not necessarily reflect visually grounded reasoning. We present a diagnostic evaluation framework that systematically probes whether VLMs base safety-critical driving decisions on visual evidence or instead rely on language priors, answer biases, and dataset regularities. Our benchmark comprises two task categories, Adverse Weather and Junctions \ Intersections, with a nuScenes spatial reasoning extension, and evaluates open-source (general, driving, and spatial-reasoning) VLMs along four diagnostic protocols: image versus no-image evaluation, counterfactual weather sensitivity, linguistic robustness under paraphrasing and negation, and spatial reasoning. All ground-truth labels are derived programmatically from BDD100K and DriveLM annotations, enabling reproducible evaluation without manual labeling. Across all models, we find that competitive overall accuracy conceals systematic failure modes, including strong affirmation bias, limited sensitivity to counterfactual scene changes, near-chance performance on negated questions despite high paraphrase consistency, and confident driving decisions even without visual input. These results demonstrate that current open-source VLMs remain insufficiently grounded for safety-critical autonomous driving and underscore the need for diagnostic evaluation beyond aggregate accuracy.

[CV-233] WA-JEPA: Goal-Aware Predictive Feedback Action Refinement for a World-Action Model

链接: https://arxiv.org/abs/2611.19183
作者: Tao An,Weihang Liang,Tianqi Wang,Chuqing Cao,Wenming Tan,Kai Zhang,Haibo Lei,Tao Xie,Lijun Zhao,Ruifeng Li
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 16 pages, 5 figures, including supplementary material. Code available at this https URL

点击查看摘要

Abstract:JEPA-based world models predict action consequences in latent space, but representative planning approaches rely on candidate-action search, requiring repeated evaluation of action sequences and incurring substantial online computational cost. Direct action generation avoids this search, yet how to leverage latent predictions to guide action generation given task goals remains a key challenge in translating learned physical dynamics into effective decisions. We introduce WA-JEPA, a world-action model that couples action generation, action-conditioned latent prediction, and goal-aware feedback. The model first generates an action proposal and predicts the resulting future scenes and proprioceptive states, then refines the proposal through two complementary feedback paths, using the task goal as a reference. Goal-discrepancy residual feedback (GDRF) uses differences between the latent representations of predicted scenes and the task goal to guide chunk-wise action refinement. Future-scene context feedback (FSCF) preserves fine-grained predicted scene representations to guide subsequent action refinement. Jointly learned from offline demonstrations, this process iteratively updates the same plan before execution, making future predictions directly participate in action generation without candidate-action search. We evaluate short-horizon goal reaching, long-horizon tasks, and full-task execution. The results show that goal-aware predictive feedback supports competitive task performance while maintaining a compact model and low online computational cost. Code is available at this https URL.

[CV-234] EviRef: Evidence-Specific Memories for Logical Anomaly Detection across Capture Conditions

链接: https://arxiv.org/abs/2611.19172
作者: Yaoting Huang,Wenqi Cai,Yijin Wei,Jianjian Qin,Chunzhi Gu,Haoran Xie,Chao Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Logical anomalies arise when every component looks normal but their combination violates the scene rules. In practical inspection, these rules must hold under varying backgrounds, illumination, and imaging quality, with only a few dozen normal images per scene, whereas existing methods couple normality with appearance and merge different constraints into one reference. We model logical normality as decomposable constraints and propose EviRef, a training-free method that builds a separate reference memory for each family of evidence. Component prototypes built once per scene parse images under any capture condition into one label space, on which occupancy and relation constraints are verified. Semantic constraints are verified with whole-image embeddings independent of parsing, and additional evidence such as text is admitted through a task-level gate without modifying existing memories. On VID-AD, EviRef attains a mean image-level area under the receiver operating characteristic curve (AUROC) of 92.1% (88.9% with visual evidence alone), the highest among the compared methods under each of five capture conditions, with a standard deviation of 0.5. Even with memories built from only 20 normal images per task, it surpasses all compared methods that use 50 normal images per task.

[CV-235] From zero-target transfer to target-site personalization: multi-source training across fall-detection models

链接: https://arxiv.org/abs/2611.19156
作者: Jaeseung Ryu,Jae-Hun Jung
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Video-based fall detection could support timely assistance when a person cannot call for help. Unfamiliar cameras and scarce local labels challenge deployment. We examine whether joint training improves cross-dataset detection and subsequent local training. We analyze 2,697 clips from eight OmniFall datasets and one AI-Hub dataset. Positive clips contain a fall or collapse transition, while negative clips do not. We hold out each dataset in turn, training on the other eight. We evaluate three vision-language models (VLMs), Qwen, LLaVA, and InternVL, and three visual baselines, CLIP-temporal, VideoMAE, and Pose-TCN. We evaluate the stored source model without target data for training or selection. For each label budget, we update a separate copy using labeled clips from the held-out target dataset. We call this target-site personalization. We compare the Source-initialized model with Target-only, which skips source training but uses identical target clips and updates. Across nine targets, source training raises mean area under the receiver operating characteristic curve (AUROC) from frozen-model values of 0.821 to 0.912 for Qwen, 0.883 to 0.947 for LLaVA, and 0.939 to 0.962 for InternVL. With five target examples per class and 26 updates per run, 95% confidence intervals for the AUROC benefit over Target-only lie above zero for Qwen, CLIP-temporal, VideoMAE, and Pose-TCN. With full-target training, they lie above zero only for Qwen and LLaVA. Separate VLM analyses reveal threshold sensitivity and post-fall false detections. The benefit varies by model and label budget.

[CV-236] Puzzle: One Pass Is Not Enough for Open-Vocabulary Remote Sensing Segmentation

链接: https://arxiv.org/abs/2611.19146
作者: Wensong Xu,Siqi Lu,Jibao Yuan,Yongbin Zheng
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Open-vocabulary remote sensing image segmentation (OVRSIS) enables the segmentation of unseen classes without additional annotation or retraining. By analyzing the predictions of existing OVRSIS methods, we find that many misclassified patches coexist with visually similar, correctly classified patches of the same class within the same image. Unfortunately, most of these methods stop after a single inference rather than using their predictions to guide further inference. Motivated by this observation, we propose \textbfPuzzle, a plug-and-play iterative re-inference framework that combines prediction history with a visual-similarity prior to progressively refine patch-level attention. Puzzle employs a lightweight history-conditioned controller to regulate each patch’s contribution as an information source and its reliance on visually similar patches, while an adaptive stopping rule determines when to stop. It keeps the segmentation model frozen and reuses the same controller across different models and datasets without retraining. Extensive experiments demonstrate consistent performance improvements across diverse OVRSIS models and datasets, highlighting the strong transferability of Puzzle. Moreover, Puzzle can further improve the performance of SOTA methods.

[CV-237] DecoEdit: Beyond Flat Control for 3D-Aware Unified Image Editing

链接: https://arxiv.org/abs/2611.19121
作者: Xiaotang Gai,Jinbo Xing,Zijie Meng,Jiayi Yi,Jimin Xiao,Yixuan Qian,Yu Liu,Chaojie Mao,Zuozhu Liu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:While recent advances in multimodal image editing have achieved impressive results, current frameworks remain largely confined to ``flat" 2D controls (e.g., text prompts, bounding boxes, or masks). Precise 3D-aware geometric manipulation – such as coordinating an object’s location, physical scale, and 3D orientation while strictly preserving its fine-grained identity – remains an elusive challenge, typically suffering from severe detail degradation or requiring expensive per-instance fine-tuning. In this paper, we present DecoEdit, an all-in-one image editing framework that goes beyond flat control by seamlessly integrating 1D textual, 2D spatial, and 3D volumetric inputs under a unified architecture. First, to power this unified model, we establish a data curation pipeline and curate a large-scale dataset, namely DecoDataset, containing 1.6 million samples, encompassing diverse multi-dimensional (1D/2D/3D) tasks and their complex compositions. Furthermore, we establish a comprehensive image editing benchmark named DecoBench to systematically evaluate multi-task editing performance. Finally, to address the fundamental trade-off between complex spatial transformations and identity preservation, we propose a novel decoupled pipeline: DecoEdit factorizes the editing process by utilizing structural guidance from SAM-3D to anchor the target geometry, scale, and spatial orientation, while concurrently leveraging a detail-preservation path to lock in high-fidelity textures. Extensive experiments demonstrate that DecoEdit achieves state-of-the-art results. Notably, on the challenging 3D spatial pose and orientation control tasks, DecoEdit significantly outperforms leading closed-source models such as Seedream 5.0 in both geometric accuracy and identity consistency.

[CV-238] DEFDVO: Deformable Deep Visual Odometry for Equirectangular Images WACV2027

链接: https://arxiv.org/abs/2611.19095
作者: Burak Sevsay,Egor Bondarev,Erkut Akdag
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted to WACV 2027

点击查看摘要

Abstract:Omnidirectional ( 360^\circ ) cameras capture the full surrounding scene in every frame, providing a strong advantage for visual odometry (VO) over pinhole sensors with a limited field of view. However, their equirectangular projection (ERP) introduces latitude-dependent distortion, an artificial seam at the image border, and a sampling bias toward the poles. Existing 360^\circ VO methods address these challenges only partially. DEFDVO closes this gap with a patch-based deep VO method that operates directly on ERP imagery, combining spherically initialized deformable feature extraction, end-to-end seam-aware estimation, and spherical-gradient patch selection. The deformable feature extractor combines spherical geometric priors with learned sampling offsets, while seam-aware processing preserves ERP continuity, and spherical-gradient selection reduces polar sampling bias. Trained exclusively on synthetic TartanAirV2 ERP imagery, DEFDVO improves the best prior average absolute trajectory error (ATE) by 4.8\times (0.201 m to 0.042 m) on the TartanAirV2 evaluation set defined by 360DVO and achieves the lowest average ATE on both splits of the real-world 360DVO benchmark. Code is available at this https URL

[CV-239] FDAM: A Frequency-Direction-Aware Multi-Task Framework for Pavement Crack Grid Localization

链接: https://arxiv.org/abs/2611.19089
作者: Haiyang Xu,Bin Pan,Mengze Wang,Qiaoying Qu,Zhenwei Shi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Pavement crack grid localization remains challenging for thin and tiny cracks due to weak visual cues and background interference. Box detection often misses these cracks, while supervised semantic segmentation faces limited pixel-wise annotations and high labeling costs. In this paper, we propose a Frequency-Direction-Aware Multi-Task framework (FDAM) with box detection and grid classification branches. The core grid classification branch comprises Frequency-Domain Enhancement (FDE) and Directional Feature Refinement (DFR). FDE adaptively weights spectral bands and fuses reconstructed features with spatial features. DFR captures thin crack patterns through multi-scale strip convolutions, direction-sensitive modulation, and feature recalibration. Sparse positive cells and ambiguous local predictions motivate adaptive grid supervision. Adaptive Importance Reweighting emphasizes uncertain predictions and local label transitions, with focal supervision for all cells. Our contribution is to couple the box detection and grid classification branches through shared feature learning for pavement crack grid localization. We validate FDAM on our RoadDistress2026 dataset of thin and tiny cracks and further evaluate it on public datasets Crack500 and DeepCrack. The code is available at this https URL

[CV-240] SearchChord: Prompt Preselection and Field Recomposition for One-Step Image Editing

链接: https://arxiv.org/abs/2611.19076
作者: Boyang Deng,Yuzhen Lu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 3 figures, conference

点击查看摘要

Abstract:Finding effective prompts and seeds for one-step image editors requires balancing edit fidelity, source preservation, and computational cost. To address this challenge, we explore a large collection of candidate edits, where each candidate is defined by a unique prompt-seed pair, and generate corresponding images on demand. We present SearchChord, a training-free framework that leverages a large pool of such candidates, combining text-based preselection, image verification, and latent-field recomposition. Our ablation studies across pool sizes up to 200 candidates show that a pool of 100 offers a cost-effective operating point. Compared with feeding all 100 candidates into recomposition, TextSim preselects a subset of 10 candidates. Rather than uniform averaging, we incorporate image-verifier scores to weight each candidate during recomposition. Experiments on PIE-Bench and the proposed CropWeed-EditBench reveal a trade-off: candidate selection improves target-text alignment, while recomposition better preserves unmodified source content. The oracle verifier achieves the best performance in both preservation and target-text alignment. Practical verification partially recovers semantic alignment, but recomposition still compromises edit fidelity, including in agricultural image edits. These results motivate text-first preselection and stronger verification to balance preservation, fidelity, and cost. Code and data will be made publicly available.

[CV-241] Hard or Mislabeled? Failure Screens Enrich Vision Benchmarks for Label Errors

链接: https://arxiv.org/abs/2611.19044
作者: Haolin Guo
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 28 pages, 5 figures, 16 tables

点击查看摘要

Abstract:Hard vision benchmarks often keep images that strong classifiers get wrong. Since a classifier is also marked wrong when the label is wrong, such failure screens may collect label errors. We measure this on ImageNet’s twice-relabeled validation set with 103 classifiers. Among the 1,892 images that all 66 supervised classifiers get wrong, 61% have a wrong label, versus 11.1% overall. The fewer images a screen lets through, the more of them are mislabeled, along approximately one curve. Fit on one model family, this curve predicts the screens of two held-out model families with a mean error of 2.4 percentage points, and replays of ImageNet-D’s and ImageNet-Hard’s screens within 0.1 and 6.3 points. On new pools, ImageNet-V2 and CIFAR-10, enrichment at the strictest screens is stronger than the curve predicts: on ImageNet-V2, by human relabeling, 67% of the images that all 66 classifiers get wrong are mislabeled, against 46% predicted. Screens that let models choose among only some of the classes, as ImageNet-A’s does (200 of 1,000), fall below the curve. By ReImageNet, 32% of ImageNet-Hard’s ImageNet images have no correct label. ScreenAudit, which comprises an audit tool, a script that rebuilds corrected answer sets for the 5,464 decided ImageNet-Hard images (of 5,646) from ReImageNet’s gated release, and per-image audits of 1,198 benchmark images, will be released later.

[CV-242] Role-Separated Diffusion Transformers: Global Rule Isolation and Deterministic Sampling Manifold Compression

链接: https://arxiv.org/abs/2611.19009
作者: Zhiguang Liu,Yi Shang
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 10 pages, 5 figures, 3 tables. Github Repo: this https URL

点击查看摘要

Abstract:We present Role-Separated Diffusion Transformers (RoSeDiT), a new diffusion transformer architecture that isolates the global rule from the image tokens. RoSeViT showed in visual reasoning that a vanilla ViT spreads a shared rule over all L tokens at an entropy of order Ld , whereas an enforced controller route reduces it to order d . In class-conditional DiT and SiT, the class and the noise level are explicitly conditioned by adaLN, so a RoSeViT-style controller channel on the ViT side is insufficient to establish the full role separation. By introducing the adaLN listening mechanism, which makes adaLN compute under the controller’s command, RoSeDiT brings the entropy theorem and role separation into the latent diffusion domain. Under the same SiT training and evaluation protocol at the same level of parameter count, RoSeDiT surpasses SiT at all four model scales at 400K steps, and RoSeDiT-L/2 even surpasses SiT-XL/2’s 17.2 FID with 15.34 and 32% fewer parameters. RoSeDiT-XL/2 surpasses SiT-XL/2’s 7M-step FID of 8.3 at 1.4M steps with 7.98, using one-fifth as many training steps. We further observe that models with the same denoising loss can have a large FID gap, which indicates a misalignment between the denoising objective and the image manifold. We formalize this phenomenon into Deterministic Sampling Manifold Compression (DSMC), a two-stage training schedule with which RoSeDiT-XL/2 surpasses SiT + REPA’s 4M-step FID of 5.9 at 2.4M steps with 5.84.

[CV-243] Probing Where Surprise Happens: Localized Violation-of-Expectation Evaluation of Video World Models

链接: https://arxiv.org/abs/2611.18979
作者: Yanbin Xu,Ming bo Cai
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Image and Video Processing (eess.IV)
备注: 8 pages, 4 figures

点击查看摘要

Abstract:Violation-of-expectation (VOE) benchmarks evaluate intuitive physics in video world models by testing whether physically impossible events produce greater prediction errors than matched possible events. However, conventional VOE evaluations aggregate prediction errors across space and time, making it unclear whether successful discrimination reflects model’s understanding of physical violations at the locations and time where they occur. To address this limitation, we introduce localized VOE evaluation, a framework that measures prediction error differences between possible and impossible events within spatial regions and temporal intervals associated with physical violations. We compare global, temporally localized, and spatiotemporally localized evaluations of V-JEPA models pretrained on developmental egocentric videos or large-scale natural video datasets. Our results reveal that global and localized judgments can diverge in both directions: global aggregation can either obscure correct localized discrimination or produce apparent discrimination without corresponding localized evidence. During continued pretraining on developmental egocentric videos, agreement between global and localized judgments changes non-monotonically, while confidence in localized judgments increases despite modest improvements in accuracy. Together, these findings establish localized VOE evaluation as a complementary diagnostic for assessing intuitive physics in video world models and demonstrate that global prediction errors do not necessarily reflect the spatiotemporal evidence of physical violations.

[CV-244] Effects of Conventional Adaptive and AI-Powered Upscaling on the Legibility of License Plates

链接: https://arxiv.org/abs/2611.18949
作者: Hany Farid,Joshua D. Bucci,Emily A. Cooper
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Small details in digital images – a face, a street sign, a license plate – are routinely upscaled in hopes of making them interpretable. Upscaling options now range from conventional interpolation, whose operations are spatially invariant and fully explainable, to AI-powered methods that synthesize detail potentially absent from the source. Here, we ask which of these techniques most reliably helps human readers. In two online experiments, we measured legibility of 255 US license plates ranging in resolution from 128 to 8 pixels tall, as well as their legibility after upscaling by eight different methods. Legibility of the un-upscaled plates collapsed from ceiling to floor across a narrow band of plate heights (32 to 10 pixels). Conventional upscaling techniques (bilinear, bicubic, and Lanczos interpolation) gave modest gains of about 5-9 percentage points within this range. Nearest-neighbor interpolation, however, was harmful, rendering unreadable many plates that were readable without any upscaling. Advanced methods diverged: SwinIR, an open-source AI enhancer, raised performance by up to 38 points relative to no upscaling, while an adaptive denoiser-based image prior and commercial AI upscaler tended to impair legibility. We found no strong evidence of consistent character hallucinations across methods. Decomposing the task, conventional interpolators mainly helped readers count characters, whereas SwinIR alone made the characters themselves more identifiable.

[CV-245] Lasso: Towards Sustainable AI-Generated Media Detection with Intel QAT On-Chip Compression

链接: https://arxiv.org/abs/2611.18895
作者: Darsh Asher,Joshua Kalyanapu,Kaushal Mhapsekar,Bita Aslrousta,Amna Hassan,Debayan Sengupta,Samira Mirbagher Ajorpaz
类目: Hardware Architecture (cs.AR); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Synthetic text, images, and video are rising. Software-only detectors are power-hungry, too slow, and costly, and do not scale with demand. When software cannot scale, historically hardware often comes to the rescue. We perform the first reverse engineering of a compression accelerator–Intel QuickAssist Technology (QAT) microarchitecture and–find different amount of time on each, a novel data structure-dependent property. Using this information, we then build Lasso, a low-cost, high-performance detector for AI-generated content that turns the accelerator into a measuring instrument. For text detection, Lasso achieves 100% accuracy on original DetectRL text and retains 92% average accuracy after paraphrasing, compared with 78% for Binoculars and 55% for DetectGPT. Across original, paraphrased, and perturbed text, it achieves 82% average accuracy, outperforming all five baselines. With topic, length, and formatting matched, accuracy reaches 95% for instruction-tuned generators, including when the tested generator is unseen. On RAID, accuracy averages 82% when trained on one generator and tested on another, and 74% across eight held-out domains. Detection takes 0.74 microseconds, a speedup of at least 83,784x across five baselines. At the same time, Lasso consumes over 99% less energy per document than both Binoculars and DetectGPT. At the same time, Lasso detects AI-generated videos, software code, and images with 67.5%, 70% and 78% accuracy, respectively, with low latency and power opening avenues for practical real time and edge device detection of multimodality AI artifacts for the first time. No other compared technique supports detection across all four modalities: text, video, images, and code. Subjects: Hardware Architecture (cs.AR); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV) Cite as: arXiv:2611.18895 [cs.AR] (or arXiv:2611.18895v1 [cs.AR] for this version) https://doi.org/10.48550/arXiv.2611.18895 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[CV-246] CISyn: Human Mesh Recovery in Close Interactions with Physics-Simulated Synthetic Data

链接: https://arxiv.org/abs/2611.18894
作者: Yufan Deng,Lorenza Prospero,Christian Rupprecht
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:3D human mesh recovery (HMR) in close-interaction scenarios is hindered by the lack of large-scale, accurately annotated training data: real capture is expensive, and existing synthetic datasets rarely include physical contact. We introduce CISyn, a large-scale synthetic dataset purpose-built for close human-human interactions, generated via physics simulation and Blender-based rendering. By simulating collision dynamics between SMPL-X-rigged humanoids, CISyn naturally produces the contact, mutual occlusion, and pose entanglement that characterize close-interaction scenes, with no manual annotation effort. The released set provides 25,000 sequences with accurate 3D annotations. We further propose an uncertainty-aware loss that allows models to learn more effectively from the heavily occluded examples in our data. Augmenting state-of-the-art HMR training with CISyn and replacing the regression loss with our uncertainty-aware variant yields consistent improvements on close-interaction benchmarks, validating CISyn as a scalable complement to real-world data. The dataset, generation pipeline, and code are available at this https URL.

[CV-247] Bring Back Language to Actions: Restoring VLM capabilities in VLA Models via Parameter Merging EACL2027

链接: https://arxiv.org/abs/2611.18891
作者: Mohamed Salim Aissi,Olivier Sigaud,Laure Soulier,Nicolas Thome
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: Accept at EACL 2027 (Main)

点击查看摘要

Abstract:Vision-Language-Action models (VLAs) are obtained by fine-tuning pre-trained Vision-Language Models (VLMs) on robot demonstration data, enabling action prediction from visual observations and language instructions. However, this fine-tuning induces catastrophic forgetting, substantially degrading the semantic and visual grounding capabilities of the original VLM. We address this by merging the parameters of a VLM and its VLA derivative without retraining either model. Concretely, we extract the task vector (the difference between VLA and VLM weights) and decompose it via Singular Value Decomposition. For each component, we learn a scalar mixing coefficient controlling how much of the VLA adaptation is retained, optimised to jointly preserve action prediction and vision-language capabilities. Experiments show that our merged model performs well on both axes simultaneously, consistently outperforming established merging baselines.

[CV-248] RAVEN-Bench: A Paired EO-IR Video QA Benchmark for Aerial Multimodal Understanding NEURIPS2026

链接: https://arxiv.org/abs/2611.18880
作者: Yu Hu,Jianyang Gu,Yue Cao,Hao Liu,Kangnan Wang,Jozsef Hamari,Zheng Liu,Mohsen Zardadi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted at NeurIPS 2026 (Datasets and Benchmarks Track)

点击查看摘要

Abstract:Multimodal large language models (MLLMs) have advanced rapidly in image and video understanding, yet their ability to reason over paired aerial electro-optical and infrared (EO-IR) videos remains poorly understood. Such video poses a harder test than standard RGB video QA: targets are often small, evidence may appear only briefly, viewpoints change with the platform, and useful cues may appear in only one stream. We introduce RAVEN-Bench (Reasoning over Aerial Videos with EO-IR aligNment), a paired EO-IR video QA benchmark for low-altitude aerial multimodal understanding. RAVEN-Bench contains 100 temporally aligned EO-IR video pairs and 1,200 human-authored four-way multiple-choice questions, each independently audited for answer keys, modality requirements, evidence intervals, and group validity. It supports EO-only, IR-only, and paired EO+IR evaluation, and uses hierarchical capability levels and structured question groups to diagnose paired-input gain beyond the stronger unimodal condition, unimodal reliance, target consistency, and evidence-chain coherence. We evaluate frontier, open-weight, and remote-sensing MLLMs, showing that aggregate accuracy can overstate evidence-supported EO-IR reasoning. RAVEN-Bench provides controlled modality conditions and group-level diagnostics for auditing whether progress on RGB video benchmarks transfers to aerial cross-spectral understanding. Code and data are available at this https URL and this https URL.

[CV-249] UnSPAD: Uncertainty-aware Single Photon Burst Photography WACV

链接: https://arxiv.org/abs/2611.18869
作者: Sachin Shah,Kalyan Bhetwal,Levi Burner,Christopher A. Metzler
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Accepted at The Winter Conference on Applications of Computer Vision (WACV) 2027

点击查看摘要

Abstract:We introduce UnSPAD, an uncertainty-aware burst photography framework capable of fusing and denoising arbitrarily long sequences of binary single photon avalanche diode (SPAD) measurements. Unlike prior methods that require fixed burst lengths or rely on powerful generative priors, UnSPAD combines learned motion estimation with uncertainty-informed measurement fusion and denoising to produce temporally consistent, hallucination-free reconstructions. UnSPAD first forms an initial scene estimate by averaging flow-aligned binary frames. UnSPAD then uses per-pixel uncertainty maps to perform uncertainty-aware denoising; the denoiser is informed about which regions accumulated the fewest observations (and thus contain the most relative shot noise) due to motion, occlusion, or alignment errors. This two-step reconstruction pipeline helps mitigate hallucinations—outside of the structure implicitly introduced by the flow estimates, the initial scenes estimates are entirely prior-free. Experiments on real and simulated SPAD and quantum image sensor (QIS) benchmarks demonstrate state-of-the-art reconstruction quality, improved temporal consistency, and superior color fidelity relative to the state of the art.

[CV-250] FiReIQ: Field-Level Confidence Estimation with Reasoning and Image Quality in Document Extraction

链接: https://arxiv.org/abs/2611.18865
作者: Xiao Liu,Yage Wang,Qi Yu,Giri Iyengar
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: Work in Progress

点击查看摘要

Abstract:Vision-Language Models (VLMs) have demonstrated strong performance in structured document extraction, yet their deployment in high-stakes applications requires reliable confidence estimates for individual fields. Existing uncertainty estimators often aggregate signals across an entire generation, obscuring field-specific reasoning uncertainty while overlooking visual degradations that may lead to confident extraction errors. To address these limitations, we introduce FiReIQ, a field-level confidence estimation framework that combines field-aligned reasoning signals from a single VLM generation with local and global image-quality features through quality-conditioned visual fusion. To facilitate systematic evaluation, we further introduce FiReIQ-Bench, a benchmark spanning financial receipts and legal contracts. Experiments with two reasoning VLMs show that FiReIQ consistently improves confidence discrimination, selective prediction, and probabilistic accuracy over existing estimators. These gains extend to private financial service data, where FiReIQ improves AUROC by 6.8% and coverage at 95% precision by 10.6% over an existing baseline. Together, these results highlight the potential of FiReIQ to reduce manual review while enabling more reliable document automation.

[CV-251] Guided Group Featurizers (GGF): Improving Block-Sparse Featurizers with Object-Guided Learning

链接: https://arxiv.org/abs/2611.18861
作者: Waldemar Chang,Andrii Krutsylo
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Block-sparse featurizers (BSFs) use groups of learned directions to capture the multidimensional structure of visual concepts. However, grouping alone does not ensure that information about a concept is consolidated within a single group. We introduce Guided Group Featurizers (GGF), a novel object-guided extension that uses object annotations for initialization and training while preserving the BSF architecture. We assess whether each group’s strongest response regions share an object type, part or material across images. Blinded reviews give GGF groups an average semantic coherence score of 94.1\pm0.9 %, versus 83.0\pm1.9 % for SLiCS and 65.3\pm2.2 % for supervised Vanilla BSF. GGF’s consolidation advantage extends to unguided categories: one selected group per category covers 65.6\pm2.4 % of target-containing crops, versus 58.2\pm2.2 % for supervised Vanilla BSF, at approximately 5% false-positive rates for both. Across COCO and VOC, GGF also improves single-group object Capture and localization over Vanilla BSFs, SLiCS and Concept-SAE. Its groups produce more focused and reproducible object responses while supporting recognition and retaining associations under rotation, scale and brightness changes. GGF also uses fewer learned parameters than Concept-SAE and shorter recorded VOC training loops than both supervised alternatives. These findings support object guidance for organizing more interpretable BSF groups and show that response concentration extends to categories excluded from guidance.

[CV-252] chnical Report of Aurigas 1:10 Scale Autonomous Vehicle Robot

链接: https://arxiv.org/abs/2611.18855
作者: Marya Sabouri Dodaran,Ali Taheri Garakani,Mandana Shabanzadeh Nasrollah-Abadi,Yousef Ashrafi,Mohammad Mahdi Azimi,Erfan Nekouei Ghahfarokhi,Amirhossein Kheiri Holighi,Seyyed Sobhan Hosseini Hajibakandeh,Matin Jamshidi,Amirhossein Gholizadeh Behbahani,Mohammad Hossein Moaiyeri
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV); Systems and Control (eess.SY)
备注: 62 pages, 32 figures, 11 tables, 1 algorithm. Technical report of Team Auriga’s 2025 1:10 scale autonomous vehicle platform (Robotics and Intelligent Automation Laboratory, Shahid Beheshti University)

点击查看摘要

Abstract:This report presents the design and implementation of Auriga Robotics’ 1:10 scale autonomous vehicle, developed at Shahid Beheshti University’s Robotics and Intelligent Automation Laboratory. The platform integrates a custom mechanical chassis, a distributed embedded system based on ESP32 microcontrollers, and a Jetson Orin NX module for high-level perception and control. The embedded layer manages sensor interfacing, real-time data acquisition, and actuator control, while the Jetson handles environment perception through deep-learning–based lane, sign, and traffic light detection, as well as decision-making and trajectory control. System communication is achieved via structured UART links, ensuring synchronized operation between perception and control subsystems. The vehicle’s software stack incorporates ROS, modular firmware design, and multiple control algorithms—Stanley, PID, and Pure Pursuit—to achieve stable autonomous navigation in structured driving scenarios. The overall system demonstrates robust performance and reliability in urban-like environments, serving as a modular research platform for future autonomous driving studies.

[CV-253] Scaling Channel-Adaptive Self-Supervised Learning

链接: https://arxiv.org/abs/2611.18849
作者: Alice V. De Lorenci,Seungeun Yi,Theo Moutakanni,Piotr Bojanowski,Camille Couprie,Juan C. Caicedo,Wolfgang M. Pernice
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Recent advances in self-supervised pre-training of foundation models for natural images have made them a popular choice for various visual systems and applications. Self-supervised strategies are also promising in non-RGB scientific imaging domains such as in biology, medical and satellite imagery, but their broader application is hampered by heterogeneity in channel composition and semantics between relevant datasets: two datasets may contain different numbers of channels, and these may reveal distinct aspects of an object or scene. Recent works on channel adaptive strategies report substantial advantages for those that account for variable channel compositions without sacrificing the ability to jointly encode channels; yet, how these strategies behave at scale remains unclear. We here show that, surprisingly, trained across large-scale datasets, independent-encoding of channels outperforms joint-encoding methods by a substantial margin. We validate this result along an extensive set of experiments on various datasets from cell microscopy to geospatial imagery. Our DINO BoC approach sets a new state-of-the-art across challenging benchmarks, including generalization to out-of-distribution tasks and unseen channel combinations. We open-source code and model weights for a new general-purpose feature extractor for fluorescent microscopy.

[CV-254] LongCat-Video-Avatar 1.6 Technical Report

链接: https://arxiv.org/abs/2611.18822
作者: Meituan LongCat Team:Xunliang Cai,Yiqi Chen,Feng Gao,Xiaoyu Kong,Le Li,Hongyu Liu,Xiaoming Wei,Tianyu Yang,Yong Zhang,Sijie Zhao
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Recent advances in audio-driven avatar video generation have enabled increasingly realistic synthesis, yet achieving precise lip-audio synchronization, natural, human-like motion, and visual consistency over long sequences remains challenging. We present LongCat-Video-Avatar 1.6, which builds on LongCat-Video-Avatar 1.5 to improve expressive naturalness and reduce error accumulation during extended generation, e.g., exaggerated lip movements and overly red, temporally unstable lip coloration, enhance the naturalness of facial expressions and coordinated head and body movements. These improvements are driven by a refined post-training pipeline that integrates reinforcement learning, improved classifier-free guidance strategies for DMD-based few-step distillation, long-video rollouts, and on-policy distillation. Together, these refinements improve expressive realism and enable the generation of videos exceeding 10 minutes while maintaining visual quality and temporal consistency. Arena evaluations show that LongCat-Video-Avatar 1.6 outperforms its predecessor and competing commercial audio-driven avatar models, demonstrating the effectiveness of the improved post-training pipeline.

[CV-255] RescueVLM: Model-Agnostic Vision-Language Framework for Rescue Drone Imagery

链接: https://arxiv.org/abs/2611.18813
作者: Niklas Voigt,Tim Sibum,Julien Meine,Niklas Digakis,Hartmut Surmann
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 7 pages, 6 figures

点击查看摘要

Abstract:A single unmanned aerial vehicle (UAV) mission over an incident site produces hundreds of images, far more than a rescue commander can review under time pressure, and human screening of destruction imagery carries a documented psychological cost. The two obvious remedies both fail in search and rescue (SAR) practice: cloud vision-language model (VLM) services are excluded because the imagery must not leave the incident network under GDPR and German civil-protection (BOS) rules, and fine-tuned specialist detectors age quickly while retraining and re-certifying them exceeds most SAR organizations resources. We present RescueVLM, an open-source, training-free and model-agnostic triage tool built on SketchVLM: a general-purpose open-weight VLM runs fully on-premise and is swapped for a newer model through a simple dropdown selection, so triage quality grows with the underlying models because all task knowledge lives in prompts and an output contract rather than in trained weights. One natural-language query over a mission folder returns a ranked shortlist with per-image justifications, editable SVG overlay annotations, and a GPS map of the hits. In a qualitative evaluation on German fire-brigade field data from three real deployments and a live fire exercise, a single query condenses an entire survey mission into a handful of annotated frames that a commander can review at a glance.

[CV-256] CARET: Clinical Assertion-Aware Residual Calibration for Medical Image-Text Alignment

链接: https://arxiv.org/abs/2611.18811
作者: Yifan Jin,Tao Liu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Medical vision–language models align radiographs with free-text reports, but radiology language distinguishes explicit findings, negation, uncertainty, and omission. Treating these report states as interchangeable during feature adaptation can change image predictions without supporting visual evidence. We introduce CARET, a parameter-efficient image-side adaptation method that retains pretrained image and text encoders and uses paired positive and negative clinical statements to define a low-dimensional semantic correction space. A bounded residual adjusts image features along these directions, while assertion-aware supervision distinguishes explicitly present and absent findings from uncertain or unmentioned ones. At inference, the adapter requires only an image; textual references remain fixed. In zero-shot classification, CARET achieves 67.80% accuracy on CheXpert 5 \times 200 and 65.20% on RSNA. These scores exceed the published CLEFT-LoRA results by 1.06 and 0.27 percentage points, respectively. Comparisons to literature baselines are informative but not controlled for differences in pretraining and evaluation protocols. Ablations across eight components identify the four-state distinction and assertion objective as major contributors, with task-dependent effects for nonmention preservation and pair alignment.

[CV-257] From Retouched Images to Paired Videos: A Synthetic Benchmark for Face Video Retouching

链接: https://arxiv.org/abs/2611.18791
作者: Peiyuan Zeng,Fengchuang Xing,Shiming Wu,Hongwei Yue,Chao Zhou
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Paired unretouched/retouched face videos are costly to produce because professional editing must remain consistent over time, while releasing identifiable face footage raises privacy, consent, and governance concerns. We present DeepVFR, an initial controlled synthetic benchmark-not a new retouching architecture-for paired face video retouching. Starting from paired synthetic DeepFRR images, Wan2.2 generates mild-motion source clips, and backend-specific face transfer constructs synchronized proxy retouching targets, followed by automatic validity screening and manual visual inspection. DeepVFR contains 1,024 pairs of 150-frame clips at 30 fps, evaluated under an identity-disjoint protocol with explicit frame correspondence; release metadata will record the construction route. Within-benchmark baselines, external checkpoints applied to DeepVFR, qualitative analysis, and a 17-participant ranking study establish reference behavior in the controlled domain. The constructed targets exhibit a substantial low-preference tail in the human study. Because both streams inherit synthetic content and the targets additionally inherit commercial-API or face-transfer artifacts, the benchmark neither replaces artist-retouched captured video nor establishes synthetic-to-real generalization. DeepVFR provides a starting point for paired supervision and controlled evaluation. Data, individual study rankings, and training logs will be made publicly available upon acceptance. Broader motion coverage and real-video validation remain future work.

[CV-258] Same Evidence Different Verdicts: Delivery of CT Images and DICOM Metadata Changes VLM Quality-Control Decisions

链接: https://arxiv.org/abs/2611.18790
作者: Nicolas Avelar
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 4 pages, 2 figures, 1 table. Submitted to IEEE ISBI 2027

点击查看摘要

Abstract:Vision-language models (VLMs) are being proposed as automated checks on clinical imaging inputs, where a scan and its DICOM metadata can disagree; a truncated CT series wrongly admitted to volumetric follow-up, such as lung-nodule growth assessment in screening, biases the measurement. A natural safeguard is to let the model judge the image before it sees the metadata. We find this backfires. On 60 NLST chest CT pairs with coverage defects whose ground truth is computed from scan geometry, a false header that Claude Haiku 5.5 resisted when shown with the image (2% of its detections overruled) overturned 91% of them when it arrived after the model’s answer (Nemotron-3.5-VL 43% to 100%). With the order reversed, the image overturned 98-100% of the header’s errors (Claude models): whichever source came second won. This is not generic sycophancy: a neutral “please confirm” reversed 0 of 1,072 correct refusals, and the effect persisted when the header arrived as a tool result, as in agentic pipelines. One warning sentence cut sequential reversal to at most 9%, and the pattern held on 24 real mismatched pairs. Contributions: (i) evidence that VLM input-check decisions on identical evidence depend on delivery; (ii) a testbed with exact geometric ground truth; (iii) a validation protocol for clinical VLM input checks. Code, prompts and all call logs: this https URL

[CV-259] hird-Place Solution for the UCF UrbanTwin V2X-Real Track at ECCV 2026: Synthetic Point Cloud Generation and Sim-to-Real 3D Object Detection ECCV2026

链接: https://arxiv.org/abs/2611.18778
作者: Antong Deng,Yuxin Zhang,Zhongxiang Xing,Zhenzhen Lin,Jiong Mu
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 1 figure, 4 tables. 3rd place, Track 7 (V2X-Real), UCF UrbanTwin Sim2Real LiDAR Challenge, 6th DriveX Workshop at ECCV 2026

点击查看摘要

Abstract:This report presents a complete technical solution for the UrbanTwin V2X-Real Sim2Real LiDAR Challenge, covering both synthetic point cloud generation and 3D object detection. For data generation, we adopt per-axis quantile calibration. Public V2X-Real training data are used to construct the real reference distribution, and independent one-dimensional quantile mappings are estimated for the x, y, and z axes of the base synthetic point clouds. The calibrated point clouds are then resampled using a fixed random seed and clipped to the valid spatial range, producing 50 frames with 2,000 float32 XYZ points per frame. For detection, we use a two-stage Voxel R-CNN (VoxelRCNN) detector trained with RangeLDM-enhanced synthetic point-cloud data, point clouds downsampled to 50k points per frame, and reconstructed pedestrian data. Training is performed on four NVIDIA GeForce RTX 4090 GPUs with a global batch size of 16 for 40 epochs. Final testing uses direct single-model inference with the original VoxelRCNN model. On the official evaluation server, the normalized Realism score is 0.7905 and the normalized Detection score is 0.5583, with mAP@0.5 of 0.3908, BEV mAP@0.5 of 0.4219, and mAP@0.7 of 0.1047. According to the official weighting rule, the final combined score is 0.6 x 0.5583 + 0.4 x 0.7905 = 0.6512. This report documents the complete implementation details of data construction, model training, inference, submission formatting, and official evaluation.

[CV-260] Quantum-inspired orthogonal adapters for certifiably robust medical image classification

链接: https://arxiv.org/abs/2611.18775
作者: Olga Tsurkan,Laia Domingo,Victoria Orlova,Asel Sagingalieva,Iris Zalaudek,Raffaele Dante Caposiena Caro
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Quantum Physics (quant-ph)
备注:

点击查看摘要

Abstract:Medical image classifiers require not only high accuracy but also reliable predictions under small input perturbations. Certified robustness provides mathematical guarantees of prediction stability, but preserving these guarantees when adapting pretrained models to new tasks remains challenging. We propose a fine-tuning approach based on quantum-inspired orthogonal butterfly layers that preserves the Lipschitz bounds of pretrained certified networks. Inspired by quantum circuits of reconfigurable beam splitters, these layers preserve orthogonality by construction using only \mathcalO(n\log n) trainable parameters. Their structured design enables efficient classical implementation and offers a path toward future quantum implementations with logarithmic circuit depth under suitable connectivity assumptions. Controlled experiments on MNIST and MedMNIST demonstrate improved certified robustness compared with conventional alternatives and, in the DermaMNIST transfer, greater resistance to the tested attacks than low-rank adaptation (LoRA). We then apply it to trichoscopic image classification for alopecia diagnosis and alopecia areata activity assessment. By adapting a certified ImageNet-pretrained network, butterfly fine-tuning achieves approximately twice the certified accuracy of conventional fine-tuning and LoRA, with comparable classification accuracy on the same backbone and data split and similar empirical resistance to the tested adversarial attacks. Larger robustness guarantees, including certificates covering certain image-processing transformations, can be obtained at the cost of reduced accuracy. These results highlight the potential of quantum-inspired butterfly layers for parameter-efficient adaptation of medical image classifiers while preserving the mathematical guarantees underlying certified robustness.

[CV-261] Agri-MRAG : Parameter-Enhanced Multimodal Retrieval-Augmented Generation for Agricultural Visual Question Answering

链接: https://arxiv.org/abs/2611.18771
作者: Fuyong Liu,Xiantao Lei,Xining Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Multimodal Retrieval-Augmented Generation (MRAG) has shown promise for Agricultural Visual Question Answering (Agri-VQA), but existing methods still face challenges in retrieving relevant knowledge while avoiding redundancy and in integrating retrieved evidence beyond simple prompt concatenation. To address these issues, we propose Agri-MRAG, a multimodal retrieval-augmented framework for agricultural VQA. Specifically, we first introduce a semantic-consistency-guided clustering mechanism that combines multimodal clustering and weighted relevance scoring to select high-confidence knowledge subsets and adaptively determine the retrieval size k . We then design a parameterized knowledge injection strategy, where retrieved agricultural knowledge is encoded into LoRA adapters and merged into the feed-forward network (FFN) layers of the MLLM, allowing external knowledge to influence generation through parameter-level adaptation. In addition, a perplexity-guided (PPL-guided) routing mechanism is employed to switch between implicit parametric reasoning and explicit retrieval-assisted generation according to model uncertainty, balancing inference efficiency and answer reliability. For evaluation, we use PlantVillageVQA as a broad agricultural benchmark and construct PlantVQA, a region-specific forestry and fruit crop VQA dataset containing 7,514 images and 15,028 expert-verified question-answer pairs. Experiments show that Agri-MRAG achieves ROUGE-L scores of 41.21 and 46.58 on the two datasets, outperforming mRAG by 11.70 and 4.51 points, and VisRAG by 8.80 and 3.39 points, respectively.

[CV-262] AerialOcc: Unified 3D Aerial Perception without Task-Specific Head

链接: https://arxiv.org/abs/2611.18770
作者: Fan Liu,Xiang Gu,Baoyu Lv,Chuanyi Zhang,Liang Yao,Yuye Ma,Yuhui Zheng
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Aerial perception systems are increasingly expected to support diverse tasks, yet existing solutions remain fundamentally tied to task-specific outputs. Detection, segmentation, and other perception objectives are usually learned either by separate models or by sharing a backbone with multiple task heads, leaving the model itself unaware of the common structure underlying these tasks. We are interested in asking: can a single perception representation support different levels of understanding without being designed for any specific task? We notice that occupancy is a promising task-agnostic representation for drone perception. By modeling the spatial existence and semantics of the environment, it naturally bridges different perception granularities, from coarse scene understanding to fine-grained object localization. Therefore, we introduce AerialOcc, a unified occupancy-based perception model that enables multiple drone perception tasks from a single inference process, without maintaining task-specific prediction heads. To facilitate this direction, we construct AerialOcc-20K, a large-scale multi-class occupancy dataset for aerial environments. Extensive experiments show that AerialOcc achieves competitive performance across diverse perception tasks while preserving a unified architecture. Our findings demonstrate that occupancy can serve as a universal perception interface, offering a new perspective toward flexible and scalable visual intelligence for autonomous drones from a common 3D scene representation. Models and data are publicly available at this https URL

[CV-263] Frame First Decode Once: Object-Centric Coordinates and Single-Pass Decoding for System-1 World Models

链接: https://arxiv.org/abs/2611.18761
作者: Abu Anas Ibn Samad
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Robotics (cs.RO)
备注: 12 pages, 6 figures, 7 tables. Code, frozen configurations and per-seed results will be released

点击查看摘要

Abstract:Single-pass primitive decoders such as LocateAnything and VFlash predict all quantized coordinate tokens of a geometric primitive in one forward pass, enabling fast, System-1 visual planning. We study what matters when such a decoder serves as a System-1 world model: a fast action-conditioned predictor that ranks candidate pushes by their predicted outcome. On ManiSkill3 PushT-v1 (512 held-out test families, 8,192 rollouts) and on human demonstrations from lerobot/pusht (2,384 test windows), we compare decoders, training losses, rigid-body projections and coordinate frames, using matched checkpoints, hierarchical bootstrap confidence intervals and Holm correction. Two factors of similar size dominate. First, expressing inputs and outputs in the SE(2) frame of the object’s initial pose lowers decision regret by 1.9-2.2x for each of five decoders, and the object-frame model trained on 128 families matches the world-frame model trained on 512. Second, single-pass decoding outperforms iterative decoding: in the object frame, single-pass heads reach 0.020-0.021 regret at 1.6-1.8 ms per decision, against 0.046 for 10-step DDIM diffusion and 0.059 for KV-cached autoregression at 10.3 and 12.9 ms. The choice among single-pass heads has no measurable effect on regret, and our registered hypothesis that the frame effect exceeds the decoder spread is not supported. Soft-token expectation with an end-to-end 2D Kabsch projection gives the lowest trajectory error (11.35 px on ManiSkill3, 3.78 px on lerobot/pusht), exact rigid-body consistency, and lower regret than diffusion at every budget from 4 to 50 DDIM steps, at 3-29x lower latency.

[CV-264] Conv-Reluplex: A Verification Framework For Convolution Neural Networks

链接: https://arxiv.org/abs/2611.18753
作者: Jin Xu,Zishan Li,Miaomiao Zhang,Bowen Du
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 6 pages, 8 figures, 5 tables. Published in SEKE 2021

点击查看摘要

Abstract:In recent years, machine learning has demonstrated impressive performance in many real-world tasks, especially in computer vision and natural language processing. However, to apply them in safety-critical systems one needs formal guarantees on the neural network outputs. The Reluplex tool is proposed to verify the safety of deep neural networks (DNNs), and in case the DNN fails to give a correct output, can generate adversarial examples. Since the tool can only handle DNNs, it is necessary to extend the tool to process image data. Therefore, in this paper, we propose the Conv-Reluplex framework, which is designed to verify the convolutional layer and pooling layer in convolutional neural networks (CNNs), and generate adversarial examples when classification is misguided. We conduct several experiments on MNIST to evaluate our approaches. The results show that the original CNN is improved using the adversarial examples generated by our tool, and the precision of classification can be increased significantly.

[CV-265] From Hosted Service to Field-Owned Infrastructure: Open Self-Hosted Real-Time Camera-Trap Alerting with Regional Open-Weight Models

链接: https://arxiv.org/abs/2611.18751
作者: Paul Fergus
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 31 Pages, 8 figures

点击查看摘要

Abstract:Cellular camera traps can alert field teams within minutes, supporting anti-poaching, threatened-species protection and human-wildlife conflict mitigation. Most systems use central servers, sending images of people and sensitive species locations outside the collecting organisation and often across national borders. We ask whether this capability can run on infrastructure owned by field organisations. Trap Tracker RT is an open-source, self-hosted application that receives images by email or locally via FTP/SFTP, classifies them with a project-specific model and sends alerts through a mail server chosen by the organisation. We evaluated DeepFaune-SSA v1.2, pairing MegaDetector v6 with a DINOv3-based classifier for 33 sub-Saharan taxa, against three open alternatives on 31,131 out-of-distribution images from Kenya, Namibia and Malawi, independent of its training data. DeepFaune-SSA achieved image-level species accuracy of 0.901 (95% CI 0.894-0.908), versus 0.530 for the global SpeciesNet classifier and 0.556 and 0.547 for YOLOv10x and YOLO26x detectors trained on the same regional data. DeepFaune-SSA had the highest recall for all 26 species evaluated. The single-stage detectors missed animals in about a quarter of images; SpeciesNet assigned species labels in only 57%, usually falling back to coarser labels otherwise. In bench tests of the fully local path, alerts were issued about 6 s after upload completion, including with slowed and interrupted uploads. The two-stage pipeline processed 19 images per minute on eight CPU threads without a graphics processor. Trap Tracker RT is used by five organisations in the United Kingdom and Kenya, including for human-elephant conflict mitigation. We discuss how data paths, rather than inference location, determine data sovereignty, the responsibilities of self-hosting and the ethics of real-time monitoring.

[CV-266] Feature-Space Latent Forecasting for Endoscopic Surgical Video Understanding

链接: https://arxiv.org/abs/2611.18750
作者: Songtao Hu,Hang Su
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Surgical workflow assistance is inherently anticipatory: systems must reason about how procedures evolve, not just recognize the current frame. However, many prior approaches perform prediction in pixel space by synthesizing future frames, which primarily focuses on appearance fidelity rather than the semantic representations required for downstream reasoning. To address this limitation, we propose a feature-space semantic forecasting framework that models temporal dynamics directly within the representations of a frozen Vision Foundation Model (VFM). Specifically, we build a hierarchical target representation from multi-layer VFM features, replace future-frame feature tokens with learnable \texttt[MASK] tokens, and train a masked transformer to forecast the corresponding future latent features from observed context, thereby formulating procedural evolution as masked future feature prediction. Experiments on public laparoscopic benchmarks demonstrate consistent improvements in phase anticipation, multi-label tool presence prediction, and remaining-time estimation over RGB-based forecasting baselines, while enabling a plug-and-play representation interface for lightweight downstream heads.

[CV-267] Video-as-Experience: Learning Computer Use via Audio-Visual Understanding

链接: https://arxiv.org/abs/2611.18747
作者: Guangzhi Sun,Zhonghao Zhai,Yixuan Li,Chao Zhang
类目: Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Multimodal large language models (LLMs) have enabled capable computer-use agents that perceive graphical user interfaces, reason over instructions, and execute actions. However, existing approaches to learning from video demonstrations typically convert videos into screenshot-based action trajectories, rather than directly leveraging their rich audio-visual information. We introduce VEX (Video as EXperience), a framework that uses audio-visual understanding to unify learning from external demonstrations and reflecting on an agent’s own behavior. Following a watch-act-rewatch paradigm, the agent extracts procedural guidance from tutorial videos containing visual demonstrations and spoken explanations, executes tasks, and verifies its progress by rewatching its execution as video. We instantiate VEX with video-SALMONN-VEX, a 9B-scale audio-visual computer-use agent built on the video-SALMONN 2 architecture with a Qwen3.5-9B backbone. We further introduce Action-Grounded Guide Optimization (AGGO), a reinforcement learning method that optimizes video-derived guidance for downstream action prediction, and augment AgentNet, OSWorld, and CUA-Gym with over 24K real-world tutorial videos. Experiments on OSWorld and a complementary subset of CUA-Gym demonstrate consistent improvements, with video-SALMONN-VEX achieving a 22% relative increase in task success over its Qwen3.5 backbone on OSWorld. Moreover, its generated guidance improves Gemini-3.5-flash by 10% relatively on OSWorld. These results highlight the potential of audio-visual understanding as a unified interface for procedural learning and self-verification in computer-use agents.

[CV-268] VTG-RAG : Long-Video Reasoning via Query-Adaptive Coarse-to-Fine Sampling and Temporal Graph Retrieval

链接: https://arxiv.org/abs/2611.18746
作者: Yao Lu,Yunhao Wang,Yaxin Gao,Feifei Teng,Chenxiao Liu,Zhe Tang,Jiaheng Wei,Zhaowei Zhu,Shanqing Yu,Qi Xuan
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Long-video reasoning requires a system to localize sparse evidence over extended temporal sequences, preserve temporal structure, and generate answers despite substantial multimodal distractors. Existing video retrieval-augmented generation methods typically build indexes from uniformly sampled video segments. In long-video settings, this design readily introduces irrelevant context and is unreliable for questions with explicit temporal constraints. We propose VTG-RAG, a framework for long-video question answering that combines query-adaptive coarse-to-fine sampling with temporal graph retrieval. First, a two-stage sampling strategy selects core segments with high mean query relevance or high intra-segment relevance variance. The framework then integrates subtitles, speech transcripts, visual descriptions, entities, and relations from these segments into a temporally ordered multimodal graph. Finally, it performs temporally filtered, two-pass retrieval and reasoning across frame, text, and graph evidence spaces. A lot of experiments have proven the effectiveness of the VTG-RAG. Specifically, we evaluate the framework on Video-MME and LongVideoBench. On LongVideoBench, the reported Qwen2.5-VL-32B and 72B results improve from 48.16% to 50.56% and from 52.13% to 53.32%, respectively, with 24 answer-input frames. These correspond to gains of 2.40 and 1.19 percentage points. The ablation experiments of VTG-RAG further support that the contributions of each component are very important.

[CV-269] Neural Latent Gaussian Process for 3D reconstruction of Mouse Brain Lipids

链接: https://arxiv.org/abs/2611.18744
作者: Daniel Trejo Banos,Artiom Casapu,Luca Fusar Bassini,Halima Hannah Schede,Giovanni D’Angelo,Gioele La Manno,Ekaterina Krymova
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 24 pages, 14 figures, 5 tables

点击查看摘要

Abstract:We address the problem of reconstructing continuous three-dimensional lipid profiles of the mouse brain from lipid measurements acquired only on a sparse set of two-dimensional tissue sections. Using MALDI-MSI measurements from coronal sections of two densely sampled and six sparsely sampled mouse brains, registered to the Allen Mouse Brain Common Coordinate Framework (CCF), we seek to interpolate the measured hundreds of lipid profiles across the unobserved voxels of the reference brain. We introduce a Neural Latent-Space Gaussian Process (LGP) architecture for this task and critically evaluate its ability to reconstruct spatial lipid distributions, using held-out mouse brain sections to assess generalization. We propose a Gaussian Process prior for the distribution of low-dimensional lipid profiles across the mouse brain (with Matérn or a Riemann manifold kernel). This allows us to encode 3D coordinates in a small number of latent 3D volumes that are then mapped to particular lipids using a multilayer perceptron (MLP). We performed an exhaustive analysis over a range of algorithms from simple linear regression, XGBoost, MLP to state of the art models such as Neural Transcriptomic Field, graph-based SPA3D for spatial omics reconstruction and DEEPSPATIAL. We additionally evaluate GPLFR as an alternative Gaussian-process-based approach. Across 173 measured lipids, LGP outperforms all evaluated baselines, including recent spatial-omics models and GPLFR, achieving a mean correlation of 0.613 and demonstrating the effectiveness of our approach for sparse 3D lipid interpolation. Comments: 24 pages, 14 figures, 5 tables Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.18744 [cs.CV] (or arXiv:2611.18744v1 [cs.CV] for this version) https://doi.org/10.48550/arXiv.2611.18744 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[CV-270] Recognizing Events Is Not Enough: Why Streaming VLMs Struggle in Compositional Instruction Triggering

链接: https://arxiv.org/abs/2611.18743
作者: Yehua Huang,Penglei Sun,Xiang Li,Jinhao Zhou,Jiawen Wen,Qingqing Xiao,Zihan Yang,Wenqi Wang,Haolan Chen,Yu Xu,Xiaowen Chu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Recognizing individual events does not guarantee correct execution of compositional instructions in streaming video. For example, given the request “Notify me when the pot boils after someone turns on the stove and leaves the kitchen”, existing models may recognize the individual events yet notify too early or miss the intended moment, failing to track how their temporal conditions are satisfied over time. To evaluate this gap, we introduce CoTriggerBench, which composes annotated video events into temporally constrained instructions and verifies notification timing and silence through executable event plans and source-event replay. To address this challenge, we propose a real-time VLM agent, TEMPI. TEMPI decomposes instructions into atomic event queries and temporal constraints, and uses Scenario Memory to continuously maintain task progress as event evidence arrives, allowing earlier observations to support subsequent decisions beyond the detector’s visual window. Even when no notification is required, relevant events continue to update the scenario state. On CoTriggerBench, TEMPI achieves 44.4% instruction-level success, outperforming the strongest closed-source and fine-tuned open-source VLM baselines by 15.6 (Qwen3.8-Omni-Flash-Realtime) and 18.1 (MOSS-VL-Realtime) percentage points, respectively. To support efficient on-device streaming inference, we further optimize the streaming inference pipeline by overlapping partial visual encoding and LLM prefix prefill with frame arrival. On an 8 GB RTX 5060, our optimized detector implementation achieves a 4.90 \times speedup.

[CV-271] arget-Free Autocalibration of Stereo PIV Camera Systems via Persistent-State Motion Representations

链接: https://arxiv.org/abs/2611.18741
作者: Nabil Madali
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Stereo Particle Image Velocimetry (PIV) recovers three-component velocity fields by triangulating tracer particles observed by synchronized cameras. It requires an accurate extrinsic calibration, which is still obtained by inserting a physical target into the flow and must be repeated whenever the rig is moved. We present a target-free method that calibrates a stereo PIV rig directly from particle image sequences. Because tracer particles are visually indistinguishable, we match motion rather than appearance: building on CUT3R, each camera stream is processed by a recurrent transformer whose persistent state accumulates particle motion over time. The state is decoded into per-camera optical flow and queried by a cross-view decoder that regresses bidirectional point correspondences, from which the relative pose is recovered with a five-point RANSAC solver. We train on a new synthetic stereo PIV dataset with ground-truth poses and evaluate on standard PIV optical flow benchmarks and on 200 real rig placements. Our model achieves lower flow error than existing PIV networks, more accurate cross-view correspondences than COTR, and higher pose accuracy than the strongest detector-matcher pipelines on both synthetic and real data, without any calibration target.

[CV-272] yphoonWorld: A Residual Token World Model for Probabilistic Forecasting of Typhoon Imagery

链接: https://arxiv.org/abs/2611.18740
作者: Xiaoyu Chen,Tianyi Li,Xuanyi Zhang
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 15 pages, 3 figures, 7 tables

点击查看摘要

Abstract:A satellite forecast of a typhoon should give a low-error picture of the next two days, together with an ensemble whose members are coherent fields and whose spread responds to the storm at hand. A small direct regression network already gives the lowest squared error on held-out storms, so a generative model has to earn its cost through the part of the future that the regression smooths away. We present TyphoonWorld, a discrete-token world model that generates only that part. Conditioned on tokens of the regression forecast, the latest frames and the best-track state at forecast time, it samples residual fields lead by lead, and members add these residuals to the regression center. We compare it under one protocol with per-pixel Gaussian networks, a diffusion model of the same residuals, residuals from other storms and an unconditional token world model, on 276 forecast windows from 46 held-out storms of Digital Typhoon. TyphoonWorld matches the squared error of the regression, lowers the CRPS of the area colder than 220 K, a direct measure of deep convection, by 14 to 39% against every other ensemble, and ties its spread to forecast error as closely as the Gaussian network, with a correlation of 0.78 across windows. It uses 56% of the parameters of the unconditional world model.

[CV-273] AITHER: Authority-Informed Evidence Routing for Mixed-Authority Vision-Language Reasoning

链接: https://arxiv.org/abs/2611.18735
作者: Xinjin Li,Haoyue Zhang,Xinyi Gou,Yuruo Gong,Yuanjian He,Muzi Li,Yujian Long,Yu Ma,Yiliu Xu,Tianxin Zhou
类目: Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)
备注:

点击查看摘要

Abstract:Reliable vision-language reasoning requires a control decision that conventional grounding methods leave implicit: which input is entitled to determine the answer? Auxiliary text may be a defeasible claim about an image, a binding task instruction, or compatible context. Correcting every image-text conflict toward the image therefore repairs false claims but overwrites valid task rules. We formalize this setting as mixed-authority vision-language reasoning and introduce AITHER, an Authority-Informed Text-Image Hierarchy for Evidence Routing. Its final policy, the Semantic-Frame Router, is a training-free black-box control layer that (i) compiles explicit task directives, (ii) recovers standardized spatial frames from metadata-scrubbed pixels, (iii) resolves the semantic polarity of remaining notes, and (iv) routes among frozen observable candidates with fail-closed semantics. This ordered decomposition converts authority arbitration from an unconstrained prompt preference into a reproducible inference program. Across 1,620 MATRIX predictions, AITHER reaches 95.1% accuracy, exceeding the strongest task-and-frame baseline by 9.1 points with a 95% record-cluster interval of 6.8-11.5 points and no harmed records. It further reaches 100.0% on 1,800 controlled FRAME predictions and 93.3% on GQA-MATRIX. The results show that explicit authority routing can simultaneously improve false-claim correction, preserve task-defined answers, and eliminate clean-context overcorrection.

[CV-274] Is It Still There? Place-Registered Verification Makes VLMs Reliable Scene-Change Detectors

链接: https://arxiv.org/abs/2611.18734
作者: Jinjie Xie,Baihua Li,Qinggang Meng
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 23 pages, 13 figures, 11 tables. Code: this https URL

点击查看摘要

Abstract:A robot that returns to a space must tell which objects were added, removed or moved since its last visit, while its trajectory, viewpoints and lighting differ between the two visits. We find that multi-image vision-language models (VLMs) fail at this task mainly because they cannot find the place where an object was. A VLM recognises what each frame shows but cannot check whether an object is still at its place in the other capture. It therefore reports objects that only one capture shows and misses most removed and moved objects. We propose STILL, which reformulates change detection as object verification at a registered place. 3D geometry registers the place of each change hypothesis in the other capture, and one VLM answers, from a scene view and a local place view, whether the object is still at that place, trained with supervision of the same form. Under the official SceneDiff protocol, STILL raises the main metric obj/sc AP from 24.0 to 37.8, and from 11.6 to 20.6 on the egocentric kitchen subset without any kitchen training data. It improves F1 by more than 21 points over a VLM fine-tuned on the same annotations and adds only about 11% to the runtime.

[CV-275] LiNO: Lightweight Network for Occupancy Estimation

链接: https://arxiv.org/abs/2611.18733
作者: Ruan Bispo,Tim Brophy,Reenu Mohandas,Anthony Scanlan,Ciaran Eising
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 13 pages / 6 figures / 10 tables

点击查看摘要

Abstract:Occupancy estimation for automated vehicles is a crucial task that builds a 3D map using vehicle sensors. This map allows the vehicle to identify drivable areas and various obstacle classes. The task involves complex subtasks such as depth estimation, object classification, and segmentation to align semantic information with the 3D grid. Moreover, producing these estimations from multiple sensors demands significant computational power and memory, as high-dimensional data fusion requires complex feature extraction and processing. Consequently, most existing methods rely on large and computationally intensive architectures that are not suitable for real-time or real-world applications. We present a real-time, lightweight architecture that leverages camera and radar sensors through a residual multi-stage head. Unlike prior methods that rely on heavy camera-only blocks, our architecture employs camera-radar fusion and lightweight components. Additionally, we introduce a task-specific data augmentation strategy that enhances data diversity and improves generalisation. Our method achieves state-of-the-art (SOTA) performance, attaining 45.7 mean Intersection over Union (mIoU) on Occ3D and 41.9 Ray-level Intersection over Union (RayIoU) (41.9) on OpenOcc datasets, running at 17.2 frames per second (FPS) and using only 5.3 GB of memory, compared with 4.1 FPS and 7.7 GB for STCOcc.

[CV-276] ShiftGuard: Evaluating Confidence and Feature Distance for Risk-Controlled Image Classification

链接: https://arxiv.org/abs/2611.18732
作者: Mitra Varun
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:A classifier that is usually right still makes confident mistakes, and a deployed system needs some way to decide when to answer and when to abstain. Cheap post-hoc scores such as maximum softmax probability (MSP) and predictive entropy are the standard answer. Distances in feature space are a popular alternative, and fusing several signals into one score is a natural next step. We ask whether adding a feature distance signal to MSP and entropy improves error detection, and whether it increases the share of predictions that can be accepted at a certified error rate when only a small labelled calibration set is available. We train a CIFAR-adapted ResNet-18 from scratch on CIFAR-10 (92.92% test accuracy, one seed) and compare five reliability scores: MSP, entropy, distance to the predicted-class centroid, a logistic fusion of MSP and entropy, and a logistic fusion of all three signals. Acceptance thresholds are chosen on a held out calibration set with Clopper Pearson upper bounds, Bonferroni corrected over a prespecified family of 315 rules at 95% simultaneous confidence. Confidence based scores were best. Error AUROC was 0.9230 for MSP, 0.9233 for entropy and 0.9235 for MSP plus entropy, against 0.8814 for the three signal fusion and 0.8224 for distance alone. At a 5% risk target with 1,000 calibration images, entropy accepted 84.82% of test images with 2.05% observed error, while three signal fusion accepted 80.61% with 2.21%. With 250 calibration images, neither distance nor the three-signal fusion could certify the 5% target at all. The results come from one seed, one dataset and one architecture, with no corruption experiments, so they say nothing about distribution shift detection. Within these limits, the distance signal did not earn its added complexity.

[CV-277] SCALE: Towards whole-slide counterfactual explanations in pathology

链接: https://arxiv.org/abs/2611.18729
作者: Weiyi Qin,Yingci Liu-Swetz,Shiwei Tan,Yi Wang,Mohammad Alexanderani,Dimitrios Korentzelos,Peizi Li,Soo Yeon Sohn,Xintong Wang,Hao Peng,Yue Guo,Yuzhe Yang,Han Zhao,Hao Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Cancer diagnosis relies on histopathological evaluation, where tumor morphology guides disease classification and informs treatment decisions. Deep learning (DL) models excel at predicting diagnostic properties from histopathology images. Unlike typical medical images featuring small critical areas, histopathological images are textural, with relevant texture spanning the entire image; this makes it challenging to explain DL predictions by locating relevant image regions. Counterfactual explanations offer a promising alternative by asking “what changes in the image could have altered the model prediction.” To achieve this, we propose SCALE, a framework for generating such counterfactual explanations for histopathological DL models. Across nine counterfactual directions spanning multiple cancer types, SCALE generates whole-slide counterfactual explanations that consistently shift predictions toward the intended target while maintaining biological feasibility. Blinded assessment by six board-certified pathologists showed higher morphological plausibility for SCALE (with a win rate of 91.7% over baselines) and promising counterfactual effectiveness (with a target consistency of 96.7%).

[CV-278] Latent Fusion: Channel Stacking for Structure-Preserving Image Editing

链接: https://arxiv.org/abs/2611.18727
作者: Mohit Dulani
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注: 12 pages, 5 figures, 4 tables. Code: this https URL

点击查看摘要

Abstract:Instruction-based image editing models must condition on the source image, either by sequence concatenation, which appends source tokens to the target sequence (FLUX.1 Kontext), or by channel-wise concatenation, which stacks the source latent onto the noisy latent and widens the input projection (InstructPix2Pix, FLUX.1 Fill). The FLUX.1 Kontext report found channel-wise concatenation worse in initial experiments at 1K resolution with a 12B model and a large corpus. We explore channel-wise concatenation as the lower-cost option for a common production regime: structure-preserving edits that change appearance (sky, exposure, foliage, cleanup) but not layout, about 1,400 training pairs, a 4B non-distilled backbone, and 2K frames. It keeps the sequence at its original length, where sequence concatenation doubles it and roughly quadruples attention cost, and needs only single-scale classifier-free guidance (two forward passes per step instead of three). Starting from FLUX.2 Klein 4B base, we widen only the input projection and train both schemes with the same trainer, data, and hyperparameters. Channel-wise concatenation keeps the output pixel-aligned with the source; in our runs sequence concatenation, which must learn the correspondence between 10,880 target and 10,880 source tokens from fewer than 1,500 pairs, displaced structure, an observation we report without claiming a general ranking. On 29 held-out images the editor reaches 20.3 dB PSNR and 0.198 LPIPS against the human edit (unedited source: 15.0 dB), and counterfactual conditioning tests show it largely learned a global grade of the source rather than a selective edit, which we report as the open problem of this recipe. We also compare against an SD3.5-Medium editor trained with the same surgery and report the full training recipe, including failure modes. Code: this https URL

[CV-279] Detection Is Not the Finish Line: Multiscale Decoding and Constraint-Aware Global Association for Multi-Camera 3D Tracking

链接: https://arxiv.org/abs/2611.18725
作者: Abdullah Naeem,Anav Katwal,Ayon Day,Md Tamjidul Hoque
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:A query-based multi-view 3D detector can transfer well to a new calibrated camera network and still produce badly fragmented tracks: on the AI City Challenge 2026 Track 1 validation set, the frozen Sparse4D detector behind the official challenge winner splits each Person identity into a median of eight fragments, and half of the gaps between fragments exceed the horizon of a causal tracker. We show that most of the remaining headroom of such a system lies after detection. Keeping the public detector frozen, we add (i) class-wise source selection between two input scales, which recovers small and distant people from a 0.75x pass while leaving large objects at the scale the detector was trained for; (ii) a geometry-to-appearance bridge that projects every 3D box into all calibrated views, matches it to 2D person detections, and aggregates robust track-level re-identification descriptors; and (iii) constraint-aware consolidation, which combines a causal short-gap stitch with offline centroid-linkage clustering under two hard constraints, temporal exclusion and metric 3D reachability, plus class-aware long-gap stitching for rigid objects. Every association stage changes object identities only. On the challenge evaluation server the system scores 62.3947 3D HOTA (DetA 57.4501, AssA 56.8131, LocA 80.1272), first on the general leaderboard snapshot of 30 October 2026 (second at the time of writing, after a competitor’s detector update) and 5.63 points above our reproduction of the official winner; the constrained association stage alone contributes +3.73 HOTA through a +5.49 AssA gain at unchanged DetA. Controlled fixed-box validation studies, a constraint ablation, sensitivity curves, and a catalogue of negative results support the design.

[CV-280] Acting by Intent Reacting by Sight: Visual Conditioned Reflexes for Real-Time VLM Agents

链接: https://arxiv.org/abs/2611.18723
作者: Qiancheng Zhou,Wenhua Zhang
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Vision-Language Models (VLMs) have shown strong capabilities as generalist agents that perceive, reason, and act in interactive environments. Yet real-time environments such as video games demand both deliberate control for exploration and split-second reactions to fast-moving hazards. Existing agents struggle to handle both: relying solely on direct VLM actions is too slow for sudden events, whereas delegating execution to fixed scripts leaves the agent blind at the moment of action, leading to frequent misfires whenever the scene shifts. Humans overcome this by pairing normal conscious control with conditioned reflexes-while the mind handles deliberate actions and shapes reflex rules, the reflex itself still relies on live vision to make the final instant decision. Emulating this mechanism, we propose ViCoR (Visual Conditioned Reflex). Alongside normal gameplay actions, the main VLM defines and updates lightweight situational guidance for recurring events, while a fast visual reflex branch (invoking the VLM in no-thinking mode) inspects live local views with motion cues to make the final sub-second action decisions. By offloading 67.8% of high-frequency reactions to the reflex branch (680 ms per check), ViCoR reduces average per-action decision latency by 2.6x (from 14.0 s to 5.4 s) and cuts token consumption by 36.9%. Across 17 diverse 2D and 3D games in the GameWorld benchmark, ViCoR achieves state-of-the-art performance, improving task success rate by +11.8% (from 52.9% to 64.7%) and average normalized progress by +13.9% (from 66.5% to 80.4%) over the strongest deliberative VLM and script baselines. In fast-paced real-time games, ViCoR refines its reflexes across episodes to boost progress by 1.67x, and even with a low-effort main planner, surpasses high-effort deliberative baselines by +8.1% at one-third of the token cost.

[CV-281] Editable Video Context: Mid-Context KV-Cache Reuse for Context-Managing Video LLM s

链接: https://arxiv.org/abs/2611.18718
作者: Dayu Wang,Shasha Xu
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Context Language Models let a model edit its own context, while Suffix Cache Reuse makes edits to text efficient by adjusting the positions of cached keys for surviving tokens. We extend these ideas to video language models with multimodal rotary position embeddings (M-RoPE). We represent a long video as timestamped segments. After viewing a sparsely sampled version, the model selects segments for decoding at 4x the frame rate, and the remaining segments are removed from its context. Our rebuild method reuses the KV cache of surviving segments by rotating their keys with a single offset shared by the temporal, height, and width components of Qwen2.5-VL’s M-RoPE. A content-hashed cache also avoids repeating vision-encoder computation for those segments. In preliminary experiments with Qwen2.5-VL-7B on synthetic needle-in-a-video-haystack questions over public-domain films, the method reduces prefill tokens and end-to-end rebuild latency relative to full recomputation while usually producing the same answer. Densifying the segment containing the answer improves accuracy when that segment is known. When the model selects segments itself, accuracy does not improve: it rarely identifies a needle that was absent from the sparse sample. We release the code, raw logs, and scripts used to reproduce the paper’s results.

[CV-282] LVRE: A Local Visual Reasoning and Editing Agent for Optical Chemical Structure Recognition

链接: https://arxiv.org/abs/2611.18717
作者: Bowen Lin(University of Minnesota),Lei Fan(Shanghai Jiao Tong University)
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 14 pages, 4 figures, 3 tables. Accepted at IFTC 2026

点击查看摘要

Abstract:Optical chemical structure recognition (OCSR) demands strict all-or-nothing graph fidelity, where a single misidentified bond order or formal charge makes the predicted graph differ from the reference structure. While vision-language models (VLMs) can reinspect ambiguous visual regions, unconstrained full-graph regeneration often introduces regressive errors into previously correct substructures. We present LVRE, a training-free visual agent that refines molecular graphs via targeted visual inspection and reversible transactional editing. Rather than regenerating the entire graph sequence, LVRE examines a high-resolution crop alongside the global diagram and local graph neighborhood, proposing modifications to named graph records that a deterministic controller checks. A controller executes each proposed modification as an isolated transaction, and a visual review module arbitrates which edits to commit or roll back. Across four foundation VLMs, LVRE improves graph accuracy by 8.0-8.5 percentage points over single-pass recognition and by 3.0-3.3 points over fixed local review. In a paired evaluation of 1,000 diagrams using Gemini 3.8 Flash, LVRE achieves 100 successful repairs while introducing only 15 regressive errors, in contrast to global reflection which yields 65 repairs against 30 introduced errors. On this paired set, LVRE introduces half as many errors on initially correct predictions as global reflection.

[CV-283] chnical Challenges for Ultra Fine Pitch Hybrid Bonding

链接: https://arxiv.org/abs/2611.18713
作者: Charles G. Woychik,Jonathan Abdilla,Chris Scanlan,Victor Vilar,Robert Patti,Dadi Setiadi,Himanandhan Reddy Kottur,Navid Asadizanjani
类目: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
备注: Symposium manuscript version of the paper published in the IMAPSource Proceedings 2026. The final published version is available through IMAPS. 10 pages, 14 figures

点击查看摘要

Abstract:Hybrid bonding is rapidly emerging as a foundational technology for next-generation advanced packaging, enabling ultra-high-density interconnects required for heterogeneous integration of chiplets, logic, memory, and specialized accelerators. As the industry pushes toward ultra-fine pitch interconnects approaching single- digit microns and below, significant technical challenges must be addressed across the process flow. These challenges include achieving extremely tight wafer and die alignment tolerances, maintaining surface planarity and cleanliness at the nanometer scale, controlling copper dishing and dielectric erosion during CMP, and ensuring defect-free oxide and copper-copper bonding interfaces. In addition, scaling hybrid bonding to ultra-fine pitch requires improvements in metrology, defect inspection, thermal management, and yield control to support high-volume manufacturing. Integration with complementary technologies such as silicon interposers, wafer-to-wafer and die-to-wafer bonding architectures, and heterogeneous material stacks further increases process complexity. This presentation reviews the key technical barriers to ultra-fine pitch hybrid bonding and discusses emerging process solutions and equipment innovations required to enable reliable, high-density interconnects for future high-performance computing, AI, and advanced system-in- package applications.

[CV-284] Whole-Slice Misses Escape Round-Trip Filtering and Volumetric Dice in a Source-Free Flow-Matching Segmenter

链接: https://arxiv.org/abs/2611.18710
作者: Guobiao Wang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 3 figures

点击查看摘要

Abstract:SymmAdapt, a source-free flow-matching segmentation pipeline, relies on two quality gates: a round-trip filter on pseudo-labels and volumetric Dice in evaluation. We audit both on 120 AMOS CT patients unseen in training and adaptation. The model often misses the spleen on whole slices, severely so in 30.0% of patients. The round-trip filter passes 83.31% of these lost slices, because its round trip reproduces the miss and the filter scores only organs present in the pseudo-label. Volumetric Dice depends on how much volume is missed, not where, and self-training leaves 96.5% of lost slices lost (96.6% on a second noise stream). A model-free repair, Chain-and-Fill, met all pre-registered endpoints on 119 further patients and could not be distinguished from SAM 2 propagation in severe-group gain, with 1/357 rather than 63/357 harmed patient-seed pairs. Reporting lost slices next to volumetric Dice makes the error visible at no cost.

[CV-285] opSDF: Topological Signed Distance Field for High-Quality Triangle Mesh Extraction from 3D Gaussian Splatting

链接: https://arxiv.org/abs/2611.18708
作者: Khoi Do,Yu-kun Lai,Binh-Son Hua
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated its effectiveness for photorealistic novel view synthesis. However, extracting clean, watertight, explicit surface meshes from 3DGS remains a critical challenge. Previous methods approximate geometry via discrete Marching Tetrahedra, where each grid vertex is assigned a single scalar signed distance value. At complex geometric boundaries and thin structures, this suboptimal parameterization forces severe compromises across shared vertices, causing gradient collisions and geometric frustration. Furthermore, existing optimization objectives neglect element shape conditioning, producing degenerate triangles. We introduce TopSDF, a framework that models the signed distance state as a multivariate Gaussian random field. By learning latent topological covariances across tetrahedral vertices, our formulation expands local degrees of freedom for SDF estimates, allowing the candidate mesh vertices to be placed flexibly along the bipolar edges in the tetrahedra as driven by such topological correlation. Concurrently, we directly integrate variational meshing energies into the joint optimization, coupled with an energy-driven Gaussian densification strategy to improve triangle mesh quality. Our experiments across Mip-NeRF 360, Tanks Temples, and ScanNet++ show that TopSDF achieves state-of-the-art mesh-based novel view rendering (+0.86 dB PSNR on Mip-NeRF 360) and triangle-mesh quality (from 0.37 to 0.49 aspect ratio).

[CV-286] Different Timescales Different Signals: Multimodal Landmark-Visibility Prediction in a Long-Term Webcam Network ACCV2026

链接: https://arxiv.org/abs/2611.18707
作者: Bryceton Bible,Hairong Qi,Masahiro Toyoura
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 18 pages, 5 figures, 4 tables. Accepted to the MM4SciReal Workshop at ACCV 2026

点击查看摘要

Abstract:Scenic visibility is not a single forecasting problem: planning a trip months ahead, deciding whether to go outside today, and judging whether tomorrow will be clear rely on different information. Using a year-scale multimodal archive of Mount Fuji webcams and meteorological observations, we identify a clear hierarchy across timescales. Seasonal climatology provides useful long-range guidance, webcam imagery is the strongest same-day observation (AUC 0.95), and the numerical forecast valid on the target day carries most of the day-specific signal one to three days ahead. Revealing this hierarchy requires evaluation that respects the dependence among cameras observing the same regional weather state: permissive splits substantially overstate multi-day performance, while date-grouped and forward-in-time evaluation reduce the apparent contribution of the current scene and current weather. On 6,575 camera-days, the current visual summary provides no measurable one-day gain once the target-day forecast is available. The remaining opportunity lies at shorter horizons, where the present visual representation discards cloud geometry and motion. These results recast scenic visibility prediction as a multi-timescale sensing problem rather than a single multimodal forecasting task.

[CV-287] Robustness as a Transferable Skill: Cross-Task Adversarial Defense for Vision-Language-Action Policies

链接: https://arxiv.org/abs/2611.18702
作者: Jinquan Zhang,Run Yang,Dongfu Yin,Qiang Lai,Yufeng Yan,Pengshuai Yin,F. Richard Yu
类目: Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Vision-Language-Action (VLA) policies are vulnerable to adversarial visual perturbations, while defenses often require task-specific adversarial fine-tuning. We investigate whether robustness learned on source manipulation tasks can protect independently specialized target policies without target-side adversarial training. We first show that even a single source defense update improves robustness on held-out tasks, establishing cross-task defense reuse independently of multi-source aggregation. To combine source updates, we introduce Source-Norm Transferable Robustness Vector (SN-TRV), which retains the direction of their mean while calibrating its magnitude using source-update norms. The resulting update is applied to the target visual encoder without target adversarial data, target-side gradient optimization, or additional inference modules. Evaluations across three VLA backbones, fixed and adaptive attacks, and different source defense recipes show improved robustness over naive averaging and identify limits in attack coverage. Physical PiPER trials increase average success under printed adversarial patches from 20.3% to 45.8%.

[CV-288] ARC-VLN: Adaptive Recursive Computing-based Visual-Language Navigation

链接: https://arxiv.org/abs/2611.18696
作者: Xiang Zhang,Xiangyu Dong,Yang Lu,Xiaoguang Ma
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Efficient aerial visual-language navigation (VLN) requires reducing model storage and inference costs while maintaining instruction-conditioned action prediction. We present ARC-VLN, a token-type-aware recursive computation framework built on a language backbone with shared parameters. ARC-VLN retains full-depth computation for text and action tokens while adaptively selecting exit depths for visual tokens. Representations of early-exiting visual tokens are retained, and active-token compaction shortens the sequences processed by subsequent recursive blocks. To mitigate routing collapse to a single exit depth, we design a dedicated learning strategy that encourages the model to distribute visual tokens across multiple exit this http URL OpenFly’s Test-Seen and Test-Unseen subsets, ARC-VLN achieves the lowest navigation error and the highest success rate and success weighted by path length among all compared methods. Meanwhile, it reduces the average number of decision steps per trajectory, total floating-point operations, mean model inference time per trajectory, and average GPU memory usage during inference, indicating its great potential as a VLN backbone for lightweight applications.

[CV-289] Beyond Raw Similarity: Prior-Aware Fusion for Out-of-Distribution Detection

链接: https://arxiv.org/abs/2611.18692
作者: Lakpa Tamang,Mohamed Reda Bouadjenek,Saman Forouzandeh,Kamal Berahmand,Mahdi Jalili,Sunil Aryal
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 9 Pages, 3 Figures, 5 Tables

点击查看摘要

Abstract:Out-of-distribution (OOD) detection in vision-language models presents unique challenges due to the reliance on cross-modal similarity. Even when embeddings are well separated for in-distribution (ID) data, similarity scores can remain overconfident for OOD inputs. A common remedy is to adapt the model on ID data, but adaptation discards the pretrained geometry that made the similarity meaningful. We propose multi-modal logit fusion (MM-Log), which keeps both by pairing a gradient-free zero-shot prior with a task-adapted branch in each modality and fuses their four cross-modal similarity channels into a single OOD score. Because the zero-shot priors are never adapted, the cross-channels act as a consistency check between the pretrained and adapted geometries: an OOD input that fools the adapted branch through a shortcut must additionally agree with frozen CLIP to retain a high score. We formalize this as a margin bound showing that the fused score is provably penalized by disagreement between the zero-shot and adapted branches, a signal to which adapted-only scores are structurally blind. MM-Log requires no auxiliary OOD data, negative prompts, or test-time adaptation, and trains on few-shot ID data alone. Across Far-OOD, Near-OOD, and Hard-OOD benchmarks, MM-Log consistently outperforms strong CLIP-based baselines.

[CV-290] Perceptual Weight Space Programming for Diffusion Super-Resolution

链接: https://arxiv.org/abs/2611.18690
作者: Chenzhi Zheng,Zhuoran Zheng
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Diffusion models for image super-resolution can recover scene content while producing distracting artifacts or unbalanced colors. We study whether a large language model (LLM) can improve perceptual quality by editing a small subset of existing U-Net weights without a clean reference image. The proposed framework provides the editor with architecture code, selected weights, activation summaries, and sampling previews. It permits up to 1,020 candidate trials while limiting accepted changes to three layers and fewer than 0.1% of model parameters. The evaluation protocol uses ResShift, SinSR, and SUPIR and compares GPT-6 Astra with a specified Qwen3-VL model under matched access. Feedback metrics guide editing, while separate quality metrics and blinded human comparisons evaluate the final weights. We also report an executed color calibration experiment using real photographs and two small residual networks. This experiment demonstrates correction of a fixed color shift and reveals that an improving feedback score can accompany worse performance on unseen scenes. A separate lightweight super-resolution experiment isolates modest changes in luminance detail while preserving color. The findings motivate restricted updates and independent evaluation of perceptual weight editing.

[CV-291] PACTHub: a large-scale structured experimental dataset for three-dimensional photoacoustic computed tomography

链接: https://arxiv.org/abs/2611.18689
作者: Zhibo Xiao,Xuanhao Wang,Hao Huang,Fan Meng,Xuran Zhou,Xiali Gao,Yang Xiao,Junhui Shi
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Three-dimensional photoacoustic computed tomography (PACT) provides volumetric contrast related to optical absorption for studying tissue anatomy and dynamics. Method development requires experimental datasets that retain measured signals and acquisition context. Public resources remain limited in their combined coverage of varied specimens, volumetric acquisition modes and reusable reconstruction information. Here we present PACTHub, a large-scale structured experimental dataset comprising 1,464 acquisition groups and 99,810 photoacoustic exposures collected over nearly four years. Animal, phantom and human measurements span single-position imaging, rotational compounding, panoramic scanning and time series. A common PAdata/JSON representation links acoustic signals to detector geometry, exposure selection, signal preparation and physical reconstruction coordinates. Public reconstruction implementations, numerical volumes and maximum-intensity projections provide practical entry points for reusing the measurements. Spatial and temporal examples demonstrate how retained acquisition relationships support panoramic reconstruction and regional signal analysis. Detector and exposure subsets enable studies of angular sampling and measurement budget. Volumetric evaluation combines PSNR, SSIM and L1 with Q, a complementary measure of local-contrast agreement within reference-defined structures. By connecting measured signals to reconstruction settings and numerical outputs, PACTHub makes its experimental scale and diversity available for reuse. It provides an experimental foundation for developing and evaluating artificial intelligence (AI) methods in volumetric PACT.

[CV-292] 3D Cell Morphology–Clinical Text Contrastive Learning with Voxel Swin Transformer and Finding Encoder Domain-Adaptation

链接: https://arxiv.org/abs/2611.18686
作者: Takato Yasuno
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 16 pages, 8 figures, 6 tables

点击查看摘要

Abstract:Our prior work adapted a 2D CLIP dual encoder with LoRA to a multi-source cell-microscopy corpus, reaching anomaly-scoring ROC-AUC up to 0.985. This paper advances the same contrastive-learning domain-adaptation methodology from 2D microscope images to 3D microscope volumes, pairing a 3D Swin Transformer voxel encoder (MONAI’s SwinUNETR backbone) with ClinicalBERT as the finding-text encoder. We train on 3,000 single-cell 3D confocal fluorescence volumes (4 tagged organelle structures) from the Allen Institute’s public hiPSC single-cell image dataset, paired with template-generated finding text, with mitotic (dividing) cells oversampled as an abnormal-like proxy class validated by ROC-AUC. Naively reusing work’s freeze backbone, LoRA the attention projections recipe for the Voxel Encoder collapses voxel embeddings to near-input-invariance, since LoRA’s low-rank delta cannot compensate for an otherwise-frozen, randomly initialized (not pretrained) 3D Swin backbone, even as InfoNCE loss decreases healthily. Switching the Voxel Encoder to full fine-tuning resolves this: 4-way zero-shot structure accuracy rises from 25.0% (chance) to 48.6%, and voxel anomaly-scoring ROC-AUC improves from 0.486 to a peak of 0.628 (still below the 0.70 sanity threshold), while bidirectional retrieval recall stays low (R@10 \le 3.8%). We further introduce a nearest-neighbor diagnosis-reason retrieval evaluation, reaching 29.8% hit-rate@1,5 at its best scale while scoring 0.0% elsewhere. The central finding of this MVP is architectural: LoRA’s parameter-efficiency assumption presumes a pretrained backbone worth preserving, and breaks down when the backbone is randomly initialized.

[CV-293] hinkWise: Learning When to Think in Large Multimodal Reasoning Models via Visual and Difficulty Perception

链接: https://arxiv.org/abs/2611.18685
作者: Wenkang Zhou,Chen Yang,Jiaheng Peng,Xitong Li,Hongzhi Yin,Yunke Zhang,Lu Yu,Peng-Fei Zhang,Xuhui Ren,Shijie Zhang
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Large Multimodal Reasoning Models (LMRMs) extend large language models to heterogeneous modalities, demonstrating strong potential across a wide range of open-ended tasks. Reinforcement Learning with Verifiable Rewards (RLVR) has been widely adopted to improve reasoning performance in LMRMs; however, existing approaches overlook the critical role of visual perception uncertainty in guiding reasoning processes. We term this limitation Reasoning Allocation Misalignment, where the misalignment between visual perception uncertainty and reasoning depth leads to over-allocation of deep reasoning under uncertain visual grounding, resulting in unnecessarily long reasoning trajectories and degraded performance. To address this issue, we propose ThinkWise, a perception-aware RLVR framework that jointly models task difficulty and visual perception state to regulate reasoning depth. Specifically, we introduce an instance-adaptive visual perception modeling mechanism that enables context-aware calibration of perception uncertainty, leading to more accurate perception estimation. We further design a structured reward formulation that jointly considers perception, correctness, and task difficulty. In particular, an answer reward provides supervision for correctness, a perception- and difficulty-conditioned length reward adaptively regulates reasoning depth, and a perception reward encourages trajectories with stronger visual grounding. Extensive experiments on multiple multimodal reasoning benchmarks demonstrate that ThinkWise consistently improves the trade-off between accuracy and reasoning efficiency across both 3B and 7B models. These results highlight the importance of perception-aware reasoning and establish a principled framework for advancing LMRMs.

[CV-294] PolarGait: Constructing Complementary Views from a Single Human Sketch for Gait Recognition

链接: https://arxiv.org/abs/2611.18682
作者: Chao Zhang,Zhuang Zheng,Zhanyong Mei,Huayan Pu,Jun Luo,Jinglei Zhao
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Existing multimodal gait recognition methods exploit complementary information by fusing heterogeneous representations. Whether useful complementarity can be constructed from a single representation before fusion remains less explored. We propose PolarGait, which constructs complementary views of Human Sketch through filtering and spatial reparameterization. Filter-Oriented Sketch Decomposition (FOSD) uses Gaussian smoothing to extract coherent structures and absolute residuals to capture local response differences. Gait-Oriented Polar Transformation (GOPT) combines silhouette-guided sketch normalization, a gait-oriented center, and lower-body-biased non-uniform angular sampling to construct a polar view complementary to the Cartesian representation. Silhouette provides auxiliary global body-shape information. Experiments on CCPG, SUSTech1K, CASIA-B, and Gait3D validate the framework. Under the single-Sketch setting, FOSD+GOPT improves the original Sketch baseline by 6.6% on CCPG and 4.6% on CASIA-B. The complete PolarGait achieves Rank-1 accuracies of 87.9%, 95.2%, 96.6%, and 86.4% on the four benchmarks, respectively. These results show that complementary gait information can be constructed within a single structured representation and integrated for robust gait recognition.

[CV-295] Spread Too Thin: How Higher Resolution Dilutes a Vision Transformers Attention and a Training-Free Rescaling

链接: https://arxiv.org/abs/2611.18680
作者: Neil D. B. Bruce
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 10 pages, 3 figures, 11 additional appendix pages

点击查看摘要

Abstract:A vision transformer read at a higher resolution than it was trained at spreads each attention row over many more patches, yet every row must still hand out exactly one unit of weight. A law of softmax attention predicts that a growing crowd of bounded scores dilutes the attention on any patch or register whose score stays fixed. The same law gives a remedy that needs no training: multiply every score, the sink’s included, by the ratio of the logarithms of the test and training token counts. Given sufficient margins at the training resolution, this keeps the attention on a match when one is present, and on a sink such as a register when none is, though trained models need not meet those margins. We test these predictions in public vision transformers and in controlled models read far above their training token count. Attention spreads as predicted. Registers lose their share, though production models compensate in part, and the rescaling holds the share of DINOv2’s registers nearly level. Applied at test time without training, the rescaling recovers part of the lost accuracy in nearly every model, with absolute or rotary positions and with registers, and on images with native detail. The calibrated multiplier rises with the logarithmic ratio on average and varies across models. Its rise is steeper than the square-root rule used in image generators; in controlled models the log ratio is also more accurate, while in the public models the two rules’ accuracy difference remains unresolved. The rescaling does not restore the attention a model learned, and why it helps, beyond countering the dilution, remains open.

[CV-296] Shared Mobile AR Without Depth Sensors from Monocular Point Maps

链接: https://arxiv.org/abs/2611.18679
作者: Mohammad Rashid
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 15 pages, 8 figures, 4 tables

点击查看摘要

Abstract:Depth-aware augmented reality on phones needs dense geometry of the surrounding room. iPhones and iPads without LiDAR expose only sparse feature points and detected planes, and ARKit often takes tens of seconds to report the first wall. We describe a scan-free pipeline for these devices. A point-map network (MoGe-2) predicts metric geometry from each camera frame, the remaining depth shift is recovered from the camera intrinsics, and each frame’s scale is aligned to views already in the map. A persistent voxel map with visibility-based expiry answers occlusion and collision queries. Co-located devices share one coordinate frame by solving yaw and floor-plane translation from sightings of each other. We evaluate the deployed system on 309 device-sessions from 19 device models without LiDAR, with 179 device-sessions from LiDAR phones in the same deployment as a reference population. A first dense depth map was recorded within 10 s in 80% of sessions, while ARKit had reported a wall plane by then in 25%. The median session scale was 0.98 against ARKit’s wall planes and 1.00 against its camera translation. Scale changed by a median of 9% between nearby frames, against 0.4% with LiDAR, and depth points spread three times as widely around walls. In multi-device sessions, 69% of devices without LiDAR reached a converged shared frame, against 84% of LiDAR phones. The main remaining error is scale that changes from frame to frame.

[CV-297] Not All Steps Matter: Temporally Adaptive Spectral Control for Concept Removal in Diffusion Models

链接: https://arxiv.org/abs/2611.18678
作者: Chanaka Thushitha Bandara,Lakmali Nadeesha Kumari,Sen-Ching Samson Cheung
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:As text-to-image diffusion models see widespread deployment, removing unsafe or undesirable concepts without retraining has become a pressing need. Spectral methods are attractive because they are training-free and reversible, yet current variants apply the same output-value (OV) circuit intervention at every denoising step, assuming concept information is temporally uniform. This wastes intervention where concept energy is low while degrading image quality, an inefficiency that compounds across concept types. Through timestep-resolved spectral analysis, concept-specific energy is found to follow a consistent temporal hierarchy: content concepts decay fastest, object concepts are intermediate, and style concepts persist longest. This paper proposes Timestep-Adaptive Spectral Nullification (TASN), which adapts suppression strength and spectral components to the concept energy profile. Cross-attention at each timestep is decomposed into concept-aligned and concept-orthogonal components, and it is proved that when the orthogonal component is approximately stable, the collateral-minimizing schedule sets suppression proportional to concept energy. TASN implements this with a timestep-adaptive spectral subset tracking only active concept directions, strictly generalizing uniform Spectral Nullification (SN). On five red-teaming benchmarks, TASN lowers attack success rate below uniform SN while recovering the majority of its quality loss; ablations confirm gains arise from temporal targeting rather than reduced suppression, generalizing across concept categories and architectures.

[CV-298] SynPanZ: A Synthetic Dataset to Improve Multi-Animal Pose Estimation in Chimpanzees (Pan troglodytes)

链接: https://arxiv.org/abs/2611.18674
作者: Flora Talyigas,Lindsay Murray,Sonya Hill,Richard Stocker
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Pose estimation underpins a wide range of automated monitoring workflows in animal behaviour research. Unfortunately, endangered great apes such as chimpanzees suffer a chronic shortage of the type of hand-annotated training data that has made robust pose estimation in humans and other model organisms possible. This paper introduces SynPanZ, a synthetically generated dataset for pose estimation in great apes. Our pipeline renders multi-individual scenes from rigged, behaviourally animated chimpanzee models with randomised poses, lighting conditions, environmental contexts, and camera motion, producing animated RGB videos with per-instance 2D keypoint locations without manual labelling. To test whether synthetic augmentation improves cross-dataset generalisation, four DeepLabCut models are trained under matched settings on data sourced from DeepWild with varying real-to-synthetic augmentation ratios. Models are evaluated on ChimpACT, a visually disjoint zoo dataset. We find that augmenting the DeepWild training data with approximately three times as many synthetic frames nearly doubles instance coverage (12.34% to 22.54%) while maintaining keypoint localisation accuracy on detected individuals (PCK@0.20: 49.71% to 52.25%). Synthetic-only training fails, showing that rendered data complements rather than replaces real annotated data. This work represents a step towards reliable pose estimation for great apes.

[CV-299] Robust-GCD

链接: https://arxiv.org/abs/2611.18673
作者: Daniel Wang
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Generalized Category Discovery (GCD) classifies an unlabeled pool that contains both known and novel classes, given labels for the known classes only. Existing GCD methods assume that every un- labeled image is clean, but unlabeled data collected in the wild is often degraded by blur, sensor noise and compression artifacts. We formalize the setting in which an unknown fraction of the unlabeled pool is corrupted by an unknown operator as Degraded Generalized Category Discovery (D-GCD). To tackle it, we propose Robust-GCD, a framework that estimates the corruption directly from the unlabeled pool and uses the estimate to build labeled anchors for known classes and pseudo-anchors for novel classes inside the corrupted region of feature space. A trust curriculum controls how much the unlabeled pool is trusted while the anchors take hold, and a paired alignment loss keeps a clean image and its corrupted twin close in feature space. Across five datasets and multiple corruption families, Robust-GCD recovers between 43% and 96% of the All accuracy that SimGCD loses under contamination, and more than 60% on four of the five datasets, with the largest gains on the corrupted images themselves.

[CV-300] Exists Valid: Calibrated Validity of Stored 3D Geometry Across Long Observation Gaps

链接: https://arxiv.org/abs/2611.18672
作者: Anwaar Ulhaq
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 3 Pages, 1 Figure

点击查看摘要

Abstract:Repeated airborne LiDAR surveys are costly, while stored 3D geometry can become outdated between observations. Existing change-detection methods compare observed epochs. However, they do not quantify the validity of geometry that has not been revisited. We distinguish observation, existence and geometric validity using a probabilistic filter over five geometric transitions: persistence, modification, growth, removal and replacement. The framework learns class-specific continuous-time transition priors from interval-censored observations and combines shape residuals with LiDAR free-space evidence. Functional anchors enable prior sharing across sites with different survey frequencies. In controlled experiments with misspecified dynamics and 60% visibility, our method reduces combined error from 36.9% to 16.5% and expected calibration error (ECE) from 0.19-0.32 to 0.017. Cross-site sharing reduces ECE at two-survey sites from 0.156 to 0.029. On Urb3DCD-v2, the method achieves 9.8% combined error and 0.016 ECE, outperforming baselines at 12.6% and at least 0.070, respectively. Re-survey prioritisation improves only modestly over random selection with two observations. Code and an evaluation protocol are released at this https URL.

[CV-301] hinking from Another Side of the Box: Spacetime Images for Computer Vision

链接: https://arxiv.org/abs/2611.18671
作者: Jade Lambert,Sébastien Piérard,Anthony Cioppa,Marc Van Droogenbroeck
类目: Computer Vision and Pattern Recognition (cs.CV)
备注: 18 pages, 6 figures, 1 table, 42 references

点击查看摘要

Abstract:Computer vision methods for video understanding, such as object detection and tracking, inherit the way videos are captured and displayed: as successive frames processed sequentially. However, a video is an inhomogeneous 2D+time volume of pixels, named voxels, and frames are only one family of cuts through it. In this frame-based representation, static backgrounds introduce computational redundancy, and motion can only be understood across several frames. In this position paper, we study another family of cuts: spacetime images (STIs), obtained by sampling any set of pixels other than frames, for instance a line of pixels followed over time. Such images are familiar in narrow forms, from line-scan inspection and kymographs to spatiotemporal slices for motion analysis, but have received little attention as a general input representation for computer vision and are often limited to (x,t) or (y,t) cuts. We define new cuts leading to STIs in their general form, explain how to read their content, and show that a family of cuts tiles the video so that existing dense annotations transfer at no cost. We also detail how motion appears in an STI, spatially and in the frequency domain, and discuss applications such as walking pedestrians, deposited objects, and dynamic scenes. Two pilot experiments illustrate the representation. The first trains a segmentation network entirely in the spacetime domain, on transferred annotations, and performs background subtraction on CDnet 2014 close to supervised frame-based methods, without explicit temporal modeling. The second recovers the pan speed of a camera as a moving cut, identified by a simple criterion on the frequency spectrum of the resulting image, turning a temporal estimation problem into the choice of a cut. Our position is not that STIs should replace frames, but that slicing the video volume is a design variable worth studying in its own right.

[CV-302] Evidence-Calibrated Selective Decoding for Reliable Visual Question Answering

链接: https://arxiv.org/abs/2611.18670
作者: Shiyu Gu,Luis Miguel Santana,Rafael Enrique Cruz
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Visual question answering systems can produce confident responses that are weakly supported by their input images. Contrastive decoding can reduce this dependence on language priors, but a fixed correction may also overturn a correct answer, while selective answering can conceal errors by reducing coverage. We develop evidence-calibrated selective decoding, a framework that separates three decisions: whether local visual evidence supports a candidate, how strongly to correct the original distribution, and whether the resulting answer should be released under a verification budget. Evidence Verification across Crops (EVC) measures visual support and disagreement; Adaptive Contrast Correction (ACC) applies a bounded modification within the backbone’s plausible candidate set; Budgeted Selective Answering (BSA) calibrates an answer-admission threshold on held-out images. We provide a complete mathematical specification, bounded-perturbation analysis, and an evaluation protocol that distinguishes forced answering from selective risk. Our quantitative analysis re-examines published baselines on POPE instantiated over MSCOCO, A-OKVQA, and GQA, together with four MME subtasks. These public measurements reveal variation across backbones, tasks, and correction settings that motivates adaptive control. The proposed framework has not yet undergone model-level evaluation; its effectiveness remains a falsifiable hypothesis rather than an established empirical result.

[CV-303] LogicEye: Agent Driven Multimodal Dataset for Interpretable Ophthalmic Diagnosis

链接: https://arxiv.org/abs/2611.18669
作者: Jiajie Mu,Lu Chen,Zheyun Qin,Xiaoming Xi,Lishan Qiao,Xiushan Nie,Yilong Yin
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Clinical adoption of ophthalmic Artificial Intelligence remains limited by low interpretability and a lack of finely annotated multimodal data. Most existing datasets provide only coarse disease labels without the granular pathological findings required for differential diagnosis. We present LogicEye ^Agent , a dual-pool inference and verification agent that emulates clinical reasoning by jointly verifying positive inclusion signs and negative exclusion criteria. Using LogicEye ^Agent , we curate LogicEye ^Data , a large-scale multimodal dataset of 804,461 fundus and OCT images paired with structured pathological reports spanning 38 diseases and 29 subtypes, with quality assured through joint verification by multimodal large language models and clinical experts. Building on this dataset, we further develop LogicEye ^Bench , which incorporates a category-constrained loss for hierarchical disease recognition and report generation. Grounding coarse labels into verified visual findings effectively curtails diagnostic hallucinations and establishes a solid foundation for interpretable vision-language modeling in ophthalmology. The code is available at:this https URL.

[CV-304] UMM-JEPA: Generative Pretraining via Masked Next-Scale Representation Prediction for Unified Multimodal Models

链接: https://arxiv.org/abs/2611.18668
作者: Yucheng Zhou,Yunjia Xi,Di Yin,Jianbing Shen
类目: Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Unified multimodal models can support both visual understanding and generation, yet generative training does not necessarily improve understanding. We argue that this limitation may arise from a mismatch between the representation targeted by generation and the perceptual representation used for understanding. Conventional generation objectives primarily optimize synthesis-oriented spaces such as discrete visual codes or generative latents. We therefore introduce UMM-JEPA, which reformulates generation as progressive prediction of continuous visual-understanding representations. The model constructs a multi-scale representation hierarchy from coarse to fine, while pixel rendering is decoupled from representation pretraining and attached only afterward. Controlled comparisons show that the representation-prediction objective accounts for most of the understanding improvement, while a data-matched control leaves the group means nearly unchanged. UMM-JEPA reaches a 16-benchmark visual mean of 51.71, against 46.36 for the understanding-only Qwen-VL baseline, and the text mean also rises. Compared with discrete-code and latent-flow generation objectives, perceptual representation prediction yields stronger and more consistent transfer to understanding. Linear probes and visual-evidence interventions further reveal richer object and spatial structure and stronger use of image-specific information. These results suggest that generation can benefit understanding when it directly supervises the perceptual representations that understanding consumes.

[CV-305] ReMoVe: Refining a Video Models Own Motion for Coherent Human Video Generation

链接: https://arxiv.org/abs/2611.18666
作者: Jie Li,Muxin Pu,Hongyi Cai,Mingkang Dong,Jianyuan Guo,Tao Huang
类目: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Graphics (cs.GR)
备注:

点击查看摘要

Abstract:Video generation models have become increasingly capable of synthesizing realistic human motion, yet they often lose track of the very structure that makes the motion plausible. As generated human motion becomes more complex, videos can suffer from progressive body deformation and appearance drift across frames. Existing motion-controlled methods use explicit representations to stabilize generation, but typically rely on pre-defined motion sequences as references. This raises a fundamental question: if the base generator already produces a usable coarse motion, why replace it with an external one? ReMoVe answers this question by treating the coarse output as a motion hypothesis rather than a final result. ReMoVe recovers and refines the underlying motion, then renders the refined motion as 3D mesh and transfers facial appearance onto it, thereby improving motion accuracy and identity consistency. We further derive depth from the same 3D scene and camera trajectory to impose scene-level spatial consistency beyond the mesh. By jointly conditioning video generation models on mesh and depth, ReMoVe synthesizes high-quality motion videos with controllable camera trajectories and enhanced spatio-temporal consistency. Extensive experiments demonstrate that ReMoVe improves human structural stability, appearance consistency, temporal coherence, and motion-following fidelity, while retaining the base generator’s motion prior and offering accurate camera control.

[CV-306] ECOD: Experience-Calibrated On-Policy Distillation for Multi-Turn Agents

链接: https://arxiv.org/abs/2611.18656
作者: Jie Jin,Shengqin Wang,Yan Gao,Xin Tan,Yuan Xie
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Sparse and delayed rewards make reinforcement learning challenging for long-horizon language agents. On-policy distillation (OPD) complements environmental rewards with dense teacher feedback on student-generated contexts. Standard OPD can misalign supervision with relative teacher competence, student trajectory outcomes, and the temporal structure of task execution. Recent variants adjust guidance through divergence-based weighting and temporal schedules, while joint calibration from task-completion experience remains underexplored. We introduce Experience-Calibrated On-Policy Distillation (ECOD), which allocates teacher supervision at the task, trajectory, and turn levels. ECOD compares teacher and student success-rate posteriors under matched task conditions, restricts direct supervision to failed student trajectories, and derives a temporal prior from successful teacher lengths. The resulting weight applies to teacher losses or teacher-derived advantages. We evaluate ECOD on ALFWorld and WebShop with Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct. Under matched teacher settings, ECOD improves standard OPD by up to 4.06% in relative ALFWorld unseen success. Across both model scales, ECOD-assisted SOD, TCOD-F2B, and ATOD configurations yield average paired relative gains of 6.20% on ALFWorld unseen and 7.05% on WebShop, with WebShop gains ranging from 2.29% to 21.60%. These results support experience-calibrated supervision allocation as a reusable mechanism across distinct distillation objectives.

[CV-307] HIMX: mixed-format matrix multiplication based on customized HIFLOAT4 representations

链接: https://arxiv.org/abs/2611.18641
作者: Zhiyuan Zhang,Yanzhao Li,Hui Dong,Sifan Cheng,Jingwei Zhang
类目: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:4-bit weight-activation quantization (W4A4) is an important approach for efficient large language model (LLM) inference. Recent hierarchical formats such as HiFloat4 (HIF4) improve the representational range of 4-bit quantization through multi-level scaling. However, the same local mantissa representation in HIF4 may not be equally suitable for weights and activations, which exhibit distinct distribution characteristics. To address this limitation, we propose HiMX, a mixed-mantissa framework built upon the hierarchical scaling structure of HIF4. Our key observation is that the two HIF4-compatible mantissa representations, S1P2 and E2M1, exhibit complementary quantization behavior: the denser S1P2 format provides higher resolution for concentrated weight distributions, while the wider-spaced E2M1 format provides complementary quantization characteristics for activation tensors. HiMX enables operand-tailored mantissa representations while converting heterogeneous mantissas into fixed-point integer operands with operand-specific widths, allowing different operand formats to share a unified integer multiply-accumulate (MAC) array. Block-level analysis on Qwen3-8B shows that HIF4-S1P2 is better suited for weights while HIF4 E2M1 is better suited for activations, and their mixed-format combination achieves lower output MSE than homogeneous configurations. Across multiple benchmarks, HiMX reduces the average end-to-end accuracy loss from 2.38 to 1.77 points on Qwen3-8B and from 2.23 to 1.85 points on Qwen3-30B-A3B, demonstrating the potential benefit of heterogeneous mantissa representations for W4A4 LLM inference.

[CV-308] A Survey of Image-to-Code Generation for Software Engineering

链接: https://arxiv.org/abs/2611.18575
作者: Licheng Zhang,Xihong Wu,Dingsheng Luo,Siew-Kei Lam
类目: oftware Engineering (cs.SE); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:A large fraction of software intent is expressed visually before it is implemented. A designer hands a screenshot or a Figma mockup to a front-end engineer, an architect sketches a UML diagram or a flowchart, and a bug report ships with a picture of what went wrong. Turning such visual artifacts directly into executable software has therefore become an active problem at the intersection of computer vision and software engineering. The present survey organizes the literature on image-to-code generation for software engineering, defined as the family of methods whose input contains a visual modality and whose output is software an engineer would otherwise write by hand. We deliberately scope the survey to software-engineering code, namely user-interface and front-end code, code recovered from software diagrams such as UML and flowcharts, program repair and test code driven by visual signals, and general software-code targets, and we exclude adjacent code-generation domains whose home communities lie elsewhere, such as chart plotting scripts, vector-graphics programs, hardware description languages, and robot policies. Rather than cataloguing systems in isolation, we introduce a unified processing pipeline of visual understanding, component recognition, layout reasoning, code planning, code generation, and verification, and organize the survey around its stages. We trace five overlapping eras from rule-based reverse engineering to agentic generation, survey each pipeline stage and each software-engineering task in depth grounded in what individual systems actually do, consolidate the datasets, benchmarks, and metrics, and lay out the open challenges of design-system generalization, long-horizon and interactive generation, cross-framework deployment, accessibility, hallucinated components, and unreliable evaluation.

[CV-309] GOOSELoc: A Multimodal Benchmark for Ground-to-Aerial Localization in Unstructured Off-Road Terrain

链接: https://arxiv.org/abs/2611.16162
作者: Fabian Deuser,Peter Mortimer,Robin-Nico Kampa,Anna Bößendörfer,Mirko Mählisch,Martin Werner,Norbert Oswald
类目: Computer Vision and Pattern Recognition (cs.CV); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Cross-view pose estimation estimates a ground camera’s geographic pose by matching onboard sensor data with geo-referenced overhead imagery. This is particularly relevant when GNSS is unreliable, for example under forest canopy or in urban canyons. Existing metric cross-view benchmarks, such as VIGOR and cross-view KITTI, are dominated by urban and road scenes with strong human-made structure without off-road terrain. We introduce GOOSELoc, a multimodal benchmark for ground-to-aerial localization in unstructured environments, including forests, fields, and rural tracks. GOOSELoc comprises 9,356 scenes recorded across all four seasons around Munich, Germany, pairing geo-referenced aerial orthophotos with surround-view RGB imagery and 3D LiDAR point clouds. Paved/unpaved labels enable analyzing the impact of terrain structure, while the seasonal recordings allow studying the appearance gap to mostly summer-captured orthophotos. To assess the transferability of existing methods, we adapt and fine-tune four state-of-the-art cross-view localization baselines on GOOSELoc. Despite task-specific fine-tuning, the performance degrades substantially in off-road environments. We therefore propose a multimodal baseline combining surround-view imagery and LiDAR geometry, improving localization performance and providing a stronger reference for future work. Data and baselines are released at this https URL.

[CV-310] RoboACE: Adversarial Co-Evolution for Robot Self-Improvement

链接: https://arxiv.org/abs/2611.13034
作者: Delin Zhao,Zhengrong Yue,Junlin He,Yali Wang
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: 13 pages, 5 figures

点击查看摘要

Abstract:Robot agents can improve reusable skills, planning procedures, and recovery strategies from execution feedback, yet a fixed practice source leaves the search for new weaknesses unchanged. Useful practice depends on the agent’s current capabilities, so the procedure that chooses practice must also be revisable. We present RoboACE (Robot Adversarial Co-Evolution), a framework that jointly evolves curriculum and execution harnesses around frozen foundation models. The curriculum harness uses accumulated capability evidence to propose informative challenges, while the execution harness responds with reusable revisions that expand execution capabilities. Alternating updates hold one side fixed as the other changes, and admitted challenges form fixed snapshots for comparing execution revisions. Starting from a compact set of LIBERO practice conditions, the resulting frozen deployment outperforms the evaluated controllers and agent systems across four LIBERO-family panels. On LIBERO-PRO, it reaches 74.0% success, exceeding the initial deployment by 15.0 percentage points and direct execution by its VLA controller by 13.5 points. Case studies document changes to evidence-guided practice selection, drawer-skill grounding, and controller intervention.

[CV-311] Parkinsons Disease Freezing of Gait (FoG) Symptom Detection Using Machine Learning from Wearable Sensor Data

链接: https://arxiv.org/abs/2506.12561
作者: Mahmudul Hasan
类目: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Signal Processing (eess.SP)
备注:

点击查看摘要

Abstract:Freezing of gait (FoG) is a special symptom found in patients with Parkinson’s disease (PD). Patients who have FoG abruptly lose the capacity to walk as they normally would. Accelerometers worn by patients can record movement data during these episodes, and machine learning algorithms can be useful to categorize this information. Thus, the combination may be able to identify FoG in real time. In order to identify FoG events in accelerometer data, we introduce the Transformer Encoder-Bi-LSTM fusion model in this paper. The model’s capability to differentiate between FoG episodes and normal movement was used to evaluate its performance, and on the Kaggle Parkinson’s Freezing of Gait dataset, the proposed Transformer Encoder-Bi-LSTM fusion model produced 92.6% accuracy, 80.9% F1 score, and 52.06% in terms of mean average precision. The findings highlight how Deep Learning-based approaches may progress the field of FoG identification and help PD patients receive better treatments and management plans.

[CV-312] How Much Does Saliency Accuracy Matter for Saliency-Guided 360-degree Image Compression?

链接: https://arxiv.org/abs/2611.21532
作者: Sakib Ahmed,Sidi-Mohammed Senouci,Philippe Brunet,Guillaume Caron,Nevrez Imamoglu
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
备注: 8 pages, 7 figures, 1 table. Accepted at the 2027 IEEE/SICE International Symposium on System Integration (SII 2027)

点击查看摘要

Abstract:Saliency-guided coding allocates bits to the regions viewers are most likely to attend, but accurate 360-degree saliency prediction is computationally expensive, an obstacle for low-latency immersive streaming applications such as robot teleoperation. We quantify how much saliency-prediction accuracy actually translates into compression gain. Building on a saliency-guided rate-distortion-optimized (RDO) learned codec, modified so that the training-time latent mask is driven by ground-truth saliency, which yields a single, predictor-agnostic codec, we evaluate saliency predictors spanning two orders of magnitude in compute on both the learned codec and H.265/HEVC. The results show that higher saliency-prediction accuracy yields diminishing compression gains: a lightweight predictor fast enough for live use (52 frames per second) captures 65% of the ground-truth oracle’s bit-rate saving, a 3.3x heavier one reaches 74%, and a further 38x increase in compute adds no measurable benefit, on either codec. Therefore, lightweight predictors might also be a practical choice for latency-constrained 360-degree streaming.

[CV-313] Rotation-Invariant Machine Learning for Highly Symmetric Molecular Configurations NEURIPS2026

链接: https://arxiv.org/abs/2611.21308
作者: Logan Bolton,Tomas Karella,Nicholas Lubbers,Nathaniel Philip Gorski,Emily Shinkle,Alice Elisabeth Anastasia Allen,Pieter J Swart,Roxana Bujack
类目: Chemical Physics (physics.chem-ph); Computer Vision and Pattern Recognition (cs.CV)
备注: NeurReps 2026 Extended Abstract Track (NeurIPS 2026 Workshop on Symmetry and Geometry in Neural Representations)

点击查看摘要

Abstract:Atomistic machine learning can approximate expensive quantum mechanical calculations at a fraction of their computational cost, but accuracy depends strongly on how the three-dimensional atomic environments are represented. Invariant representations enable fast computation but can fail to distinguish geometrically distinct configurations, especially in highly symmetric environments. We introduce a lightweight benchmark to evaluate expressivity on n -fold symmetric configurations, and use it to expose degeneracies in both invariant and equivariant architectures. Guided by these results, we construct a new invariant basis for HIP-HOP which resolves most of the identified failures. On a methane simulation dataset, the extended basis achieves the lowest error among the HIP-HOP variants and has similar accuracy to EquiformerV3 while running 3.2\times faster at inference. These results show that carefully constructed invariant representations can mitigate key expressivity limitations while balancing accuracy and computational efficiency.

[CV-314] Investigating the Benefit of Transfer Learning from the Locus Coeruleus for Substantia Nigra Pars Compacta Segmentation

链接: https://arxiv.org/abs/2611.21229
作者: Tayebeh Bahador,Max Dünnwald,Matthew J. Betts,Emrah Düzel,Steffen Oeltze-Jafra
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
备注: 5 pages, 3 figures, 1 table. Submitted to IEEE ISBI 2027

点击查看摘要

Abstract:The substantia nigra pars compacta (SNpc) is a small midbrain nucleus that degenerates early in Parkinson’s disease (PD). T1-weighted FLASH MRI enables its visualization and may support early detection of neurodegeneration. However, manual segmentation is difficult and time-consuming, and deep learning models often suffer from the limited availability of labeled data. Transfer learning (TL) may help by adapting pre-trained models to small datasets. We investigate whether TL from a model trained on the locus coeruleus (LC), a nucleus with a similar appearance in this sequence, improves SNpc segmentation. Several TL strategies were evaluated using a 3D U-Net. With as few as 5-14 labeled subjects, TL and training from scratch achieved similar performance, reaching a Dice similarity coefficient of up to 0.77, close to reported inter-rater agreement. No TL strategy differed significantly from training from scratch, suggesting that initialization from the LC offers no significant benefit for SNpc segmentation.

[CV-315] Scribble Once Segment the Sequence: Bidirectional Temporal Consistency for Weakly Supervised Segmentation in Echocardiography Sequences

链接: https://arxiv.org/abs/2611.21137
作者: Vincent Li Weixin,Chen Jingkun,Vicente Grau
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:Temporal medical image segmentation relies on pixel-wise annotations, which are prohibitively expensive for long image sequences. Existing weakly-supervised methods reduce the annotation burden per frame but still require repeated annotations across the sequence. Furthermore, sparse supervision can result in temporally unstable representations. Addressing both challenges requires a model that can leverage strong spatial priors from large-scale pretraining while minimizing reliance on dense annotations. Frozen 2D vision foundation models provide a promising basis for this by offering transferable spatial representations with low trainable parameters. However, their lack of explicit temporal modeling limits their ability to propagate reliable information across anatomical sequences. To address this, we present a weakly-supervised framework for temporal segmentation from a single scribble-annotated frame, with all remaining frames left unlabeled. The framework combines parameter-efficient adaptation of DINOv3 encoders present in the mean-teacher architecture. Furthermore, a set of complementary losses, including a temporal loss helps generate reliable segmentation priors for the unlabeled frames. This formulation decouples annotation costs from sequence length while retaining the transferable spatial knowledge of the pretrained encoder. Extensive experiments on the publicly available CAMUS and CardiacNet datasets demonstrate that our approach outperforms other scribble-supervised methods. These results show that accurate and temporally consistent segmentation can be achieved from extremely sparse supervision, substantially reducing annotation requirements for temporal echocardiographic segmentation and providing a practical framework for scalable clinical deployment.

[CV-316] Beyond AUC: Do Mammography Risk Models Predict When Cancer Occurs?

链接: https://arxiv.org/abs/2611.20274
作者: Svetlana Krasnova,Andrey Krylov,Dmitry Sorokin
类目: Quantitative Methods (q-bio.QM); Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
备注: 7 pages, 2 figures, 1 table. Submitted to ISBI 2027

点击查看摘要

Abstract:Longitudinal deep learning models for mammography-based breast cancer risk prediction output either per-year cumulative risks or a discrete distribution over the year of diagnosis, from which cumulative risks F1,…,F5 are derived. These models are evaluated almost exclusively with year-wise AUCs and the concordance index. We show that these metrics are largely insensitive to when cancer occurs: they can be high even for a model that carries no information about the year of diagnosis. We re-evaluate two recent longitudinal models (OA-BreaCR, LMV-Net) on CSAW-CC with a 5-year horizon using timing metrics computed on cancer cases only. Despite C-indices up to 0.75, neither orders cancer cases by time to diagnosis significantly better than chance (c_time 0.51 and 0.55, both 95% CIs covering 0.5). Weighted absolute timing error favours a constant “year 3” prediction over both models, so we use c_time as the primary timing metric. For Mirai’s additive-hazard head, also used by LMV-Net, we analyse how per-year increments passed through a ReLU can collapse under severe class imbalance. We also propose a breast-level, two-stage model that separates whether from when. It matches LMV-Net in discrimination (C-index 0.745 vs. 0.746) but orders cancer cases better (c_time 0.58 vs. 0.51, paired 95% CI of the difference 0.01-0.13), on par with Mirai’s published weights (0.57).

[CV-317] Streaming across viewing conditions - The effect of viewing distance and display luminance on video quality

链接: https://arxiv.org/abs/2611.20226
作者: Dounia Hammou,Lukáš Krasula,Christos G. Bampis,Zhi Li,Rafał K. Mantiuk
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV); Multimedia (cs.MM)
备注:

点击查看摘要

Abstract:The visibility of streamed video distortions is affected by the user’s display and viewing conditions, such as their display’s luminance or the viewing distance. Distortions become less visible at lower luminance levels or farther viewing distances. However, most image and video quality metrics assume fixed optimal display and viewing conditions, making them insensitive to the viewing conditions. In this work, we demonstrate how existing metrics can be adapted to account for both factors and compare such an adaptation with the metrics that natively model viewing conditions. First, we collected a new video quality dataset, HDR-VDC, which contains streaming distortions in 4K HDR videos. The dataset consists of 12 4K and 4 FHD HDR source video content, each distorted with AV1 encoding and Lanczos upscaling from a lower resolution. The quality loss due to the distortions was measured at two viewing distances, corresponding to 60 and 120 pixels per visual degree, and two display mean luminance levels, varying by a factor of 8 in terms of luminance. Next, we developed a pre-processing framework that lets us adapt many quality metrics so that they can evaluate content across the viewing distances and luminance levels. The framework employs a display model and perceptually uniform encoding, PU21, to account for the effect of display luminance, and perspective projection scaling (PPS) to account for the effect of viewing distance. We used the framework to compare the performance of 35 state-of-the-art metrics. This allowed us to identify a selection of existing quality metrics that can be effectively adapted to account for the viewing conditions.

[CV-318] Pixel-KWS: Rendered-Glyph Representations for Orthographically Robust Open-Vocabulary Keyword Spotting

链接: https://arxiv.org/abs/2611.19389
作者: Adarsh Arigala,Ramesh Gundluru,Srinivasan Umesh,Ludovick Lepauloux,Hakim Hacid
类目: Audio and Speech Processing (eess.AS); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Sound (cs.SD)
备注: 5 pages, 4 figures, 3 tables

点击查看摘要

Abstract:Open-vocabulary keyword spotting (KWS) allows users to enroll arbitrary keywords at inference time. However, character and phoneme-based text encoders rely on fixed symbol vocabularies learned during training, limiting their ability to handle unseen or visually similar symbols such as homoglyphs, l33tspeak substitutions, and accented characters. We propose PIXEL-KWS, an open-vocabulary KWS framework that represents characters through rendered glyphs rather than learned embedding lookups. We render each character as a grayscale image and encode it with a lightweight CNN, enabling the text encoder to generalize from visual form rather than a fixed vocabulary. On LibriPhrase, PIXEL-KWS and its matched character-based counterpart achieve 9.57% and 9.55% EER on LibriPhrase-hard, 0.43% and 0.41% EER on LibriPhrase-easy, outperforming all published baselines without pre-training or synthetic data. PIXEL-KWS provides substantially greater robustness to both homoglyph and l33tspeak symbol-level perturbations. These results highlight glyph-based representations as a simple and effective approach for robust open-vocabulary KWS in practical settings.

[CV-319] XLSAR-Bench: Towards Extremely Large SAR Imagery Understanding of MLLM s

链接: https://arxiv.org/abs/2611.19364
作者: La Jiang,Lingfeng Chen,Jiaqi Zhu,Fengrong Guo,Jiehua Zhang,Panhe Hu,Li Liu,Zhen Liu
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:The rapid development of multimodal large language models (MLLMs) has highlighted the need for comprehensive benchmarks to assess their capabilities and limitations in remote sensing (RS). However, evaluating MLLMs in large-size synthetic aperture radar (SAR) scenarios remains challenging, as understanding complete scenes requires both global semantic interpretation and fine-grained target perception. Existing SAR benchmarks predominantly rely on small image patches, limiting their ability to assess these capabilities jointly. To address these limitations, we present XLSAR-Bench, a comprehensive benchmark for evaluating the perception and reasoning capabilities of MLLMs in large-scene SAR imagery. XLSAR-Bench comprises 405 images with an average size of 7153 \times 7153 pixels, ten tasks spanning visual question answering, visual grounding, and detailed image captioning, with all evaluation samples meticulously annotated manually, assisted by a semi-automatic captioner. Experimental results from general-purpose and domain-specific MLLMs reveal persistent challenges in understanding large SAR scenes and accurately perceiving small targets. We have open-sourced XLSAR-Bench to support further research in developing more powerful MLLMs for complete large SAR scenes.

[CV-320] Learned Regularization for Sparse-View CT: Accuracy Lesion Detectability and Distribution Shift of a Plug-and-Play Prior against Tuned Total Variation

链接: https://arxiv.org/abs/2611.18936
作者: Mohammed Abo Oraig
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV); Signal Processing (eess.SP)
备注: 62 pages, 22 figures, and 30 tables. Preprint (not peer-reviewed); also available on Zenodo: this https URL . Code and stored results are available at: this https URL

点击查看摘要

Abstract:Learned image priors are often proposed to replace hand-crafted regularizers in sparse-view computed tomography (CT), yet they are usually judged by PSNR on data resembling their training set. We evaluate one noise-conditioned DnCNN denoiser used as a Plug-and-Play prior against filtered back-projection (FBP), Tikhonov and total-variation (TV) regularization, with every method tuned on validation data and all differences paired. On synthetic 28x28 digit images, the prior beats box-constrained TV by about 3.4 dB at moderate noise (sigma = 0.05, 20 angles), but the advantage shrinks at low noise and reverses under distribution shift: an MNIST-trained prior is 3.1 dB (Fashion-MNIST) and 4.3-4.6 dB (Shepp-Logan phantom) below TV, whereas a Fashion-MNIST-trained prior beats TV by 2.2 dB on Fashion-MNIST, consistent with prior mismatch rather than a failure of PnP itself. In a lesion-insertion test the prior detects lesions worse than TV in distribution despite its higher PSNR, and under mismatch it produces more lesion-like false blobs than box-constrained TV for seven of eight lesion types. On real LoDoPaB-CT slices (patient-disjoint split; 30, 60 and 120 angles; three dose levels; 40 test patients), the prior does not consistently outperform a well-tuned TV in PSNR (-0.93 to +0.33 dB) and has lower SSIM in eight of nine conditions. Its lesion-detection AUC is higher than this TV’s in all nine conditions (significant in six) but never significantly higher than that of Hann-filtered FBP, which the PSNR-tuned TV falls below in detection in all nine conditions (seven significant). On non-anatomical phantoms its PSNR falls below TV in all six conditions. PSNR and lesion detectability thus rank the methods differently. Our conclusions concern one small denoiser, and the CT analysis was developed after seeing results, without pre-specified primary comparisons.

[CV-321] Foundation Model Fine-Tuning with Attention-based Multi-View Fusion for Multi-Task Cardiac Function Assessment MICCAI-2026 MICCAI2026 MICCAI

链接: https://arxiv.org/abs/2611.18816
作者: João Freitas,Warre Snaet,Jan D’hooge,Sandro Queirós
类目: Image and Video Processing (eess.IV); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: 9 pages, 1 figure. Original pre-peer-review manuscript. Accepted for publication in the MICCAI 2026 Satellite Events (Workshops and Challenges) proceedings. The accepted version includes methodological revisions and is available at this https URL

点击查看摘要

Abstract:In cardio-oncology, left ventricular ejection fraction (LVEF) is the standard biomarker for monitoring cardiac function during cancer therapy. However, its estimation from echocardiography is highly operator dependent, and clinically meaningful declines are often detected only after myocardial injury has progressed, motivating automated methods for standardized assessment and earlier detection of LV dysfunction. To address this need, we propose a multi-task framework integrating cardiac cycle-based temporal standardization, multi-view spatiotemporal feature extraction, and joint optimization for LVEF estimation and LV dysfunction classification. Apical four- and two-chamber echocardiographic sequences are temporally standardized by sampling a clip over an estimated cardiac cycle before being processed by the EchoPrime foundation model, which is fine-tuned end-to-end within the proposed architecture. View-specific features are fused through an attention-based aggregation module to generate a patient-level embedding, from which two task-specific prediction heads estimate continuous LVEF and binary LV dysfunction. To broaden the LVEF distribution encountered during training, samples from the MIMIC-IV-ECHO dataset are incorporated using a balanced sampling strategy. Evaluated on the EchoRisk-MICCAI Challenge, our method achieves an MAE of 4.003 for LVEF estimation (Task 1) and an AUC of 0.764 for LV dysfunction detection (Task 2) on the validation set, demonstrating the effectiveness of foundation model fine-tuning with attention-based multi-view fusion for automated cardiac function assessment in cardio-oncology.

[CV-322] Accelerating Variational Cone-Beam CT Reconstruction with Learned Warm Starts

链接: https://arxiv.org/abs/2611.18801
作者: Héloïse Dudoignon,Clément Delmas,Laurent Facq,Clair Poignard,Baudouin Denis de Senneville
类目: Image and Video Processing (eess.IV); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV)
备注: Submitted to IEEE ISBI 2027

点击查看摘要

Abstract:Variational reconstruction provides a rigorous framework for cone-beam computed tomography (CBCT) by combining an explicit acquisition model with image priors. However, iterative optimization remains computationally demanding, particularly under severe noise and sparse-view conditions. Although learning-based methods achieve high reconstruction quality, they often replace the original variational formulation with learned optimization schemes or nonlinear equilibrium models. We propose a neural network-guided warm-start framework that accelerates variational reconstruction while preserving its original formulation. Projection data are first reconstructed using the Feldkamp-Davis-Kress (FDK) algorithm, and the resulting image is refined by an image-domain neural network. The network output is used solely to initialize the subsequent model-based optimization, leaving the forward model, data-fidelity term, regularization, and optimization scheme unchanged. A pre-regularization step further improves compatibility with the variational priors. Experiments on pelvic CT images from the SynthRAD2025 challenge show a more than threefold reduction in reconstruction time while maintaining comparable image quality.

[CV-323] Personalized and Adaptive Federated Learning for Canine Cardiomegaly Keypoint Prediction

链接: https://arxiv.org/abs/2611.18789
作者: Jialu Li,Youshan Zhang,Honggang Wang
类目: Image and Video Processing (eess.IV); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV)
备注:

点击查看摘要

Abstract:In the field of medical imaging, limited data and cross-institutional heterogeneity remain major challenges for training robust deep learning models. Federated learning (FL) has emerged as a promising approach that enables collaborative training across institutions while preserving data privacy by keeping datasets local. In this work, we propose a structure-aware personalized and adaptive federated learning framework, PAFL, for automatic vertebral heart score (VHS) estimation and cardiomegaly classification from canine thoracic X-rays. Our method treats the task as anatomical keypoint prediction followed by geometry-based VHS computation, introducing implicit structural consistency and improving interpretability. To address cross-institutional heterogeneity, we adopt a parameter-decomposed architecture that separates shared feature representations from client-specific adaptation modules. Furthermore, we introduce an adaptive, quality-aware aggregation strategy that dynamically weights client contributions based on both dataset size and model quality, thereby enhancing robustness to non-IID data distributions. Experimental results demonstrate that the proposed PAFL framework improves models’ predictive performance compared to isolated training and to conventional and existing personalized federated learning (PFL) methods. Our results demonstrate the potential of integrating structure-aware modeling with PFL for reliable and interpretable medical image analysis.

[CV-324] RouteDiff: Low-Dose CT Denoising via Fidelity-Refinement Routing and Hypergraph State Regulation

链接: https://arxiv.org/abs/2611.18786
作者: Xinyuan Wang,Ruili Li,Tianyang Li,Yan Li,Xuepeng Liu
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV)
备注: 16 pages, 10 figures, 7 tables. Submitted to Knowledge-Based Systems

点击查看摘要

Abstract:Diffusion models have become a strong prior for low-dose computed tomography (LDCT) restoration. However, existing samplers apply one structure-detail trade-off to the whole image, set the restoration emphasis of each step by the timestep, and derive each next state recursively from the current one, carrying its errors forward. We propose RouteDiff, a state-aware dual-route diffusion framework that routes detail to where it can be trusted, regulates each reverse step by the state it has reached, and re-enters the trajectory from an anchored estimate. Fidelity-Refinement Routing (FRR) equips one shared U-Net with a conservative fidelity route and a detail-oriented refinement route, fuses them pixel-wise with a confidence map, and exposes their disagreement as route evidence. Hypergraph State Regulation (HSR) organizes the current features and this evidence into a state hypergraph, whose soft hyperedges group scattered regions in the same state. Its readout yields an ordered stage judgment that modulates the bottleneck, and its incidence guides feature interaction in the skip connections. Estimate-Anchored Re-entry (EAR) re-projects a route-aware anchor onto the trajectory at a fixed step and hands the new state back to HSR to be judged again. With the ten network evaluations of its baseline, RouteDiff achieves the best peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) on four datasets with inserted, real and simulated noise (40.95 dB and 95.13% on Mayo 2020; 45.18 dB and 97.23% on Mayo 2016). It also generalizes to unseen doses of real scans and yields the highest Dice (80.92%) in downstream segmentation.

[CV-325] Manifold-Constrained PET Reconstruction with Learned Flow-Matching Priors

链接: https://arxiv.org/abs/2608.29158
作者: Hengjia Ran,Jie Luo,Yutao Zhu,Rui Hu,Fenglin Zhan,Jianyong Jiang,Huafeng Liu,Bo Zhao
类目: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV); Medical Physics (physics.med-ph)
备注:

点击查看摘要

Abstract:Image reconstruction for positron emission tomography (PET) is an ill-posed Poisson inverse problem. In this work, we introduce an unsupervised, optimization-based reconstruction framework that employs a flow-matching generative model as a learned manifold prior. We train the generative model on high-quality PET images to learn a differentiable ordinary differential equation (ODE)-based transport from a Gaussian latent distribution to the PET image distribution. We incorporate this generator as an explicit manifold constraint into a regularized Poisson likelihood formulation. To solve the resulting problem, we develop an alternating direction method of multipliers (ADMM) algorithm that integrates an expectation-maximization (EM)-type surrogate update for Poisson data fidelity with gradient-based latent-space optimization. Numerical simulations and in vivo PET experiments demonstrate improved reconstruction accuracy compared with conventional reconstruction methods and recent deep learning-based approaches, particularly in low-count settings, with competitive computational efficiency.

[CV-326] An XAI Approach to Deep Learning Models in the Detection of DCIS

链接: https://arxiv.org/abs/2106.14186
作者: Michele La Ferla,Matthew Montebello,Dylan Seychell
类目: Image and Video Processing (eess.IV); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
备注: 12 pages, 5 figures

点击查看摘要

Abstract:The results showed that XAI could indeed be used as a proof of concept to begin discussions on the implementation of assistive AI systems within the clinical community.

人工智能

[AI-0] Mitigating Catastrophic Forgetting with Physics-Indexed Specialist Memory: the Evolutionary Neuroplastic Agent

链接: https://arxiv.org/abs/2611.21598
作者: Blas Eugenio Vicco Zarate,Miguel Hernán Vicco
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Neural and Evolutionary Computing (cs.NE)
备注: 25 pages, 7 figures, 23 tables. Code and data: this https URL

点击查看摘要

Abstract:Continual reinforcement learning requires adaptation to changing conditions while preserving performance under previously encountered dynamics. It was investigated whether specialists generated by evolutionary search can provide a reusable memory when combined with environment detection and failure-driven search. The Evolutionary Neuroplastic Agent (ENA) evolves small neural networks without gradients, archives selected policies by detected dynamic regime, and retrieves them through Bayesian policy reuse. The architecture was evaluated in non-stationary CartPole and HalfCheetah environments. In the 30-replicate CartPole ablation, the complete architecture achieved a mean success rate of 83.2% across the five test environments. Among variants that learned the harder CartPole environments, the complete architecture showed the most consistent retention. A selection ablation also showed that the retrieval rule affected performance. In the evaluated sequences, ENA adapted more slowly than PPO on CartPole and achieved lower returns than SAC on HalfCheetah. The evidence concerns the complete architecture and does not isolate gradient-free optimization as the cause of retention or establish superiority over other continual-learning methods.

[AI-1] ScenarioForge: Resource-Grounded Generation and Execution Validation of Cyber Exercises

链接: https://arxiv.org/abs/2611.21579
作者: Jaime C. Acosta,Anantaa Kotal,Aritran Piplai
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Cyber-range and capture-the-flag scenario construction remains labor-intensive because authors must coordinate infrastructure, vulnerable services, attack paths, artifacts, and success conditions. LLMs can propose detailed exercises, but plausible descriptions do not guarantee deployable software or verifiable objectives. We present ScenarioForge, which turns high-level requirements into resource-grounded topologies, services, traffic, segmentation, and challenge chains, then instantiates and validates them in the Common Open Research Emulator (CORE). The ScenarioForge Execution Dataset contains 226 materialized scenarios covering 294 Vulhub images and all 85 custom-built challenge-resource generators developed with our AI-assisted scaffolding. All 226 passed construction, preview, sequencing, deployment, strict runtime validation, and live artifact checks; their 867 steps, each an individually verifiable challenge objective, comprise 457 vulnerability-backed and 410 generated-resource-backed instances. In a separate 50-prompt matched benchmark, ScenarioForge+Qwen3.8 and ScenarioForge+GPT5.6 each achieved 50/50 (100%) topology/build success and 160/160 intent fields; direct frontier generation reached 41/50 builds and 147/160 fields only with the full cumulative scaffold: structure, vocabulary, policy, and container-image guidance. The test still omitted ScenarioForge’s downstream challenge-resource generation, dependency/flag sequencing, deployment, and live validation. The artifacts support operationally rich, reproducible exercises for training and evaluation.

[AI-2] When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents

链接: https://arxiv.org/abs/2611.21577
作者: Zeynab Anbiaee,Sajjad Dadkhah,Ali A. Ghorbani
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language model (LLM) applications increasingly act as agents that select and invoke external tools. The Model Context Protocol (MCP) standardizes this interaction by allowing clients to compose tools advertised by multiple servers. However, tool metadata used for discovery and selection is not, by itself, necessarily bound to a verified provider identity. This paper studies this gap as tool-provider misbinding: a failure in which a request intended for a legitimate provider is instead resolved to an attacker-controlled tool. The attacker does not compromise the host, LLM, legitimate server, or network channel; it only operates another MCP server that advertises tools with chosen metadata. We formulate tool provider misbinding as an attacker-controlled tool-selection optimization problem and evaluate it in a controlled MCP environment across five LLM-based clients. The results show that interactions drive misbinding. The strongest failures occur when exact-name collision, task-aligned description, and compatible schema reinforce one another, but authority-enhanced descriptions with compatible schema can also cause wrong-provider execution or unresolved provider ambiguity without name collision. Different models expose the same provider-binding weakness differently, including attacker-controlled provider selection, clarification requests, and unstable or inconsistent provider resolution.

[AI-3] Severe but Harmless: Stress-Testing Shift Severity for Frozen Time-Series Forecasters

链接: https://arxiv.org/abs/2611.21562
作者: Korede R. Bishi,Dominic Okonkwo,Adetayo Okunoye
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 10 pages, 6 figures, 5 tables. Accepted at IEEE BigData 2026

点击查看摘要

Abstract:Zero-shot time-series foundation models (TSFMs) have become attractive alternatives to task-specific models because they handle diverse downstream tasks, including anomaly detection and forecasting, without per-series training. However, time-series data undergo distribution shifts, and the dissimilarity between two regimes is commonly scored with statistical measures such as Wasserstein-1, where a higher score is read as a more severe shift. In systematic experiments using three established severity measures on nine regime shifts from three standard traffic datasets, we show that these measures do not reliably predict degradation for instance-normalized TSFMs. The regimes rated most severe are not the regimes that degrade the models most, and degradation reaches 29% in the lowest-scoring regime. In most cases, the audit shows no positive relationship between measured severity and model degradation. We provide an equivariance-based stress-testing analysis that explains why severity readings can become decoupled from model degradation.

[AI-4] RACE: Topological Representations for Anomaly detection through Contrastive Embeddings

链接: https://arxiv.org/abs/2611.21548
作者: Sydney Johns
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: Accepted at ARTMAN 2026

点击查看摘要

Abstract:Intrusion detectors must recognize evolving malicious behavior even when confirmed attack labels are scarce or incomplete. We present TRACE, a topology-aware graph contrastive learning framework for label-efficient intrusion detection. TRACE learns from benign system-activity graphs by combining message-passing and topological representations, then assigns anomaly scores in the same projected embedding space optimized by the contrastive objective. This design captures both local interactions and graph-scale structural deviations without using attack labels during representation learning. A labeled validation set containing benign and malicious graphs is used only to calibrate the anomaly threshold. We evaluate TRACE on StreamSpot, GraSec-IoT, and Unicorn Wget against a supervised graph neural network and the benign-only GraphCL baseline. TRACE obtains F1 scores of 0.923, 0.790, and 0.583, respectively. Compared with GraphCL, this is higher on StreamSpot (0.450) and Unicorn Wget (0.203), but lower on GraSec-IoT (0.975). The supervised model leads on StreamSpot and GraSec-IoT, whereas TRACE has the highest F1 on Unicorn Wget. The mixed results show that topology-aware benign-only learning can improve detection in some graph settings but does not provide a uniform advantage.

[AI-5] SciResearch-Eval: Evaluating Scientific Judgment in Autonomous Experimental Agents

链接: https://arxiv.org/abs/2611.21541
作者: Brian T. Weston,Alondra Torres Contreras,Aldair E. Gongora
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Computational Physics (physics.comp-ph)
备注: 28 pages, 7 figures. Code and leaderboard: this https URL

点击查看摘要

Abstract:We introduce SciResearch-Eval, an interactive benchmark of scientific judgment built on a standard laboratory problem: measuring a material’s thermal properties by heating it and fitting a heat-conduction simulator. An agent designs its own experiments, uses a supplied simulator that may be under-resolved or physically wrong, and reports each property with a 95% interval. To our knowledge it is the first benchmark to combine agent-designed experiments, a deliberately flawed simulator, and proper scoring of calibrated intervals against exact ground truth. On a held-out split (93 tasks, 9 models, 2,511 episodes), models handle procedural traps. When the simulator is missing physics, most report wrong values while stating they are 95% sure: their ranges contain the truth in 20-62% of cases, against 94% for GPT-6 Astra and 99% for Claude Fable 5.2. At default effort the failing Claude 5.5 models rarely diagnose the missing physics (Opus in 35% of episodes, against 66% for Astra), and even when they do, their ranges often miss; Astra and Fable design experiments in which the flawed simulator is accurate, such as gentle heating, which keeps Astra calibrated even when it does not name the cause. Asking for wider intervals does not help; stating the strategy improves interval scores 1.6-2.5-fold, replicated on a pre-registered fresh split, but coverage barely moves. A pre-registered effort ablation shows that the correction is latent in the Claude 5.5 models: at maximum effort Opus improves more than tenfold and is no longer detectably behind Astra at default, but only with 13-31 times more reasoning; older and open-weight models gain nothing. Fable 5.2 is the most accurate model at default effort (interval score 1.9 times better than Astra’s) but writes 6 times as many tokens, and at Astra’s token count it falls short. We release the environment, graders and harness.

[AI-6] Algorithm Distillation for Safe In-Context Reinforcement Learning

链接: https://arxiv.org/abs/2611.21524
作者: Minjae Kwon,Amir Moeini,Shangtong Zhang,Lu Feng
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Safe in-context reinforcement learning (Safe ICRL) asks a policy with fixed parameters to improve on an unseen constrained task from its interaction history while aiming to keep its cost within a budget. Prior work studied Safe ICRL on five benchmarks in three families: a grid world with a hidden goal and obstacles (DarkRoom), continuous navigation where goals and hazards stay invisible until reached (PointGoal, CarGoal), and locomotion with a hidden velocity limit (HalfCheetah, Ant). Safe Algorithm Distillation (Safe AD), trained on the learning histories of a safe RL algorithm, has shown adaptation failures in prior evaluations. For example, in the locomotion tasks its return stays below zero, and it rarely reaches the goal in CarGoal. We revisit two parts of Safe AD. First, we distill a budget-aware teacher with two phases. It searches for the hidden task, exploring more when more budget remains. After its first reward or cost, an expert policy exploits the task within the remaining budget. We collect the teacher’s histories under fixed or dynamic budgets. Dynamic budgets increase the occupancy entropy of the return and budget inputs in every environment, and at test time they improve return and lower cost in DarkRoom and HalfCheetah. Second, we propose A-Lag, a test-time adjustment that reweights the distilled policy’s actions by their predicted cost, with an adaptive Lagrange multiplier that is updated from the costs observed. In our experiments, A-Lag had lower mean episode cost than Safe AD in all five environments, with lower mean final return in four.

[AI-7] Safe Resemblance Silent Propagation: Limits of Text-Similarity Guardrails for AI Agent Safety

链接: https://arxiv.org/abs/2611.21520
作者: Charles Xiao
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:When AI agents relay information to one another, an output can preserve an injected instruction’s effect on downstream agents even after its wording changes, yet legitimate quotation and template reuse can produce high textual overlap with the source, so text-similarity guardrails that attempt to flag the former risk conflating these two cases. We test this distinction using DonkeyRail, a public classifier combining BLEU, ROUGE-L, and METEOR features, deployed in a controlled three-agent email relay, defining a behavioral measure in which an output counts as functional (the injected procedure induces fresh computations in two subsequent agents) independent of the detector’s score, alongside legitimate controls that explicitly authorize quotation and reuse. At its default threshold the classifier misses 8 of 19 functional outputs while flagging 70 of 192 legitimate replies, with every completed quotation and template-reuse reply triggering a false alert; calibrating for low false positives on one template set does not transfer, as the recalibrated threshold detects only one functional output and still produces 15 false alerts on new templates. We also identify an operating-point error in all 21 published classifier configurations, where ROC interpolation replaces direct threshold application, and show formally that high AUC against a surface-level text criterion provides no distribution-free guarantee of high AUC against the behavioral target, while the existing evaluation labels lack the joint observations needed to estimate the gap. An authorization analysis further shows that identical visible transcripts can require different safety decisions depending on the governing policy, a distinction no observation-only classifier can resolve. These results argue for behavioral and policy-grounded evaluation over text-similarity guardrails.

[AI-8] Lazy-MOSAIC: Accelerating Skill-Centric Manipulation Planning with Learned Skill-Effect World Models

链接: https://arxiv.org/abs/2611.21517
作者: Itamar Mishani,Mukund Krishnakumar,Oren Salzman,Maxim Likhachev
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注: Under review

点击查看摘要

Abstract:Long-horizon manipulation tasks require a robot to perform a sequence of physical interactions with its environment, typically combining several different skills. Planning such behaviors efficiently is challenging: given a library of parameterized skills, solutions lie in a vast hybrid space of skill sequences and their parameters, and evaluating each candidate skill’s feasibility and effect with a high-fidelity world model such as a physics simulator is computationally expensive. MOSAIC, a recent skill-centric approach, lets the skills guide a non-sequential search toward the regions of this space where they are likely to succeed; however, it pays the full cost of simulation for every parameterized skill it considers. In this work, we introduce Lazy-MOSAIC, a lazy planning framework that instead searches over hypothesized skill outcomes. A cheap learned skill-effect world model predicts, in a fraction of a millisecond, a skill’s feasibility and the world state it produces. Lazy-MOSAIC then uses these predictions to focus expensive simulation on promising candidate skills and, when a hypothesis is invalidated, repairs only the affected part of the search, while ensuring that every skill in a returned plan is validated in simulation. We evaluate Lazy-MOSAIC on cluttered-shelf retrieval tasks of increasing difficulty and demonstrate it on a physical robot, showing that it plans substantially faster than baselines with fewer simulator queries, while matching or exceeding their success rate in every setting. Visit this http URL.

[AI-9] he CYBER-BRIDGE project: An AI-Powered Platform for Cross-Border Cybersecurity Compliance and Forensics in the EU

链接: https://arxiv.org/abs/2611.21497
作者: Nikolaos Kekatos,Georgios Koutidis,Charalampos Bratsas,Stefan Bucur,Alexandra-Mihaela Chirlejan,Mihaela Curcă,Angelos Fountoulakis,Michael Ioannou,Paul Lacatus,Alexios Lekidis,Emilios Charalambous,Avgi Michael,Tom Nianios,Simona Purcaru,Antonis Voulgaridis
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
备注: 7 pages, 2 figures, 2 tables. Accepted at the 2026 IEEE International Confer Resilience (IEEE CSR 2026)

点击查看摘要

Abstract:Cross-border cyber attacks on European SMEs and critical-infrastructure operators expose a structural cyber-resilience gap: cybersecurity (threat sharing, detection, response), compliance (framework assessment, certification) and digital forensics (cross-jurisdictional evidence) are jointly regulated by NIS2, the CRA, GDPR and the LED, yet their tooling evolved in silos and rarely interoperates across Member States. The EU-funded CYBER-BRIDGE project integrates the three domains under one architecture and a hash-chained evidence layer: AI-driven CTI sharing with ENISA-aligned notification, multi-framework assessments with retrieval-augmented advisory, and RFC 3161-anchored custody with lawful GDPR/LED exchange, all aligned with MITRE ATTCK. A sanitised replica of the back-end image hosts a Cyber-Range for upskilling. The contribution is not the individual tools but the reconciliation mechanism that lets one incident produce, from a single evidence record, a cross-border CTI advisory, the regulatory notifications its deadlines demand and a re-verifiable forensic pack; we set out why binding three differently regulated evidence regimes onto one record is non-trivial and how the shared evidence layer resolves the resulting tensions. This paper introduces the project’s scope, requirements, pilots, architecture and toolkits; quantitative KPIs follow in later deliverables.

[AI-10] Differentiable Exact Learning NEURIPS2026

链接: https://arxiv.org/abs/2611.21486
作者: Hristo Papazov,Francesco D’Angelo,Nicolas Flammarion
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Formal Languages and Automata Theory (cs.FL); Machine Learning (stat.ML)
备注: Accepted in NeurIPS 2026

点击查看摘要

Abstract:We introduce a framework for differentiable exact learning that couples an unbounded external environment to a state-tracking neural controller, which we train on policy-trajectory observations (PTOs): expert behavior traces that expose observations and actions while hiding internal states. In addition to the learning framework, we introduce Differentiable Finite-State Transducers (DFSTs) which use multilinear updates, support log-parallel training, and emulate arbitrary finite-state controllers, yielding computational universality through environmental interaction. From tiny training datasets, compact DFSTs achieve extreme length generalization on arithmetic and universal-simulation tasks. We replicate the success of DFSTs on RNNs and GRUs, thereby supporting a broader learning recipe: State-Tracking Controller + Unbounded External Environment + PTOs = Extreme Length Generalization. Beyond length generalization, our metric-extraction protocol recovers interpretable discrete controllers from trained DFSTs, which we prove equivalent to the target arithmetic and universal-simulation algorithms. Furthermore, for our Softmax DFST models, simply taking \mathrmargmax over the transition probabilities yields discrete controllers which provably exactly perform universal Turing-machine simulation and binary arithmetic. Proving exact learning for the neural controllers requires assumptions, which we empirically support. We make our code, training recipes, model weights, and proof checkers publicly available at this https URL.

[AI-11] RESILAGENT : A Three-Pillar Architecture for Cyber-Physical Resilience in Electrical Power and Energy Systems

链接: https://arxiv.org/abs/2611.21480
作者: Nikolaos Kekatos,Georgios Koutidis,Konstantinos Antonakopoulos,Stylianos Basagiannis,Sokratis Katsikas,Georgios Kavallieratos,Alexios Lekidis,Tom Nianios,Elpiniki Papageorgiou
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Systems and Control (eess.SY)
备注: 7 pages, 5 figures, 1 table. Accepted at the 2026 IEEE International Conference on Cyber Security and Resilience (IEEE CSR 2026)

点击查看摘要

Abstract:Electrical Power and Energy Systems face hybrid threats that combine cyber intrusions, physical sabotage and climate-driven stress, and routinely outpace today’s siloed monitoring stacks. We present RESILAGENT, a concept architecture organised around three operational pillars, Sensing Data Intelligence, Resilience Analytics Reporting, and Incident Response Recovery, mediated by a Cloud-based Orchestrator and exercised through a closed Detect-Simulate-Decide-Sign loop. The framework fuses agentic artificial intelligence and multimodal large language models for cross-modal detection, real-time digital twins for what-if simulation, and a permissioned blockchain anchor for tamper-evident threat intelligence and CER/NIS2-aligned evidence. We instantiate it on two scenarios, a transmission operator under coordinated cyber compromise, sabotage and heatwave, and a distribution operator under a supply-chain compromise during a storm, and report an IEEE-bus micro-benchmark that quantifies each pillar through false-data-injection attack detection, cyber-physical fusion lift, N-1 contingency screening and anchor cost.

[AI-12] Generative AI and the Crisis of Higher Education Assessments: Examining AI-Proofing Strategies for College-Level Assessments

链接: https://arxiv.org/abs/2611.21476
作者: Bonam Mingole,Davis Yadav,Hui Min Lee,S. Shyam Sundar,Amulya Yadav
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Higher education is confronting an academic-integrity challenge unlike any in its history. Students can now outsource homework, take-home assignments, and online examinations to large language models (LLMs) that complete them in seconds. Existing research on evaluating LLM performance on academic assessments has focused largely on standardized examinations, or on questions from college-level courses drawn from a single discipline. Far less is known about how LLMs perform on college-level assessment questions spanning multiple disciplines, that instructors have purposely designed to be AI-resistant, and about which features of such questions, if any, meaningfully constrain LLM performance. We address this gap by collecting 75 questions from authentic college coursework assessments, spanning 17 disciplines, which were contributed by 37 faculty members from ten U.S. universities and one U.K. university. Of these, 61 questions were intentionally designed by faculty members to be ‘LLM-proof’, while the remaining 14 questions were sourced from online courses. We then (i) evaluate three state-of-the-art LLMs’ ability to correctly answer these questions with and without supplementary course materials; (ii) identify five strategies, through thematic analysis, that faculty used to make their questions harder for LLMs to solve; and (iii) test whether these strategies impact LLMs’ ability to answer the questions accurately. Our experimental results show that the best-performing model scored approximately 83% without supplementary course materials and 85% with them, and we observed noteworthy differences across models, answer formats, and instructor tweaking strategies. We discuss implications for assessment design, academic-integrity policy, and the broader ethics of deploying LLMs in educational settings.

[AI-13] Network Intrusion Detection Adaptability under Spurious Labels: A Noise-Aware Few-Shot Class-Incremental Learning Framework

链接: https://arxiv.org/abs/2611.21473
作者: Zijian Zhao,Davide Di Monda,Domenico Ciuonzo,Yipeng Wang,Antonio Pescapè
类目: Networking and Internet Architecture (cs.NI); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Network Intrusion Detection Systems (NIDSs) must incrementally adapt to emerging attacks, even when only a limited amount of labeled network traffic is available. Consequently, Few-Shot Class-Incremental Learning (FSCIL)-based NIDSs have gained growing attention. However, existing FSCIL-based NIDSs are highly vulnerable to label noise in newly-observed attack classes. To address this limitation, we propose NAFN, a noise-aware FSCIL-based NIDS framework that reduces the impact of mislabeled data during incremental adaptation. NAFN consists of two components: (i) Noise Detection, which identifies and removes mislabeled biflows using noise-detection algorithms, and (ii) Feature Sampling, which estimates class-conditional embeddings from clean biflows, calibrates them with base-class statistics, and resamples synthetic features to recover the original biflow count. The framework follows the Rethink Few-Shot (RFS) teacher-student paradigm with lightweight nearest-neighbor or logistic regression classifiers, and various pretraining memory selection strategies. Experiments on CSE-CIC-IDS2018, Edge-IIoT, and IoT-NID datasets show that NAFN effectively preserves prior knowledge and maintains strong incremental performance even under 50% label noise, consistently improving new-class F1-scores over non-denoised baselines, up to +52%, while being time-efficient and robust in cross-network and high-noise incremental settings. The implementation is publicly available to foster reproducibility.

[AI-14] Subprocess-Constrained Markov Decision Processes NEURIPS’26

链接: https://arxiv.org/abs/2611.21471
作者: Jiarui Gan,Debmalya Mandal
类目: Artificial Intelligence (cs.AI)
备注: To appear at NeurIPS '26. Code available at this https URL

点击查看摘要

Abstract:Constrained Markov decision processes (CMDPs) typically bound expected cumulative cost from an initial state. Such a guarantee can be insufficient when a policy must remain feasible after every history it may encounter. We introduce \emphsubprocess-constrained MDPs (S-CMDPs), which require the expected discounted cost of each continuation, conditional on any reachable history, to satisfy a bound determined by the current state. This stronger requirement changes the structure and complexity of the problem: we show that stationary policies can be strictly suboptimal, and finding even an approximately optimal stationary policy is NP-hard. We adapt a recently introduced value-set iteration framework to show that history-dependent policies surprisingly overcome both obstacles. Our polynomial-time algorithm represents such a policy implicitly: its expected reward is within \epsilon of the optimal feasible value, it violates each subprocess cost bound by at most \delta , and its action distribution can be computed in polynomial time for any given history.

[AI-15] Warm Starting Neural Dynamics Models From Few Jacobian Samples

链接: https://arxiv.org/abs/2611.21465
作者: Maxime Rochkoulets,Joppe De Jonghe,Mariya Ishteva
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:Model-based reinforcement learning (MBRL) leverages a model of the environment, used for planning and generating samples without interacting with the real world. Motivated by evidence that informed initialization significantly improves the efficiency of MBRL, we propose TDec-Init, an algorithm to warm-start neural dynamics models using transition Jacobians. Using a decoupling algorithm based on the decomposition of a tensor of stacked Jacobian samples, we recover weights and flexible nonlinearities, and use them to initialize a feedforward neural network. In the low-data regime, we show on several benchmark environments that our approach outperforms random initialization and value training, and either matches or outperforms Sobolev training, while requiring fewer training steps between episodes.

[AI-16] Expertise capture in workplace AI

链接: https://arxiv.org/abs/2611.21456
作者: Andreas Ekelhart,Ahmad Mohsin,Iqbal H. Sarker,Helge Janicke
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI assistants are becoming part of how knowledge workers think and perform, accumulating the prompts, exemplars, memory and interaction histories that encode a worker’s professional judgment. This infrastructure typically runs under employer or platform control. We name the resulting process expertise capture, the absorption of worker-applied knowledge and judgment into AI resources under organizational control. We develop a framework of contribution forms, technical residences and transformation families, and show that how far a contribution can be attributed, separated and revoked depends on where it resides and how it has been transformed, properties fixed by system design long before any dispute arises. When a worker departs, the organization retains a capability shaped by their judgment while the worker loses access to it, and the asymmetry compounds as contributions move from inspectable artifacts into derivatives, learned components and successor systems. We outline three measures, write-time provenance, extraction of separable contributions and an executable exit procedure, that make an agreed allocation technically performable, and make explicit what cannot be separated at all.

[AI-17] Learning to Reason with Predictive Gains for Generative Recommendation

链接: https://arxiv.org/abs/2611.21441
作者: Zhuoyang Li,Teng Shi,Jiakai Tang,Weijie Yu,Guofu Xie,Xiao Zhang,Lei Kou,Yang Yu,Lei Shen,Xin Chen,Bing Wang,Xiaoyi Zeng,Jun Xu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Reasoning-enhanced generative recommendation (GR) has attracted increasing attention for leveraging the reasoning capabilities of large language models (LLMs) to improve next-item prediction. However, existing methods primarily optimize reasoning through Group Relative Policy Optimization (GRPO) with rule-based rewards determined by final recommendation performance, which faces two key limitations: (i) unmeasured predictive contributions within reasoning, where the predictive contributions of individual reasoning steps are not directly reflected in optimization signals, and (ii) weakly discriminative learning signals within groups, where different reasoning trajectories within the group often receive indistinguishable rewards, limiting effective policy optimization. In this paper, we propose GainReason, a simple yet effective training framework that optimizes reasoning through Predictive Gain (PG). Specifically, PG measures changes in target Semantic ID (SID) log-likelihood induced by individual reasoning sentences, providing contribution-aware and discriminative rewards. We further theoretically establish the alignment between expected PG and the recommendation ranking objective. Building on this signal, we develop PG-based Reasoning Optimization (PGRO) to optimize reasoning generation with sentence-level rewards, supported by a custom causal attention mechanism for efficient PG computation. Additionally, an auxiliary PG-based Reasoning Utilization (PGRU) task adaptively strengthens reasoning-conditioned prediction to complement PGRO. Extensive experiments on three public benchmarks demonstrate that GainReason consistently outperforms competitive baselines and achieves substantial performance gains from reasoning.

[AI-18] A Reinforcement Learning Method to Navigate SMEFT Operator Space

链接: https://arxiv.org/abs/2611.21439
作者: Jacky Kumar
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); High Energy Physics - Experiment (hep-ex); High Energy Physics - Phenomenology (hep-ph)
备注: 15 pages, 3 figures, GitHub Link: this https URL

点击查看摘要

Abstract:In particle physics, discrepancies between experimental measurements and Standard Model (SM) predictions are commonly interpreted within new theoretical models. In this context, the Standard Model Effective Field Theory (SMEFT) provides a general framework encompassing a broad class of models. However, exploring the enormous space of possible SMEFT operators rapidly becomes a challenging combinatorial problem: many operators can contribute simultaneously to different observables, their effects are correlated, and distinct operator combinations can provide comparably good descriptions of the data. In this work, we formulate the exploration of the SMEFT operator space as a sequential decision-making problem using reinforcement learning (RL). Our RL agent employs a lightweight actor-critic architecture with approximately 40K trainable parameters in total. Rather than searching for a single best-fit theory, our method is designed to discover a diverse set of high-quality theory hypotheses. We test the method in a complex phenomenological environment comprising 132 observables across several sectors of particle physics, including 24 measurements that deviate from their SM predictions by more than 1.5\sigma . We show that the learned RL policy efficiently constructs multi-operator theories that substantially improve the global fit. Comparisons with independent random-subset and random-sequential searches demonstrate the ability of the RL policy to navigate this large combinatorial space. Our results establish RL as a systematic framework for exploring diverse patterns of new physics (NP) interactions within SMEFT.

[AI-19] Emergent Steering Beyond Endpoint Alignment in Chemical Reaction Models

链接: https://arxiv.org/abs/2611.21423
作者: Yili Shen,Kehan Guo,Haomin Zhuang,Xiangliang Zhang
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Machine learning models for chemical reaction prediction are usually evaluated by endpoint accuracy, but endpoint correctness alone does not reveal how a model represents chemical transformation or why it fails. We introduce UniRxnAxis, a unified forward-retrosynthesis framework that models both tasks as direction-conditioned endpoint redistribution in a shared latent space. By aligning source states, transformed states, and encoded target endpoints, UniRxnAxis enables directional auditing of learned latent updates. This shared latent geometry allows the learned displacement (from the encoder source state S_0 to the decoder-facing transformed state S_1 ) to be decomposed into an endpoint-parallel component and an endpoint-orthogonal residual, which we call emergent steering. Our evaluation shows that emergent steering is functionally important: in retrosynthesis, it recovers most of the model’s utility, remains effective under random, shuffled, and prototype controls, and directly affects which bond edits are promoted or suppressed. The residual-over-parallel ordering persists across alignment strengths, after decoder-interface calibration, against angle-matched random axes, and in a second graph-edit architecture, although its magnitude depends on the reference axis and optimization. The decomposition also supports a proxy taxonomy of model-side failures, showing how reaction models can be audited beyond endpoint accuracy.

[AI-20] Harness Leverag e Across Model Size: A Ladder of Five Local Agent Harnesses from 0.75B to 4.7B

链接: https://arxiv.org/abs/2611.21407
作者: Hao Wang,Ting Huang
类目: Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
备注: 56 pages, 12 tables, 5 figures

点击查看摘要

Abstract:A harness turns a model into a system that completes tasks: it builds the prompt, exposes tools and decides when work is done. Harness comparisons hold the model fixed. Model comparisons hold the harness fixed. We vary both: five shipped harnesses, 18 artifact-producing tasks, four locally deployable open-weight models from 0.75B to 4.7B and a 35.5B anchor, with one deterministic checker scoring all 360 ladder cells. The completion-controlling harness, Mingbird, has the highest mean at every size. Of four pre-registered paired contrasts, two survive Holm correction, at 0.75B and 4.7B. Thirteen of sixteen paired intervals exclude zero under the deployment reading, with leads from 0.075 over the strongest rival to 0.819 over the weakest. The completion gate is not the carrier: disabling all six executable checks moves the mean by +0.080, +0.117, +0.003 and -0.004, every interval including zero, and only at 1.9B does disabling it create empty runs. Removing one repeat-blocking rule alone costs 0.411 at 0.75B and leaves eight of eighteen tasks with no artifact. The harness effect is a progress effect: it decides whether a run leaves anything on disk. Of 357 usable cells, 127 leave none of their declared artifacts and 85 of those exit normally, so a scorer that reads exit status credits work that was never written. Whether the effect grows or shrinks with size, four rungs cannot settle: a floor arm produces a rise, a ceiling a decline, and a cross-family rung a peak.

[AI-21] VisualArtifactSpec 1.0: Evaluating Visual Artifacts under Task Specifications

链接: https://arxiv.org/abs/2611.21382
作者: Shuyi Ouyang,Xiaojun Tang,Zijian Zou,Xintong Zhang,Mingli Song,Guannan Zhang,Hongwei Xue
类目: Artificial Intelligence (cs.AI)
备注: 24 pages. Project page: this https URL

点击查看摘要

Abstract:Visual artifact tasks test agents’ ability to translate task specifications and source materials into outputs that meet practical needs. Evaluation across artifact families helps distinguish broadly applicable authoring capabilities from strengths confined to a particular medium. We introduce VisualArtifactSpec 1.0 (VAS), a benchmark of 211 creation and modification tasks spanning presentations, static graphics, animations and videos, and web interfaces. Each task specifies authoring objectives, source materials, constraints, and deliverables. Existing benchmarks provide limited support for systematically evaluating how authoring choices suit different audiences, purposes, and media across artifact families. VAS evaluates two aspects separately: Task Fulfillment verifies explicit requirements and applicable functionality, while Artifact Performance assesses Content, Organization, Visuals, and Delivery. Task-specific rating anchors instantiate shared criteria for the intended audience, purpose, and medium, with judgments grounded in source materials and artifact inspection records. We evaluate nine model-harness configurations on all 211 tasks. Overall scores range from 42.12 to 70.52, with Artifact Performance accounting for 73.4% of the gap between the highest- and lowest-scoring configurations.

[AI-22] Proceedings Eighth International Workshop on Formal Methods for Autonomous Systems

链接: https://arxiv.org/abs/2611.21375
作者: Matt Luckcuck(University of Nottingham, UK),Maike Schwammberger(Karlsruhe Institute of Technology, Germany),Akhila Bairy(Karlsruhe Institute of Technology, Germany)
类目: Logic in Computer Science (cs.LO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:This EPTCS volume contains the papers from the Eighth International Workshop on Formal Methods for Autonomous Systems (FMAS 2026), which was held on the 17th and 18th of November 2026. FMAS 2026 was co-located with the 27th International Conference on Formal Engineering Methods (ICFEM), and hosted at the Leonardo Royal Grand Harbour Hotel in Southampton.

[AI-23] Representational Change Under Self-Consuming Training: A Sparse Autoencoder Analysis

链接: https://arxiv.org/abs/2611.21348
作者: Suchit Gupte,Zhongteng Cai,Xueru Zhang,Mohammad Mahdi Khalili
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models are increasingly exposed to model-generated text, raising the possibility that future models will be trained recursively on the outputs of their predecessors. Prior work on such self-consuming training loops has largely focused on output quality and downstream performance, leaving their effects on internal representations poorly understood. We study how representations evolve across five generations of self-consumption and whether sparse autoencoders (SAEs) remain reliable under this shift. At each generation, we compare a lineage trained on synthetic data with a matched control trained on real data under the same training settings. We evaluate how well SAEs trained before self-consumption transfer to later generations, measuring reconstruction and sparsity across SAE widths, architectures, and model layers. We then fine-tune the original SAE on each generation and use sparse probing to test whether task-relevant information remains concentrated in a small number of latents or becomes more distributed across the representation. We find that zero-shot SAE performance degrades progressively under self-consumption, with growing reconstruction and sparsity gaps that vary substantially across layers. Fine-tuning largely restores reconstruction, but not feature-level usefulness: task-relevant information remains accessible in the full latent space while becoming less concentrated in individual features. These results show that self-consumption induces representational drift and that reconstruction fidelity alone is insufficient to establish that an adapted SAE remains equally useful for interpretation.

[AI-24] RIAGE-CPG: A Cross-Model Study of Budgeted Static Malware Triage with Large Language Models

链接: https://arxiv.org/abs/2611.21331
作者: Paul-Andrei Sava,Alexander Küchler,Daniel Kowatsch,Bernhard Grill,Jennifer Pullman,Yanick Fratantonio,Luca Invernizzi
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: Code and artifacts available here: this https URL

点击查看摘要

Abstract:When triaging unknown binaries, security analysts must decide which samples to clear as benign, flag as malicious, or escalate for deeper analysis. Large Language Models (LLMs) could support these decisions by interpreting decompiled code, but a binary can contain hundreds or thousands of functions, making exhaustive LLM analysis impractical. We study how current LLMs behave when static triage is formulated as budgeted evidence acquisition: which functions they inspect and how they trade off benign clearance, malware detection, and escalation. To enable a controlled comparison, we implement TRIAGE-CPG, a fixed pipeline over a code property graph (CPG) of a decompiled binary. An LLM-free heuristic selects seed functions, an agent expands the resulting evidence dossier under a fixed budget, and a reporting stage produces a benign, suspicious, or malicious verdict together with a behavioral report. We evaluate eight closed-source and open-weight models on 1,000 Windows PE binaries, using a 600-sample test split and the same zero-shot protocol for every model. We assess the ternary verdicts under both a strict binary reading and triage balanced accuracy (tBA), which credits malware that is flagged or escalated and benign samples only when cleared. After analyzing about 40 of a mean 1,000 functions per binary, the highest-scoring model reaches a strict balanced accuracy of 0.805 and a tBA of 0.855, flagging 65% of malware, escalating another 12%, and clearing 94% of benign samples. The models implement different triage policies: high-specificity models produce small, malware-rich escalation queues, whereas recall-oriented models escalate or flag much more broadly. Both closed-source and open-weight models appear across this spectrum. Our results identify no universal deployment configuration: model choice itself determines the clear/flag/escalate trade-off under limited static evidence.

[AI-25] Benchleaking Is Accelerating: Benchmark contamination has worsened sharply in the past month – from post-training down to pre-training base models

链接: https://arxiv.org/abs/2611.21315
作者: Yiwei Qin,Muhang Xie,Xuefeng Li,Mohan Jiang,Dayuan Fu,Tiantian Mi,Qipeng Guo,Pengfei Liu
类目: Artificial Intelligence (cs.AI)
备注: 24 pages. Project: this https URL

点击查看摘要

Abstract:Benchmark contamination is not new, but recent model releases show it in a sharper form. We use the term benchleaking for the case in which a benchmark’s own test examples enter training data, so that a score reflects recall rather than capability. This differs from benchmaxxing, which optimizes toward a benchmark without seeing its test set. We document two escalations over the past month. First, in post-training, the traces have become explicit: recent audits show an agent stating hidden test requirements it was never given in 23% of its attempts, where its predecessor did so in none, and continuing to follow them after the task is changed to contradict them (Schaeffer and Räuker, 2026; Lo and Lo, 2026). Second, the problem has reached pre-training. Using few-shot continuation on 20 benchmarks across four public base models, we find that some base models write out other test examples—complete questions, options, and answers—that appear nowhere in their input. For the recently released ERNIE-5.5-Flash-Base, 96% of HumanEval problems leak with their reference implementations and 85% of OpenBookQA questions leak with their answers, against 9% and none for DeepSeek-V4-Flash-Base. To quantify what a reported score owes to such examples, we define an Anti-Hack score: the benchmark score recomputed with every leaked example marked wrong and the number of examples unchanged. Averaged over the suite, it costs DeepSeek-V4-Flash-Base 1.5 points and Hy3-preview-Base 1.3, against 5.0 for MiMo-V2.5-Base and 16.6 for ERNIE-5.5-Flash-Base. Open base models are the foundation of much downstream research, so benchleaking now affects more than leaderboard rankings. We call for decontamination reports and contamination-adjusted scores to become standard practice.

[AI-26] Can AI Mature a Models Geometry? Training-Free Weight Baking in LLM s

链接: https://arxiv.org/abs/2611.21310
作者: Amin Golnari,Saeid Sanei
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Can an AI make a trained language model more mature, in the sense of a better organized latent space, by looking only at the geometry of that space? An analyst model reads a fixed-size digest of topological descriptors of the layers of a frozen target and proposes a transformation, which closed-form ridge regression writes into the feed-forward weights without gradient training. Using lightweight models, 126 runs show that category purity rose on unseen words and sentences, benchmark accuracy stayed within sampling error, and the analyst matched a category-aware heuristic with gentler edits. In a five-model round robin approach, 16 pairs raised purity and four left the target unchanged. The results provide a baseline for future work on LLM-guided, training-free model editing without using raw weights.

[AI-27] Programmable Semantic Interventions: Learning Causal Carriers from Behavioral Programs

链接: https://arxiv.org/abs/2611.21307
作者: Paolo Muratore,Mackenzie Weygandt Mathis
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: this https URL

点击查看摘要

Abstract:Causal abstraction explains a frozen network by aligning its activations with the variables of a simpler, interpretable model, and checking that interventions on the two agree. Finding such an alignment currently requires a counterfactual dataset: for every pair of prompts, the answer the network ought to give once the variable has been changed. In practice, those labels are produced by an executable algorithm for the behavior, written before studying the model supposed to implement it. However, the criterion these methods optimize only leverages the answer an intervention should produce. Here, we develop an approach that lets the investigator define that relationship directly: the answers the variable can produce, and which of them an intervention should yield. We call this interface a semantic program and introduce Programmable Semantic Interventions (PSI): a low-rank activation subspace learned so that interchanging it changes the model’s answer as the program declares, with the expected answers read from the frozen model itself. On the MIB benchmark PSI performs on par with supervised alignment at matched data and training budget, without curated pairs, and it needs no ground-truth labels; while on natural science questions with gold labels withheld, it tracks the model’s own answers better than a subspace trained on gold labels. The declaration also fixes a coordinate system with known semantics, enabling donor-free steering: jointly training PSI for both localization and control realizes 94.8% of held-out requests for answers the investigator chooses, without the need for donor prompts.

[AI-28] uring Agent : Agent ic Formal Verification for Mathematical Research

链接: https://arxiv.org/abs/2611.21299
作者: Yang Chen,Qianyi Wang,Ronghui Yang,Zhanhao Li,Yu Cong,Chen Zhang,Hangfan Zhang,Shao Zhang,Shuai Li,Toru Takisaka,Chao Xu,Yi Zhou,Zhihao Tang,Xiao Wang,Bo Zhang,Shuyue Hu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models (LLMs) are rapidly advancing mathematical discovery, moving beyond benchmark-level problem solving to tackle frontier research problems. This progress is ushering mathematics into an era of proof abundance, raising a fundamental challenge: how can we ensure the correctness of the ever-growing volume of AI-generated proofs and transform them into trustworthy, machine-verified mathematical knowledge? We introduce Turing Agent, an agent harness designed to orchestrate end-to-end formal verification of research-level mathematics. Its wo-scale architecture integrates theorem formalization and proof formalization agents, empowered by structured proof decomposition, parallel verification, accumulated proof experience, and selective semantic review. Leveraging our collaborations with theoretical computer science (TCS) researchers, we focus our evaluation on TCS, a particularly demanding domain characterized by intricate quantification, complex dependencies, and stringent semantic requirements. Against direct generation and general-purpose coding harnesses, including Codex CLI and Claude Code, Turing Agent achieves approximately 57-62% faithful formalization success on Formal TCS Bench and 83% overall pass@1 on a TCS-focused subset of RLMEval, using GLM-5.3 and DeepSeek-V4-Pro. Notably, Turing Agent, built on DeepSeek, successfully completes the formal verification of three previously unresolved TCS tasks from OpenAI’s October 2026 collection. One particularly challenging TCS problem required about a hundred formally verified auxiliary claims. Moreover, by actively exploring and refining the formalization process, Turing Agent can uncover simpler alternative proofs, expose unnecessary dependencies in original proofs, and generate reusable mathematical knowledge as a byproduct of verification.

[AI-29] Agent -Native by Design: A Programming Language for Agent ic Applications

链接: https://arxiv.org/abs/2611.21298
作者: Jayanaka Dantanarayana,Savini Kashmira,Xiaoyu Ma,Malitha,Malitha Prabhashana,Lingjia Tang,Jason Mars
类目: Programming Languages (cs.PL); Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
备注:

点击查看摘要

Abstract:Building agentic applications requires combining large language model (LLM) operations, workflow structure and application state, and model-guided routing. With SDKs and agent frameworks, developers connect these components through supporting code that prepares model inputs, propagates state, and dispatches model-selected steps. Workflow changes can require coordinated updates to this code, increasing development and testing effort. Splitting execution logic between application code and framework internals also limits compiler visibility into control flow and state dependencies, hindering automatic optimization. We introduce Agent-Native Programming, which unifies these three components with shared execution semantics. We realize this approach in Jac, a Python-superset language, through “by llm()” for typed model operations and tool use, Object-Spatial Programming for workflow structure and state, and “visit … by llm()” for model-guided routing. The routing construct derives the model’s choices from executable workflow steps and schedules the selected steps. This explicit connection between model decisions and program execution gives the compiler and runtime a common basis for dependency analysis and automatic optimization. Across five benchmarks and five baselines, Jac achieves a 1.8-8.7x reduction in AI-integration code relative to the OpenAI SDK, with competitive observed task quality and lower coding-agent development effort. Two case studies demonstrate speculative execution and service partitioning through configuration alone, using program analysis and runtime checks to preserve workflow behavior under stated conditions.

[AI-30] CAE-Bench: Benchmarking AI Agents for Reusable CAE Solver Development

链接: https://arxiv.org/abs/2611.21273
作者: Zhiyu Zhou,Tianyou Liu,Xinping Yang,Yiyao Liu,Jiayi Dong,Yihan Lin,Haokai Ye,Yifan Lu,Zeyu Dong,Yansong Li,Shengqi Zhang
类目: Artificial Intelligence (cs.AI)
备注: 18 pages, 6 figures, 8 tables. Zhiyu Zhou and Tianyou Liu contributed equally

点击查看摘要

Abstract:We developed CAE-Bench, a benchmark to evaluate AI agents on Computer-Aided Engineering (CAE) solver development. A CAE solver is more than a PDE solver; it must be able to generalize across various inputs, such as different boundary conditions, geometries, and material properties. CAE-Bench evaluates the generalizability of AI-developed CAE solvers by providing the same task, such as a heat transfer problem, multiple cases, with each case specifying an input. The AI agent develops a solver for a fixed task using only a few public cases, and the resulting solver is then tested on held-out cases, i.e., cases it has never seen before. We evaluate 12 models across a range of reasoning settings on 80 computational tasks. Our experiments show that even if a solver developed by an AI agent performs well on the public cases, it may even get 0 score on held-out cases. For tasks such as fluid dynamics problems, even the solver developed by GPT-6 Astra performs poorly on held-out cases. A data format and an SDK for task registration are provided, so the researcher can use the benchmark on their own infrastructure. We hope that CAE-Bench will help the community to develop AI agents that can develop reusable CAE solvers that generalize to various inputs.

[AI-31] In-Context Estimation and Prediction for Parametric ODEs

链接: https://arxiv.org/abs/2611.21268
作者: Abdullah Ahmed,Yun Yang
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Numerical Analysis (math.NA); Statistics Theory (math.ST); Machine Learning (stat.ML)
备注:

点击查看摘要

Abstract:Transformers have demonstrated the ability to perform in-context learning, that is, to adapt to new tasks from a few examples without updating their parameters. Existing theory has largely focused on emulating gradient-based algorithms for convex statistical problems or establishing general function approximation guarantees without exploiting the underlying statistical structure. We study in-context parameter estimation and trajectory prediction for parametric ordinary differential equations (ODEs), where estimation is generally nonconvex. Under suitable identifiability and regularity conditions, we construct a transformer that estimates unknown system parameters from noisy trajectory observations through a two-stage approximate maximum likelihood procedure: a coarse search provides an initialization within a locally strongly convex region, followed by gradient descent with local linear contraction. For prediction, we compose this estimator with a neural approximation of the ODE solution map. We further show that, through parameter sharing, a single transformer can perform both parameter estimation and trajectory prediction simultaneously. For these constructions, we establish nonasymptotic guarantees and finite-pretraining oracle inequalities that separate errors due to finite pretraining data, finite context length, and finite-sized transformer approximation. Numerical experiments on linear and nonlinear ODEs further show that pretrained transformers can benefit from longer inference contexts without updating their weights.

[AI-32] Conformal Expert Pools: Training-Free Probabilistic Time-Series Forecasting Beyond Seasonality

链接: https://arxiv.org/abs/2611.21267
作者: Valery Manokhin
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 21 pages, 3 figures, 11 tables. Successor to arXiv:2605.03789

点击查看摘要

Abstract:Conformal Seasonal Pools (CSP) showed that a forecaster with no training at all, built only from past values at the same point in the season and from past forecast errors, significantly outperformed DeepNPTS, Amazon’s learned non-parametric forecaster. We propose Conformal Expert Pools (CEP), a training-free successor. CEP replaces the single seasonal-naive forecast by a pool of simple forecasters, builds its predictive distribution from their past forecast errors, and adapts both the pool and the errors to the series being forecast. No model is trained and no parameter is learned across series. On the CSP benchmark of 232 datasets and 24,502 forecast windows, CEP is better than CSP on 78% of datasets by CRPS, on 84% by MSIS and on 83% by Winkler score; on a typical dataset its CRPS is 14% lower and its MSIS 23% lower (both p 1e-22, paired Wilcoxon test), and coverage of 95% intervals rises from 90.1% to 92.2%. CEP has a lower CRPS than DeepNPTS on 88% of datasets, than NPTS on 89% and than SeasonalNPTS on 83%. The improvement persists on dataset families outside the development set and on 36 additional fev-bench files. We also test the published claims that Laplace, an online ensemble forecaster, “beats (almost) everything” and “beats” CSP. On 228 datasets at multi-step horizons the claims do not hold: CSP outperforms Laplace on 60% of datasets by CRPS and on 64% by interval score, and Laplace ranks sixth of seven methods, with the widest intervals of all (69% wider than CEP’s). Its point forecast is biased low and on 54% of datasets is less accurate than seasonal naive, and from a cold start it takes about 7,700 times longer than CSP. CEP is better than Laplace on 82% of datasets by CRPS and on 80% by MSIS, with intervals about 40% narrower, and runs 271 times faster.

[AI-33] How Developer Personality Shapes the Use of AI Coding Assistants

链接: https://arxiv.org/abs/2611.21240
作者: Annemarie Wittig,Alina Mailach,Norbert Siegmund
类目: Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
备注:

点击查看摘要

Abstract:Developers differ in personality traits that shape how they approach software development tasks, yet AI coding assistants either behave uniformly or offer personalization features whose fit with individual developers must be found through trial and error. This is because it remains unclear how personality shapes developers’ experiences with AI assistants, and thus which traits personalization should account for. In other words, we provide tools with no underlying theory on what works when and why. We take a first step towards such a theory with a systematic scoping review on how users’ personality relates to their perception of LLM-based tools. We identified twelve studies through general search engines, major Software Engineering (SE) venues, AI tooling, and snowballing. They all reported associations between personality and perceptions, particularly perceived usefulness, yet none addressed this in a SE context. Building on these findings, we conducted an exploratory two-wave survey with 29 and 11 developers, collecting first empirical evidence of how developers’ personality (measured using the Big Five Model) relates to the perception of AI coding assistants. Higher Extraversion was associated with more positive private-use experiences in both waves, while Agreeableness was associated with negative professional experiences in the first wave and positive ones in the second. These results provide first indications that the personality effects found in the literature extend to developers. Although further studies with bigger samples are needed to build a theory, our work takes the first step towards an empirical grounded theory of developer personality and AI personalization traits. Subjects: Artificial Intelligence (cs.AI); Software Engineering (cs.SE) Cite as: arXiv:2611.21240 [cs.AI] (or arXiv:2611.21240v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.21240 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-34] Verifiable Random Sampling of Private Data Without a Trusted Single Party

链接: https://arxiv.org/abs/2611.21222
作者: Muhammed Karamuk,Mikołaj Glinka,Ruben Wolff
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: 12 pages

点击查看摘要

Abstract:Problem: When a party supports claims about a private dataset by disclosing a sample, auditors cannot tell whether it was drawn at random or “cherry-picked” to hide flaws in the rest of the data. Solution: The data holder commits to the hashes of every item in its dataset and registers the commitment with independent servers. Selection then waits on a decentralized randomness beacon: a network that jointly emits a fresh random value every few seconds, outside the control of the data holder and the auditors. The value emitted two beacon rounds after the commitment picks the sample. Each server returns a signed receipt, so anyone can later check that the revealed sample is exactly the one that value prescribes. Selection from the committed set stays unbiased as long as the beacon’s threshold holds and one required server is honest, and a pre-agreed name, checked with the servers, stops repeated or abandoned draws. We call this protocol NoCherry. Applications: Any audit that spot-checks a private corpus: private AI benchmarks and held-out test sets, dataset quality claims, code and security audits, test-fairness review, vulnerability handling, and compliance sampling. Auditors need not trust the data holder to hand over a representative sample of the committed data, since the protocol proves it, and data holders need not expose their entire dataset. Comments: 12 pages Subjects: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.21222 [cs.CR] (or arXiv:2611.21222v1 [cs.CR] for this version) https://doi.org/10.48550/arXiv.2611.21222 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-35] InclusionMed: Benchmarking AI for Inclusive Healthcare

链接: https://arxiv.org/abs/2611.21218
作者: TheInclusionMed Collaboration
类目: Artificial Intelligence (cs.AI)
备注: 31 pages, 8 figures, 10 tables. Project website: this https URL

点击查看摘要

Abstract:Advancing medical AI for inclusive healthcare requires evaluation across diverse healthcare roles, tasks, and care settings, but benchmark coverage remains uneven. We introduce InclusionMed, an open medical AI evaluation initiative for inclusive healthcare. The InclusionMed Atlas organizes benchmarks across three healthcare roles and three AI capabilities. For the first release, we selected four public benchmarks from 388 candidates through structured audits and independent reproduction. With InclusionMed-DxFiles, our private diagnostic reasoning benchmark, the Atlas contains five benchmarks covering three of its nine role–capability cells. We evaluated frontier large language models on the Atlas and aggregated their results into the InclusionMed Index, with overall scores ranging from 17.5 to 49.1 out of 100. The Index also reports scores by healthcare role and AI capability, showing that a model’s rank varied with the role or capability considered. A within-benchmark analysis using task-level tags showed that the ten highest-ranked models in the Index all scored higher on patient communication than on clinical decision-making. Models with similar Index scores also differed substantially in estimated inference cost. These findings highlight the need for evaluation covering more diverse healthcare needs and for context-specific model selection. Through open collaboration and community contributions, InclusionMed seeks to ensure that diverse healthcare needs worldwide shape how medical AI progress is defined and measured.

[AI-36] Scientific Coding with Large Language Models : Evidence from Finite Element Analysis

链接: https://arxiv.org/abs/2611.21176
作者: Erfan Hamdi,Joel Shor,Emma Lejeune
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI); Computational Engineering, Finance, and Science (cs.CE)
备注: 28 pages, 7 figures, 14 tables

点击查看摘要

Abstract:Large language models (LLMs) offer a potential route to automating scientific programming, but their usefulness depends on consistently producing numerically correct code. We extend FEM-Bench to 17 finite element tasks involving nonlinear and history-dependent mechanics and structural dynamics, and evaluate 15 models on implementation and test generation without execution feedback. The best model succeeds in 87.1% of single-pass implementation trials, but rankings differ between tracks. The leading implementer achieves 49.9% joint test success, while the leading tester reaches 82.6%. For six models, we compare single-pass generation with two agentic harnesses that allow code execution and revision. Agentic execution solves some tasks that never succeeded in single-pass trials and improves consistency on others, but gains vary across models and harnesses, and regressions also occur. Claude Opus 5 with SimpleAgent reaches 96.5% Pass@3, the probability of at least one success in three attempts, but 77.1% Pass ^3 , the probability that all three succeed. No model-harness combination exceeds 78.8% Pass ^3 . To understand why agents submit incorrect code after passing their own checks, we analyze 249 failed agentic trajectories and identify six categories of self-verification failure. The most common are inadequate test coverage (73.1%) and reliance on generic consistency checks (70.3%), and in 28.5% of cases the tests share an error with the implementation. Motivated by these findings, an exploratory configuration with separate coder and tester agents achieves the highest observed scores, with 97.6% Pass@3 and 88.8% Pass ^3 . These results identify consistency across attempts and independent verification as the main obstacles to trusting LLM-generated scientific code.

[AI-37] Linglong Agent Service: A Runtime for Centralized Agent Reasoning and Distributed Execution

链接: https://arxiv.org/abs/2611.21129
作者: Hua Zhang,Min Jiang
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:LLM agents can use remote tools and cloud sandboxes, but integrating reasoning, session state, and heterogeneous execution sites remains a runtime design problem. We present Linglong Agent Service (agentsvc), a self-hosted HTTP service and embeddable Python runtime that treats execution site as a configurable dimension alongside model, tools, and context. A parent service retains inference orchestration, tool descriptions, and conversation history; each inference request selects one child environment for file, shell, terminal, and MCP operations. Direct HTTP and child-initiated WebSocket tunnels carry the same application interfaces, including portable request context and centrally managed updates. The normal core path requires no third-party Python packages, with an optional Windows pseudo-terminal dependency. We further present online_debugger as the process-scale special case of execution-site separation: an instrumented program submits exception evidence, executes session-scoped inspection tools inside the incident process, and resumes the same diagnostic session. Its default prompt requires failing-then-passing reproduction evidence for source changes, while runtime patch activation and failed-call recovery remain unimplemented. We explain the design through constrained expected-utility reasoning, compare concrete mechanisms with Hermes Agent, Pi, Codex, and Claude Code, and report implementation and regression-test evidence. The current evidence establishes executable contracts, not performance advantages, repair accuracy, or production recovery. The source code is open source under the MIT license at this https URL.

[AI-38] Searching for Autoregressive Neural Architectures from Fundamental Operations

链接: https://arxiv.org/abs/2611.21111
作者: Thomas Woehrle,Linus Ericsson,Henry Gouk,Elliot J. Crowley
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Expressive Neural Architecture Search (NAS) search spaces are desirable as they do not restrict the space of potential architectures to what is already established. However, few search spaces are expressive enough to represent a broad range of architecture families, and those that are have focused primarily on image classification problems. The task of autoregressive sequence modelling has become ubiquitous in modern machine learning, but no existing search space supports the construction of and search over sequence models from fundamental operations, instead imposing strong inductive biases on the structure of potential architectures. This work describes how the expressive einspace search space can be extended to yield a search space capable of representing autoregressive models using fundamental operations. We perform a series of experiments showing that the novel einspaceAR search space can find capable autoregressive sequence models from scratch–i.e., without imposing strong inductive biases. We believe that our einspaceAR search space will be a useful part of the NAS toolbox and that the techniques described in this work can serve as a blueprint for the integration of autoregressive modelling in other expressive search spaces.

[AI-39] Inference Equivalence and Quotient Coding

链接: https://arxiv.org/abs/2611.21096
作者: Zibo Meng,Andrew C. M. Austin,Alin Achim
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Incomplete observations can admit multiple physical interpretations without making every task conclusion ambiguous. Recovering all compatible scenes is often unnecessary, whereas indiscriminate compression may change which claims are justified. We ask which distinctions among observation-compatible interpretation sets must be retained to preserve every valid conclusion of a specified task. We first define inference validity as truth across all physically compatible states and characterize it, under a closedness condition, by distances to physically realizable falsifying measurements. For a finite semantic claim family ordered by specificity, the Exact Inference Frontier forms the unique minimal antichain generating all valid claims. We then define inference equivalence between compatibility sets and construct a canonical quotient code from their frontiers. Among deterministic codes admitting exact inference decoding, it induces the coarsest partition, minimizing the number of distinguishable code values rather than storage or computation. Finally, we characterize inference errors from learned compatibility sets, derive conditions for exact inference preservation despite set errors, and evaluate an asymmetric learning surrogate under distribution shifts. On a registered finite bank, 145 compatibility sets yield 31 quotient codes with unchanged valid-claim sets. Across geometry, mechanism, and joint shifts, learned estimates exhibit different accuracy-specificity trade-offs. On 1,962 official HRSID test images, 1,913 distinct candidate sets induce only 36 seven-claim inference signatures, while a visual 3-NN baseline achieves 87.31% precision on issued claims.

[AI-40] Conflict-Aware Multi-Teacher On-Policy Distillation in Aligned Action Spaces for Cross-Game Generalization

链接: https://arxiv.org/abs/2611.21094
作者: Jing Yi,Cehao Yang,Hualei Zhang,Haitao Li
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Zero-shot cross-game generalization aims to consolidate accessible source-game experts into a single vision-language model (VLM) that controls unseen games from raw pixels without target-game parameter updates. While multi-teacher on-policy distillation (MOPD) mitigates offline exposure bias and single-game reinforcement learning (RL) overfitting, scaling it to autoregressive VLMs faces two bottlenecks. First, token-to-action misalignment from subword length bias distorts discrete action probabilities and uncertainty margins. Second, multi-teacher policy conflict arises as source teachers under incompatible dynamics issue contradictory closed-loop corrections, causing negative transfer. We propose CAMOPD (Conflict-Aware Multi-Teacher On-Policy Distillation), a post-training framework that resolves multi-teacher policy conflict within an aligned semantic action space. CAMOPD first uses complete-action alignment to normalize command completion likelihoods over each game’s legal action set, forming a calibrated categorical policy that provides reliable uncertainty margins for selective on-policy retention and visual specialist probes that can be evaluated across source games. To resolve policy conflict without target-game data, conflict-aware multi-teacher consolidation cross-evaluates these specialists across source games to capture both visual identifiability and inter-teacher compatibility, upweighting compatible teachers while suppressing conflicting corrections. On GameWorld’s temporal and directional transfer suites with Qwen3.5-4B, CAMOPD achieves 26/30 (86.7%) and 24/30 (80.0%) Task-1 zero-shot success on unseen GeoDash and Doodle Jump, outperforming prompting, weight-merging, and multi-task baselines.

[AI-41] hink2Listen: Evidence-Centric Learning For Agent ic Audio Understanding With LALMS

链接: https://arxiv.org/abs/2611.21083
作者: Jin Wang,Liang Lin,Kaiwen Luo,Weiliu Wang,Yuxuan Li,Zengxin Han,Gan Zhang,Jingtong Wu,Yufei Guo,Kun Wang
类目: ound (cs.SD); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Recent advances in large audio-language models (LALMs) have broadened audio understanding across speech, sound, and music, with tool-based agents further supporting these tasks by invoking predefined audio tools. However, a persistent gap remains between perceiving acoustic content and using it as evidence for reasoning: recognizing what is audible does not ensure that a model can identify which cues resolve an uncertain decision. To bridge this gap, we propose Think2Listen, an evidence-centric learning framework for agentic audio understanding. Guided by evidence before commitment, Think2Listen couples iterative reasoning with a targeted LALM interaction to actively seek acoustic cues that distinguish competing hypotheses and revise candidate judgments based on the observations returned. It learns what to ask, how to use evidence, and when to stop, using observation-induced changes in candidate-state to inform training credit. Extensive evaluations on two benchmarks spanning seven audio categories demonstrate that Think2Listen achieves (I) enhanced multi-step audio reasoning, with gains in overall accuracy of 5.08%–17.67% over competing audio agents; and (II) improved audio understanding and reasoning across speech, sound, and music, boosting overall accuracy by 2.29%–13.10% over the corresponding baselines. The source code is available at this https URL.

[AI-42] SemiNovo: Learning Beyond Search-Identified Spectra for De Novo Peptide Sequencing KDD2027

链接: https://arxiv.org/abs/2611.21079
作者: Pan Liu,Dongxin Lyu,Yusen Tan,Yifan Li,Hai-tao Yu,Yunhua Zhong,Zikun Wang,Tianze Ling,Jun Xia
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 13 pages. Accepted as an oral presentation at KDD 2027, AI for Sciences Track

点击查看摘要

Abstract:Sensitive peptide identification is fundamental to protein characterization and novel protein discovery, for which de novo peptide sequencing (DNPS) from tandem mass spectra provides an important database-independent approach. Although deep learning has substantially advanced de novo sequencing, existing models are predominantly trained on high-confidence peptide-spectrum matches obtained through database search. This search-conditioned supervision excludes a large fraction of unidentified spectra, creating a distribution mismatch between the selectively labeled training data and the complete spectrum stream encountered at inference time, thereby limiting peptide identification performance. To address this challenge, we propose SemiNovo, a semi-supervised framework that learns jointly from identified and unlabeled spectra. SemiNovo first constructs DarkSpec, a large-scale quality-controlled corpus of unlabeled spectra collected from diverse public proteomics projects, thereby extending training beyond search-identified spectra and mitigating the resulting shift in DNPS supervision. It then combines stepwise mass-constrained autoregressive decoding with cumulative path confidence to derive reliable supervision from unlabeled spectra. Moreover, SemiNovo introduces head-wise chemical-evidence gating to improve spectral evidence aggregation under noisy and incomplete fragmentation. Experiments show that SemiNovo performs favorably against state-of-the-art methods on multiple benchmarks. Code and dataset are publicly available.

[AI-43] Counting What Matters: Reducing Redundant Object Detections in Overlapping UAV Mapping Imagery for Disaster Response

链接: https://arxiv.org/abs/2611.21076
作者: Max Schulte,Hartmut Surmann,Julien Meine
类目: Artificial Intelligence (cs.AI)
备注: 9 pages, 6 figure, IEEE International Symposium on Safety, Security, and Rescue Robotics, SSRR 2026, Incheon, Korea, November 02-04, 2026

点击查看摘要

Abstract:UAV mapping flights enable rapid aerial reconnaissance of large-scale incident sites in civil protection. However, the image overlap required for orthophoto generation causes every physical object to be detected multiple times across images, while dynamic objects may be omitted from or duplicated in the resulting orthomosaic. Analysis based on the orthophoto alone is therefore structurally insufficient. We address this by detecting objects on the individual images and merging redundant detections across them. This task lies between multi-object tracking and re-identification, as mapping flights form temporally ordered but sparsely sampled sequences with object reappearances across flight strips. We fine-tune YOLO detectors on a new domain-specific nadir dataset from real civil protection operations, roughly doubling mAP50 over the previous system, and propose re-identification approaches based on 3D positions from sparse reconstruction, DINOv3 appearance embeddings, and their combination via constrained greedy clustering. %On four manually annotated real-world mapping flights, the combined weighted multi-pass approach achieves the best trade-off between association accuracy and count error. On four manually annotated real-world mapping flights, an association-only evaluation using ground-truth detections shows that the combined weighted multi-pass approach provides the best trade-off between association quality and count error. It is integrated into the open-source web application ARGUS and visualized as verifiable, correctable object groups for first responders. this https URL

[AI-44] Sage S2.1 Flash Technical Report

链接: https://arxiv.org/abs/2611.21072
作者: Kun Cai,Haodong Chen,Guangran Cheng,Qingxiu Cheng,Xu Dai,Guanting Dong,Qiming Ge,Lixin Gu,Tao Gui,Honglin Guo,Yuyang Hu,Zixian Huang,Jiajie Jin,Alexander Lam,Jiaqiang Li,Yining Li,Jiahang Lin,Yanjiang Liu,Xinyu Lu,Haijun Lv,Ningsheng Ma,Zerun Ma,Minlong Peng,Jianmin Qian,Huayu Sha,Junlin Shang,Shuo Shen,Qisheng Su,Jiajun Sun,Yanan Sun,Zhichao Sun,Ran Tao,Zhongbo Tian,Guoqiang Wang,Hui Wang,Rui Wang,Yiding Wang,Zhecan Wang,Hao Xiang,Xinchen Xie,Shuhao Xing,Xiaoyu Xing,Wanghan Xu,Xinyu Yang,Yajie Yang,Dingbo Yuan,Yuhang Zang,Hongjie Zhang,Chengfeng Zhao,Haoran Zhao,Penghao Zhao,Ruojun Zhou,Dongsheng Zhu,Yicheng Zou,Qiye Cai,Yifei Cao,Xinmeng Che,Hongji Chen,Jiabei Chen,Jiadong Chen,Jiahao Chen,Jiayi Chen,Xinjie Chen,Yujia Chen,Xin Deng,Yi Dong,Yichen Dong,Wenjie Fu,Jianing Geng,Chenyang Gu,Ding Han,Han Han,Feiyang Hao,Haotan He,Qiushi He,Xufeng He,Xin Hong,Jie Hou,Binze Hu,Yongkang Hu,Zijian Hu,Enfu Huang,Junhao Huang,Changhao Jiang,Huicheng Jiang,Jiazhen Jiang,Shufan Jiang,Wenqing Jing,Jiahao Kuang,Bowen Lai,Bo Li,Qi Li,Qilong Li,Zhuoqun Li,Tianhao Liang,Jiaxiang Liu,Junzhang Liu,Shaofan Liu,Shuainan Liu,Tong Liu
类目: Artificial Intelligence (cs.AI)
备注: website: this https URL

点击查看摘要

Abstract:We introduce Sage S1.5 Flash, a 320-billion-parameter general-purpose mixture-of-experts model designed to reason over scientific evidence and act in interactive environments. Sage Sci-Perception uses dedicated encoders for molecular graphs and biological sequences and represents spectra as ordered patches of measurements. Modality-specific projectors map these representations into the input space of a shared backbone, which processes them alongside text and images. To develop the model’s agent capabilities, we construct verifiable environments for tool use, search, and coding, as well as specialized workflows involving machine learning experiments and scientific software. Task completion is checked against execution results, changes in environment state, or submitted artifacts. We evaluate Sage S1.5 Flash on 33 benchmarks across seven domains. In the primary comparison against five baselines, it ranks first on OfficeQA Pro (72.2), MLE-bench Lite (90.3 HumanRank), Harvey LAB, and SpreadsheetBench~2, and second on OSWorld-Verified (85.2), WebArena-Verified (67.6), and DeepResearch Bench~II (54.2). These results show particular strengths in using documentary evidence, producing verifiable workbooks, and carrying out machine learning tasks, alongside competitive performance in desktop and browser workflows that require the model to act on visual feedback.

[AI-45] Multi-CONDOR: Learning Multistable Motion Primitives from Demonstration

链接: https://arxiv.org/abs/2611.21059
作者: Gabriele Giuseppe Di Marzo,Mariano Ramírez Montero,Maximilian Stölzle,Rodrigo Pérez-Dattari,Cosimo Della Santina
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Dynamical-system-based motion policies provide a framework for learning reactive robot motions with stability and convergence properties. Many manipulation tasks, however, admit several distinct goals or a continuous set of acceptable outcomes, requiring motion primitives that preserve these alternatives. Accommodating both goal structures within a unified framework with continuous, autonomous dynamics remains an open challenge. We propose a CONDOR-based framework, multi-CONDOR, for learning motion primitives with either multiple discrete attractors or a continuous attracting goal contour. The approach combines imitation learning with losses that encourage agreement between the learned task dynamics and latent reference systems encoding the desired attraction structure. We provide a Lyapunov-based analysis of the convergence properties of the learned motion dynamics. We evaluate the framework on simulated benchmarks with discrete goals, qualitative examples with continuous goal contours, and robotic grasping experiments. In simulation, multi-CONDOR reduces trajectory RMSE by approximately 72% compared with separately learned single-goal motion primitives selected by a state machine. For tasks with two and three goals, respectively, it reduces the fraction of evaluated initial states ending outside the goal tolerance from StableNODE’s 0.31% and 7.29% to 0% in both settings. Experiments with a Franka Emika robot arm demonstrate successful grasping of a container by either of its two handles using motions learned from demonstrations. Together, these results show multi-CONDOR combines motion reproduction with the flexibility to converge to alternative valid outcomes.

[AI-46] System 1 agents : a method for local reliable and fast agent ic AI

链接: https://arxiv.org/abs/2611.21053
作者: Eleftherios Tzanis,Michail E. Klontzas
类目: Artificial Intelligence (cs.AI)
备注: 18 pages, 5 figures, 9 tables. Code and data: this https URL

点击查看摘要

Abstract:Agentic artificial intelligence could automate analyses of sensitive clinical data, but current agents depend on large language models (LLMs). Cloud LLMs expose patient data. Local LLMs hallucinate, and an invented value in a tool call propagates through every later step. Here we present a method to build System 1 agents: local agents that call tools without generating text. A deterministic compiler turns each tool’s schema and the request into multiple-choice questions whose options come from the request and the schema, then a one-pass decision model answers them. On 1,952 requests to 157 biomedical tools, the best System 1 agent chose the right tool, or correctly declined, for 97.0% of requests, significantly ahead of all 18 local LLM-agent settings (best 95.2%). The System 1 agent filled 97.0% of 6,056 stated argument values. No System 1 call contained invented content, against 5.5-19.7% of LLM calls. The System 1 Agents framework is freely available as open-source software at this https URL.

[AI-47] SWE-OPD: On-Policy Distillation for Agent ic Coding

链接: https://arxiv.org/abs/2611.21020
作者: Yuhang He,Yanyu Chen,Chuanzhe Guo,Jie Zhu,Zichen Tian,Hongyu Ge,Zixuan Ren,Huang He,Bingjin Chen,Junyuan Shang,Shuohuan Wang,Yongqi Zhang,Irwin King,Jing Liu
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: under review

点击查看摘要

Abstract:In this work, we study on-policy distillation (OPD) for transferring agentic coding capabilities from a large teacher model to a smaller student model. Prior work has focused primarily on reasoning and short-horizon agentic tasks. We present SWE-OPD, a minimal OPD study for software engineering, where the student interacts with executable repositories and the teacher supervises the prefixes visited by the student. Across three teacher–student pairs, two training sources, and three SWE benchmarks, we find that on-policy supervision transfers coding capability: supervision from Qwen3.6-27B increases the SWE-bench Pro resolve rate of Qwen3.5-9B from 35.9% to 39.4%. Refreshing trajectories is not strictly required either, as an offline variant with cached teacher scores still improves over the student baselines while cutting end-to-end GPU-hours by 59.4%–72.6% and halving peak GPU allocation, at benchmark-dependent accuracy. Furthermore, we extend the offline variant to corss-tokenizer scenarios: distilling GLM-4.5-Air from DeepSeek-V4-Flash-0731 raises its SWE-bench Pro and Multilingual resolve rate from 35.8% to 39.7% and from 41.8% to 46.2%, around 4% and 4.6% improvement, respectively. We also find that a decreasing loss, however, does not necessarily indicate improved agent performance: under top- K reverse-KL loss, one student exhibits severe repetition despite a decreasing loss. We identify a support-mass gradient term as a potential mechanism underlying this failure.

[AI-48] K2-Horizon: A Frontier Fleet of Open Source Language Models

链接: https://arxiv.org/abs/2611.21003
作者: K2 Team:Zhengzhong Liu,Haonan Li,Xuezhe Ma,Alexander Moreno,Subham Sekhar Sahoo,Rupesh Kumar Srivastava,Mikhail Yurochkin,Hadi Abdine,Yash Akhauri,Adharsh Babu,David Bellamy,Jalaj Bhandari,Dani Bouch,Wail Bouhedja,Weili Cao,Ahmad Chamma,Po-Chun Chang,Junlin Chen,Lingjie Chen,Zhoujun Cheng,Kyle Chickering,Ryandito Diandaru,Evan Dufraisse,Chaitanya Dwivedi,Bilal El Bouardi,Desai Fan,Richard Fan,Chengqian Gao,Rifo Ahmad Genadi,Jonathan Geuter,Charles Guille-Escuret,Willis Guo,Dhawal Gupta,Seungwook Han,Tingxu Han,Shibo Hao,Besher Hassan,Guowei He,Ming Shan Hee,Joel Hestness,Yu Huang,Yuan Huang,Haolong Jia,Linghao Jin,Junu Kim,Lingjing Kong,Nurkhan Laiyk,Andrej Leban,Liqun Ma,John Maggs,Amr Mohamed,Mukhesh Narra,Omkar Pangarkar,Aldo Pareja,Khiem Pham,Varad Pimpalkhute,Nikhil Ranjan,Elizaveta Rebrova,Hector Ren,Shaurya Rohatgi,Guokan Shang,Jianshu She,Chufan Shi,Aaryamonvikram Singh,Daria Soboleva,Kartik Sreenivasan,Suqi Sun,Yuekai Sun,Bowen Tan,Liping Tang,Shengkun Tang,Muzi Tao,Mallikarjuna Tupakula,Micah Villmow,Hongyi Wang,Huijuan Wang,Renxi Wang,Yongxin Wang,Yuqi Wang,Zhenting Wang,Shicheng Wen,Yingquan Wu,Anze Xie,Andrew Xing,Matthew Yang,Shuanghai Yu,Kerem Zaman,Zenan Zhai,Hongli Zhan,Xuezhou Zhang,Anshun Asher Zheng,Zhe Zhou,Richard Zhuang,Eric P. Xing
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:We present K2-Horizon, a fleet of six fully open language models, from a 0.9B dense model to a 375B-A23B MoE, built for general-purpose reasoning and agentic work across math and science, software engineering, deep research, office automation, and beyond. The models lead or are competitive with strong open models of their size on agentic and coding benchmarks. On the modeling side, to strengthen reasoning, we introduce Mixture-of-Value Attention (MoVA), applying MoE to attention value projections, and softplus attention gating, a more flexible, theoretically motivated gate. Our 36B-A4B model uses both and nearly matches a 32B dense model trained on the same data, with one-seventh the active parameters. We also introduce Uno, lossless diffusion adapters that speed up decoding by up to 2.2x without a separate draft model. Our release follows the LLM360 definition of open source for transparency and reproducibility: the full development history, including intermediate checkpoints from every training stage, pre-training data, training logs, and all training, evaluation, and infrastructure code, notably our production-scale xLLM framework for dense and MoE training and the RL360 reinforcement-learning stack. Beyond the data itself, we release and document in detail every pipeline that built it, the part of model development that open releases most rarely disclose. These pipelines produced roughly 10T synthetic pre-training tokens, showing that synthetic data can be a major ingredient of pre-training at scale. Four of the models share one 22T-token pre-training corpus and recipe, revealing consistent scaling behavior and enabling controlled studies of scale and sparsity. All artifacts are available at this https URL

[AI-49] Explaining Cross-Label Decision Structure in Multi-Output Tree Ensembles

链接: https://arxiv.org/abs/2611.20996
作者: Sylvio Barbon Junior,Hiago F. F. Oliveira,Rafael Figueira Gonçalves,Jessica Fernandes Lopes,Ricardo Cerri
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 53 pages, 22 figures, 11 tables

点击查看摘要

Abstract:A multi-output tree ensemble predicts several binary labels through decision paths that it largely shares across those labels, yet the attribution methods commonly used to explain such models treat each output in isolation. We propose multi-sink Decision Predicate Graphs (ms-DPGs), which extend decision predicate graphs with one sink per label: each leaf contributes to the sink of every label its in-bag positive count for that label. On this representation we define Shared Predicate Flow (SPF), the share of a label’s decision mass that passes through conditions active for both labels of a pair, and a contracted view of the graph in which label-specific lifts indicate whether a shared condition raises the probability of both labels or of one at the expense of the other. Controlled experiments show that SPF captures shared decision conditions rather than mere feature overlap. On nine multi-label benchmarks, native multi-output forests share substantially more decision structure than Binary Relevance on eight, including five of the six on which the two are at predictive parity, even at equal hyperparameters, whereas per-label TreeSHAP rankings register the change of architecture without showing whether more is shared. Because a native forest routes every label through the same splits, SPF describes the fitted model rather than the data, and on eight of the nine benchmarks it rises when the feature-label signal is removed; a high SPF is therefore not evidence of shared predictive information. ms-DPG offers a global, predicate-level account of how a multi-output ensemble reuses decision conditions across its labels.

[AI-50] GRACE: Parallel Best-of-N Multi-Agent Analysis of Closed-Form Questions over Dataframes with Guardrail Selection and Adaptive Retry

链接: https://arxiv.org/abs/2611.20983
作者: Fabio Picariello,Salvatore Cipolla,Susanna Bonura,Vito Morreale
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Accepted at the 5th Italian Conference on Big Data and Data Science (ITADATA 2026). To appear in CEUR Workshop Proceedings

点击查看摘要

Abstract:LLM agents answer closed-form questions over a tabular dataframe by writing and executing Python, almost always as a single rollout: one interpretation, one script, one result, accepted without comparison. The few pipelines that produce several candidates reconcile them by answer-agreement voting, which needs a single comparable scalar and always returns one, even when every candidate is wrong. We present GRACE (Generative, Ranked, Adaptive Code Ensemble), a production-oriented pipeline that instead runs parallel best-of-N selection specialised to dataframe code. N agents run concurrently under a strict, machine-checkable output contract, diversified by persona hints, per-retry rotation, and a two-regime temperature schedule. Rather than vote, GRACE judges the executed artefacts: a pairwise tournament pre-filters the pool and a domain-specialised guardrail either selects the best candidate or, when none is defensible, rejects the whole pool and abstains with a diagnosis (an option voting structurally lacks), which in turn drives an adaptive retry with per-agent feedback. On InfiAgent-DABench (DAEval, M = 246), using Qwen3.6-35B-A3B as the backbone, GRACE raises answer match from 39% for a single-rollout baseline to 71% (+32 pp, 1.8x), recovering 53% of previously wrong answers. On the same DAEval task, the smaller dense Qwen3.5-9B lifts answer match from 38% to 63% (+25 pp, 1.7x). A controlled 3 x 3 study that sweeps the number of parallel agents against the retry budget finds, on the 35B backbone, the two levers largely substitutable: once either one is generous, the other adds little, identifying a cheap operating point for deployment. The same directional pattern replicates on the independent DataBench benchmark, again on the Qwen3.6-35B-A3B backbone, where answer match rises by 15 percentage points (from 70% to 85%) across 400 sampled questions.

[AI-51] Breaking feature collapse in self-supervised time series encoders

链接: https://arxiv.org/abs/2611.20981
作者: Özgün Turgut,Daniel Rueckert
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:General-purpose time series encoders promise a single backbone that generalises across tasks and domains; however, they currently do not scale. We trace this down to a critical point of failure: under the predominant masked data modelling supervision, the encoder’s output tokens partially collapse. This is the natural consequence of supervising encoders only in the input space, which leaves no constraint against redundant token dimensions in the feature space. We introduce OTISv2, which suppresses the collapse by complementing the reconstruction objective in input space with two objectives in the feature space: a self-distillation loss against an exponential-moving-average teacher, and a KoLeo regulariser. Under matched training, the resulting encoder reduces the cosine similarity and lifts the effective rank of its output tokens relative to its predecessor, and it scales monotonically from 7.1 M to 116 M parameters where its predecessor declines. It matches or surpasses existing general-purpose time series encoders across 143 uni- and multi-variate datasets. To foster further research, we release our code and pre-trained weights.

[AI-52] Hilbert States for Decision-Making under Partial Observability

链接: https://arxiv.org/abs/2611.20970
作者: Zibo Meng,Andrew C. M. Austin,Alin Achim
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Decision-making under partial observability can be fundamentally limited by information, even with known dynamics: distinct physical states can induce the same observation history while favoring different actions. We study three questions: when is full-state optimality unattainable, what information exactly preserves the best observation-based decision, and how does finite-data learning affect reliability? First, we characterize decision ambiguity through conditional action preferences. We give a zero-regret criterion, a positive lower bound for conflicting uniquely optimal actions on positive-probability regions with uniformly positive cost gaps, and an exact binary-action regret formula. This also shows why physical-state ambiguity alone need not cause decision loss. Second, for two actions and squared terminal-goal costs in the plane, we derive an exact inner-product representation of conditional expected cost differences using three conditional moment contrasts and a goal-dependent feature. This Hilbert state preserves observation-optimal actions without hidden-state reconstruction. Under nondegenerate history and goal variation, its three-dimensional form is minimal among exact bilinear cost-contrast representations. Third, we relate finite-data state errors to excess decision regret via goal-projected errors, action margins, and goal geometry. Physics-based simulations verify information-limited regret, exact action preservation, and the role of second moments. Held-out real-SAR experiments demonstrate feasible goal-conditioned decisions from masked images, while finite-data evaluations reveal goal-dependent reliability. Together, these results separate irreducible information loss, exact representation sufficiency, and learning-induced excess regret.

[AI-53] Razavi-Bench: Tracking AI Progress in Analog Circuit Reasoning

链接: https://arxiv.org/abs/2611.20935
作者: Zhishuai Zhang,Oscar Toomey,Taekwang Jang
类目: Artificial Intelligence (cs.AI)
备注: 6 pages, 3 figures, 2 tables. Project page: this https URL ; code and data: this https URL

点击查看摘要

Abstract:We propose Razavi-Bench, a benchmark built from the 50 questions in Razavi’s two Analog Design Experiments With AI articles and graded by GPT-6 Luna against reference answers and rubrics, and use it to track how fast AI models improve at analog circuit analysis. Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33% in 30 months, mostly before the reference answers were published. The leading models now answer almost every textbook question correctly; the hardest questions show a circuit that resembles a textbook one but works differently, or ask about behavior the obvious analysis misses. Within two and a half years, a set of questions that exposed the limits of AI models has become one they answer almost perfectly.

[AI-54] Multi-LLM Agent Agent ic AI at the Edge: Joint Agent Scheduling and Resource Allocation

链接: https://arxiv.org/abs/2611.20915
作者: Yuanyan Song,Kezhi Wang,Xinmian Xu,Yongmin Li
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:In this paper, an edge-enabled multi-large language model (LLM) agentic AI architecture is proposed, where one supervisor AI agent and multiple worker AI agents cooperatively execute users’ requests at the edge. The supervisor AI agent decouples the user request into a series of sub-tasks and dynamically dispatches them to LLM-equipped worker AI agents via two agentic workflows, i.e., parallel and serial workflows, with a two-level retry mechanism to enhance execution reliability. To prolong the service duration of worker AI agents, the energy minimization problems are formulated by jointly considering AI agent selection and resource allocation, subject to task execution latency and quality of service (QoS) constraints. A resource allocation scheme based on alternating optimization (AO) is proposed. Specifically, for the parallel workflow, the scheme decouples the original problem into bandwidth allocation, retrieval selection and AI agent selection sub-problems, which are iteratively solved by the proposed successive convex approximation (SCA), Lagrangian relaxation (LR), and Gale-Shapley (GS) or LLM-based algorithms, respectively. For the serial workflow, the problem is similarly decoupled into service time allocation, retrieval selection and AI agent selection sub-problems; the optimal service time allocation is obtained via Karush-Kuhn-Tucker (KKT) conditions, while the latter two sub-problems are addressed by extending the proposed algorithms to the serial scenario. Simulation results show that the proposed resource allocation scheme can reduce system energy consumption compared with other benchmarks, while experiments on a multi-LLM agentic AI system based on Qwen and Llama demonstrate improved task execution success rates.

[AI-55] Auditing a Moving Target: Evidence Anchoring and the Expiry of Fairness Audits of Provider-Served Language Models

链接: https://arxiv.org/abs/2611.20901
作者: Mika Facchini(Technical University of Munich)
类目: Computers and Society (cs.CY); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Organizations increasingly use large language models (LLMs) for decisions about people, such as hiring, investment and employee evaluation, and rely on them over long periods of time. Fairness audits of such systems, however, are conducted only once. We examine how the biases of LLMs change over time. In a venture screening setting, models from five providers decided whether and how much to invest in startups whose founders differed only in gender, origin and university. We repeated this audit three times over 13 months (248,400 API calls). Study 1 shows that the same model, called under the same identifier one year later, displayed a significantly different bias pattern. Study 2 shows that the successors of retired models carried the same bias as their predecessors, but in a significantly different size and form. Study 3 shows that the bias does not depend on the industry of the startup. A document analysis further shows that providers change and retire models at their discretion, with notice periods of zero to six months. Drawing on these findings, we develop a theory of evidence anchoring. We theorize that bias resides in the model class and persists, whereas audit evidence is anchored to a single model and expires when the provider decides. We derive policy advice for organizations and regulators, namely audits of model classes instead of single models, a version record for every model response, and re-audits triggered by model changes.

[AI-56] Which AI Use Cases Deserve a Pilot? A Risk and Value Framework for Funding Scoping and Scaling AI

链接: https://arxiv.org/abs/2611.20894
作者: Vishal Srivastava
类目: Artificial Intelligence (cs.AI)
备注: 18 pages. All results use synthetic data. A companion Python package reproduces the tables

点击查看摘要

Abstract:A composite score can rank AI use cases while leaving the permitted operating scope and the evidence required for expansion unspecified. We treat the unit of decision as a use case at a proposed scope (authority, human mediation, limits, volume, environment, population) and define three decisions: screen the scope, fund the experiment, and authorize expansion only when residual risk is acceptable and evidence of incremental net value applies at that scope. A risk track (R1-R4) is implemented as an explicit decision table with vetoes and documented precedence, and a value track follows four evidence stages with each transition needing its own evidence. Funding uses expected value of perfect information as an upper bound and the value of sample information for the specific experiment. A quantified expected-loss check, required where policy says so, computes the containment an expansion needs and credits a control only on evidence measured at the proposed scope; in a synthetic scenario the highest-value control design fails that test while a smaller scope passes. A reproducible Python package with 279 passing tests demonstrates the framework on synthetic refund, reconciliation and knowledge-assistant use cases, including a synthetic matched-scope fixture and a transferability stress test (value evidence is simulated, not realized), and compares it with a public AWS sample. Thresholds and appetite are prototype policy requiring field validation.

[AI-57] Accelerate Bench: Evaluating Agents for Application-Level Program Optimization

链接: https://arxiv.org/abs/2611.20882
作者: Bowen Liu,Yuhan Li,Nuo Chen,Weile Luo,Miao Peng,Qintong Zhang,Qifan Zhang,Jia Li,Ziqi Gao,Lydell Huang
类目: Artificial Intelligence (cs.AI)
备注: 49 pages, 8 figures. Work by Bowen Liu was done during an internship at Tencent Hy Team. Yuhan Li is Project Lead; Nuo Chen is corresponding author (chennuo26@gmail.com). Project page: this https URL

点击查看摘要

Abstract:Application optimization requires coding agents to discover useful changes and coordinate them across a complete program while preserving the required computation. We introduce Accelerate Bench, a benchmark of 32 CPU optimization tasks adapted from scientific applications and supercomputing competitions. Each task provides a working program and specifies the workload, correctness requirements, timing boundaries, and permitted edits, leaving agents to choose where and how to optimize. Public testing supports iterative development, and a separate verifier scores the final submission. On benchmark version 1.1, we evaluate 16 coding agents, each pairing a model with a harness. Their Score@1 ranges from 25.5% to 67.4%. Most of their programs are correct, and the size of the speedup separates them: the two leading agents are at least 2x faster on 22 or more of the 32 tasks, while the other fourteen reach 2x on 8 to 17 and leave 125 of their 412 correct outcomes below 1.25x. Controlled replays show why coordination across an application matters: identical source code runs between 0.29x and 2.12x depending only on build and runtime settings, and one code change is worth 1.01x or 6.4x depending on another. Accelerate Bench pairs executable workloads with agents’ programs and trajectories to evaluate both their gains and the process behind them.

[AI-58] Re-Reading Is Not Free: Token Traces and Cache-Aware Energy Estimation for LLM Agents in Software Engineering

链接: https://arxiv.org/abs/2611.20873
作者: Tina Vartziotis,Ketevan Peranidze,Rodopi Kosteli,Ezgi Sarikayak,George Dasoulas,Michael Keckeisen,Sotirios Kotsopoulos,Francesca Dominici
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 21 pages, 13 figures, 18 tables

点击查看摘要

Abstract:As large language models (LLMs) evolve from conversational assistants into autonomous agents embedded in software development workflows, estimating their environmental footprint requires moving beyond conventional token-based accounting. Existing energy estimators focus on isolated request-response interactions, overlooking the repeated processing of accumulated context, cache reuse, and work delegated to subagents. We analyze nearly one billion tokens collected across 42 sessions and 752 task episodes over ten weeks of real-world AI-assisted software development. Cached-context reads make up 94.8% of all API-reported tokens, while uncached input accounts for only 0.017%; for every generated token, approximately 260 cached tokens are read. Context grows 5.3-fold over a session compared with 1.4-fold within individual tasks. To quantify the energy implications, we extend an existing analytical inference-energy model to account for cached-context processing, yielding a cache-aware hybrid estimator. It estimates approximately 250 kWh for the workload, compared with approximately 10 kWh from the output-based estimator EcoLogits, a 25-fold difference largely driven by cached-context reads. Neither estimator has been validated on long cached contexts, but no modeling assumption we varied closes the gap. Starting each task from a fresh context would lower our estimate by about 60%. We also find that parent-session logs report only a subagent’s final internal call, undercounting delegated tokens by a factor of ten; for this workload the correction changes the energy estimate by about 5%, but token-volume studies are fully exposed to it. Our findings expose a substantial gap between how current energy estimators and agentic workloads treat context, highlighting the need for workflow-level energy accounting that captures context accumulation, cache reuse, and delegated execution.

[AI-59] Scenario-Based Risk Certificates for Multi-Property Vulnerability Discovery in Parametric MDPs

链接: https://arxiv.org/abs/2611.20861
作者: Zhiming Chi,Andrea Turrini,Yong Li,Lijun Zhang
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: VMCAI 2027

点击查看摘要

Abstract:Safety-critical systems such as network protocols and distributed algorithms depend on environmental parameters whose exact values are uncertain. Parametric Markov decision processes (pMDPs) model such systems by making transition probabilities symbolic functions of the parameters. For a single quantitative requirement, parametric verification asks which parameter valuations satisfy the requirement. In practice, however, engineers often monitor several safety and performance requirements at once, and one valuation may make several requirements vulnerable simultaneously. In this paper we study multi-property vulnerability discovery under requirement-wise semantics: given a pMDP and M quantitative requirements, we construct candidate regions of the parameter space that may contain violations, and study budgets for random testing under a specified parameter distribution. To solve this, we adopt a two-layered approach in which the scenario-approximation layer provides PAC guarantees on identifying conservative vulnerability regions for each requirement while the testing layer treats random parameter exploration as coverage of overlapping regions and yields a stopping certificate computed from the marginal region masses alone; a composition theorem then bounds the expected non-discovery score by the scenario approximation error plus the testing risk from the two layers. For confirmed discoveries, candidate hits are checked by model checking and budgets use independently certified discovery probabilities. Experiment evaluate these certificates using empirical surrogates and independent confidence bounds.

[AI-60] ProcReverser: Multi-Agent Black-Box Microarchitectural Reverse Engineering for Processor Vulnerability Analysis

链接: https://arxiv.org/abs/2611.20827
作者: Guanqi Lyu,Decheng Qiu,Guangyang Zhang,Fei Zhao,Zhaojun Lu,Jiaji He,Yongqiang Lyu
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:When processor implementation details are unavailable, black-box microarchitectural reverse engineering helps security researchers identify hardware conditions related to security vulnerabilities. However, this process requires substantial manual effort to design experiments, develop test programs, and analyze results for each processor. We present ProcReverser, a multi-agent system that automates this process. ProcReverser uses execution feedback to improve experimental plans and test programs and identifies microarchitectural properties from measurements. We develop a comprehensive suite of property identification tasks focused on hardware mechanisms relevant to Spectre-class attacks, covering speculative execution, branch prediction, memory dependence, and cache behavior. We evaluate ProcReverser on the XuanTie C910, the gem5 RISC-V O3 CPU model, and the SiFive P550. ProcReverser achieves Pass@1 and Pass@5 scores of 81.3% and 93.8%, respectively, on C910, and 83.0% and 100.0% on gem5 O3. On the SiFive P550, ProcReverser further characterizes microarchitectural properties covered by the proposed task suite. These findings highlight the potential of automated microarchitectural reverse engineering to support processor vulnerability analysis and defense evaluation.

[AI-61] Conversing with a Conversational Agent : A Real Conversation Not a Simulated One

链接: https://arxiv.org/abs/2611.20826
作者: Christophe Denis(UMMISCO, IHPST)
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:The claim that interactions with artificial conversational agents merely simulate human conversation rests on a philosophical distinction between reproducing the observable features of an activity and actually performing it. Although this distinction is appropriate for certain physical phenomena, its application to linguistic and symbolic activities is far less straightforward. This article argues that conversations with large language model (LLM)-based agents are genuine conversational activities rather than mere simulations, without thereby attributing consciousness, subjective understanding, or human intentionality to artificial systems. Drawing on the distinction between subjective and objective simulation, the article first examines the limitations of defining conversation exclusively through the intrinsic properties of its participants. It then develops a relational and semiotic account of conversational activity, informed by linguistic pragmatics, Lacan’s Conversing with a Conversational Agent theory of the mirror, Baudrillard’s concept of the simulacrum, Simondon’s philosophy of individuation, and Peirce’s semiotics. These perspectives make it possible to distinguish the reality of conversational interaction from the appearance of artificial subjectivity and from the epistemic validity of the meanings produced. The argument culminates in the concept of Hybrid Semiotic Fabrication, which characterizes meaning-making as a distributed process involving human participants, artificial intelligence, corpora, and technical arrangements. Recognizing the reality of human-AI conversation therefore entails neither anthropomorphizing artificial agents nor granting their outputs epistemic authority. Rather, it requires a more precise account of the relational, semiotic, and normative conditions under which conversational meaning is produced and justified.

[AI-62] Surprise-Gated Memory for Live Speaker Diarization

链接: https://arxiv.org/abs/2611.20819
作者: Louis Mouchon,Alexis Lainault
类目: ound (cs.SD); Artificial Intelligence (cs.AI)
备注: 2 pages, 5 figures, 4 tables

点击查看摘要

Abstract:We transpose the Surprise-Gated Memory recipe, a frozen encoder plus a calibrated novelty signal that gates a one-shot episodic store consolidated by periodic replay, from continual visual learning to live speaker diarization. The moment-to-moment question a streaming diarizer must answer, whether an arriving voice belongs to an already-heard speaker or a new one, is a known-versus-novel decision, exactly the decision the surprise gate is built for. On a frozen fine-tuned DiariZen backbone, our online system SGM-Diar recovers 4.8 of the 9.2 diarization-error points separating a naive online baseline from the offline oracle on AMI-SDM (52%) and 1.9 of 2.3 points on VoxConverse (82%), in the end-of-meeting revisable mode, with calibration frozen on development data and transferred cross-domain without adjustment. These are single-seed runs on three AMI meetings and five VoxConverse sessions, so the ordering of conditions rather than the absolute rates is the result. Beyond error rate, we measure the gate’s behaviour directly and find a coherent memory-like profile: one-shot speaker recall at 0.857, a novelty response that habituates as speakers become familiar, re-identification that is invariant to the length of intervening silence, and a false-novelty rate of 0.001. Two ablations reproduce, now on audio, the central prediction of the visual SGM work: a sliding replay window is worse than no replay at all. The memory is strictly intra-meeting and is wiped at the end, and no gradient touches the run-time system. We report SGM-Diar as a proof-of-concept and state its limitations explicitly.

[AI-63] ForeSay: Tell the Future in Words Branch the Details in Noise for Precipitation Nowcasting

链接: https://arxiv.org/abs/2611.20815
作者: Xudong Ling,Wenjie Luo,Chaorong Li,Tianxi Huang,Tao He,Guiduo Duan
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Precipitation nowcasting involves two distinct scales of uncertainty: macroscopic storm evolution and fine-scale local convective variations. Current generative ensemble methods entangle both scales within independent random noise draws. This coupling causes two practical limitations: every ensemble member must execute the full sampling trajectory from scratch, so an N-member ensemble costs N times a single forecast; moreover, divergent trajectories spatially misalign heavy-rain echoes across members, causing standard pixel-wise averaging to blur critical high-intensity extremes. We present ForeSay, a generative nowcasting framework that decouples macroscopic evolution prediction from local stochastic detail generation. Given historical radar frames, a vision-language model first predicts a textual scenario of future storm dynamics to condition a Rectified Flow generator. Guided by this shared macroscopic prior, ensemble members share early trunk sampling steps to establish the storm trajectory and branch with fresh noise only in later steps to explore local convective details. Because members then differ mainly in local details, self-anchored support-consensus decoding replaces pixel averaging with discrete support voting to preserve sharp heavy-rain cores. Across six datasets, ForeSay achieves the best CSI at the highest precipitation thresholds, with relative improvements over the strongest baseline on each dataset ranging from over 10% to 84%. Our 8-member ensemble retains the forecasting skill and diversity of eight independently sampled runs while reducing computational cost from 8x to 5.2x that of a single forecast.

[AI-64] he Cybersecurity Dataset Observatory: Which Data for AI in Cybersecurity?

链接: https://arxiv.org/abs/2611.20812
作者: Francesco Cerasuolo,Gabriele Mangiacapre,Antonio Pescapè
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI is increasingly used for cybersecurity tasks such as intrusion and malware detection, in enterprise, industrial, vehicular and mobile networks. Its effectiveness depends on data: models need representative network traffic, both to be trained and to be evaluated realistically. Public datasets of cybersecurity network traffic, however, are scattered across laboratory pages, repositories and aggregators, and are continuously released, corrected and withdrawn: 92 of the 234 datasets we catalog were released in 2023 or later. At this pace, surveys that list them are outdated soon after publication, since they describe the field only as it was when they were written. We present the Cybersecurity dataset observatory, a public and continuously updated catalog of cybersecurity datasets based on network traffic. New datasets are added as they are found or submitted, every record is periodically re-checked against its original source, to which it links rather than hosting the data, and every change is recorded in a public change log. This paper documents the first release: we define a traffic-centered scope and a seven-axis taxonomy, and review 234 datasets (204 original captures and 30 derivatives) across 14 domains. The review shows a field that grows quickly but unevenly, where 91 datasets distribute only processed records, a few laboratories produce a large share of the data, and UAV, wireless and V2X networks have little public evidence, while 5G has few real captures. We then relate the catalog to common uses in AI research, outline a procedure for selecting datasets, and discuss benchmark pitfalls and open gaps.

[AI-65] Auditing World-Model Explanations in Retail Forecasting: Faithfulness Accuracy and Price Response

链接: https://arxiv.org/abs/2611.20811
作者: Kathik Iyer
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Forecasting models that read the price and promotion plan explain each forecast as a baseline, what would sell with nothing planned, plus the lift the plan adds. We test such explanations with eight checks, six with a pass rule fixed before any result, and find that passing them does not make an explanation right. In generated store data with known effects, a world model passes the checks of construction, stability, price direction and agreement with sales and with LightGBM, yet its lifts are 1.33 to 1.40 times the truth when merchants plan into high-demand weeks, and 1.23 to 1.42 times in eight variants of that world. It still picks the right weeks to promote, but stock ordered to its lift is off by 0.44 to 0.48 of the true lift. Its stated response to a lasting regular-price change is 4.7 to 6.5 times the true -0.5, and only a measured cause finds this. On Dominick’s, 38 episodes in which a store’s regular prices sat apart from its zone peers give a noisy benchmark: a pooled 8-week response of -0.33 (95% interval -0.68 to 0.01). For the final tuned world models of our companion paper on nine data sets, baseline and lift add up under each model’s own plan, but a deeper price cut lowers the forecast in 29% to 83% of item-week steps on Dominick’s (23% to 80% within prices seen in training), and the world model and LightGBM both state about twice the measured response. Across 703 configurations from the settings searches, checks passed were not related to forecast accuracy (Spearman 0.02, 95% interval -0.02 to 0.05); with no known answer for those versions, this concerns accuracy, while whether an explanation is right is measured only in the generated data.

[AI-66] Regularizing Treatment Leakage for LLM -based Uplift Estimation

链接: https://arxiv.org/abs/2611.20794
作者: Bojing Feng,Jinqiu Deng,Zihan Hong,Yuan Xue,Wenqi Xiang,Shuhan Wang,Feng Xiao,Shiwei Wang,Zijiang Liu,Shaomin Li,Yongzhi Qi
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Uplift modeling estimates the conditional average treatment effect (CATE) and supports industrial decisions such as promotional pricing. When large language models (LLMs) encode product text, treatment-revealing content can introduce shortcuts that do not transfer to deployment. We propose \emphLLM-based Uplift Leakage Regularization (\textscLULR), which uses a CLUB-based penalty on the causal head’s trainable shared representation while keeping the LLM encoder frozen. Controlled proxy experiments show a positive ranking-score interaction and higher scores than the tested weight-decay configuration; treatment-dependent evaluation limits their interpretation to exploratory model behavior. Structured-only comparisons provide reference performance, without establishing a general advantage over tree-based estimators. On a million-record industrial pricing RCT, \textscLULR improves offline \textlift@0.3 in 4/5 seeds, although the difference is not statistically significant and estimated policy values are near-tied. A separate one-week online A/B experiment records 5.45% higher sales volume and 1.91% higher GMV than the control arm. These aggregate results provide short-term business evidence for the complete deployed pricing policy; they do not isolate the regularizer’s contribution.

[AI-67] GainCtrl: Predicting Remaining Gains for Efficient LLM Agent Execution

链接: https://arxiv.org/abs/2611.20736
作者: Kangwen Zhao,Jianfeng Cai,Ruopei Sun,Mingxiao Feng,Peng Zhang,Wengang Zhou,Houqiang Li
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models (LLMs) are widely used to power agents that interact with tools and environments to solve complex tasks. Repeated attempts improve the quality of tool-using agents, but continued execution incurs additional costs even when it is unlikely to improve the best result ultimately returned. This raises the question of when to stop an attempt to reduce execution cost while preserving task-level Best-of-N quality. To address this question, we introduce GainCtrl, a gain-based execution controller for LLMs. In GainCtrl, a lightweight CPU predictor uses interaction histories to estimate expected positive gains over retained scores, which are then used to guide stopping decisions with calibrated thresholds to limit the loss in mean Best-of-N task quality. We evaluate the controller through trajectory replay across three environments and three LLM actors. Comparisons of interpolated quality-cost Pareto frontiers show that GainCtrl generally uses fewer tokens than baselines at matched target quality. Fixed policies also achieve quality close to that of full execution while substantially reducing token consumption. The predictor runs at millisecond latency on a single CPU thread, supporting repeated monitoring without access to LLM internals.

[AI-68] Routing Beyond Task Categories: A Cost-Aware LAB-Bench Study

链接: https://arxiv.org/abs/2611.20732
作者: Ahmed Tariq(Independent researcher),Leo Guinan(Idea Nexus Ventures)
类目: Artificial Intelligence (cs.AI)
备注: 21 pages, 3 figures. Reproduction archive at this https URL

点击查看摘要

Abstract:We tested whether features of biology questions or of a model’s first answer can improve accuracy within an inference budget. In a LAB-Bench pilot, we compared a category router with one that adds six structural features. Neither reached the primary 80% accuracy target on development. At the 50% and 60% secondary targets, both chose Haiku for all 120 held-out questions and scored 71/120. We then froze a new comparison before collecting responses on 144 additional questions. Semantic routing scored 86/144. A policy that could defer after reading Haiku’s answer scored 87/144. Both stayed within the inference budget of USD 0.678 per 1,000 queries. Against a 95% Haiku and 5% GPT mixture, the accuracy gains were 0.66 percentage points for semantic routing and 1.35 points for deferral. The paired 95% intervals were [-0.28,2.25] and [-0.14,3.47] points. A simple rule that escalated Haiku nonanswers scored 92/144 at USD 0.710 per 1,000 queries, above the budget. A retrospective nonanswer rule with a calibration-selected escalation probability gave an expected 85.63/144. At the learned deferral policy’s observed inference cost, post hoc interpolation gave an expected 86.14/144. In a later development check, Astra and Gemini each scored 30/40. The best tested policy in each router family matched that score. The pair covered 33/40 in hindsight. The routers missed the 32/40 criterion, so the proposed 120-question evaluation did not start.

[AI-69] PEETS: Preference-aware Embeddings with Explicit Tree Structure for Auxiliary Product Bundling

链接: https://arxiv.org/abs/2611.20726
作者: Maolei Huang,Jing Tian,Dexuan Duan,Yan Chen,Huawei Cao,Shuhan Song
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Auxiliary products such as flight protection and flexible-change services are recommended alongside primary travel purchases. This setting creates a granularity ambiguity: SKUs within one product type share semantics but differ in price and coverage, while rejecting one offer does not imply disinterest in the entire type. We present \textbfPEETS (\textbfPreference-aware \textbfEmbeddings with \textbfExplicit \textbfTree \textbfStructure), a training-time framework for auxiliary product bundling. PEETS regularizes an existing click–conversion predictor through two conditional comparisons. Hierarchical Contrastive Loss uses the pair’s common ancestor as a reference; Negative Contrastive Loss conditions structural sampling and comparison shifts on logged rejection events. Both act on shared SKU representations during training. This separates catalog relatedness from individual preference supervision and leaves the predictive serving path intact. PEETS achieves 78.20% AUC, exceeding TWIN V2 by 0.90 percentage points and the matched base model by 1.15 points. In a 14-day online evaluation, gross auxiliary GMV per exposed user-day improves by 8.20% over Policies, 3.06% over Base, and 0.84% over TWIN V2, with all three user-bootstrap 95% intervals above zero. Bundling rate is 0.51 points below Policies. The cancellation interval’s upper limit exceeds the prespecified margin. PEETS adds no auxiliary-loss computation to serving and records 42.1~ms p95 latency, compared with 42.0~ms for Base.

[AI-70] Agent FaultBench: State-Verified Evaluation of Tool-Fault Recovery NAACL2027

链接: https://arxiv.org/abs/2611.20711
作者: Sachin Gupta
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 63 pages, 7 figures, including appendices. Submitted to the NAACL 2027 Industry Track

点击查看摘要

Abstract:Does recovery advice help an agent finish a task after a tool fails? AgentFaultBench evaluates LangGraph and CrewAI agents performing specified tasks in an isolated GitLab installation. We inject tool errors and check both the resulting files and issues and the required reads, verification and final response. The prompt supplies the expected issue and report text, so the task tests execution rather than content generation. In the primary comparison of 1,944 planned runs, advice changes success by -8.33, 0.00 and +1.06 percentage points relative to a neutral reminder; missing outcomes leave the latter two comparisons inconclusive. Later experiments test a revised recovery policy. GPT-4o mini improves, while Sol and Claude already complete the recoverable cases. In a separate 624-run study on six related tasks not previously used by that policy, the guided condition records 161 correct final states and one unknown outcome in 192 planned runs, versus 142/192 with neutral feedback. The gain lies between +9.90 and +10.42 pp when missing outcomes are included. Removing issue IDs gives 47/48 successes versus 46/48 with IDs; other execution information remains. A second checker agrees on 1,879 observed primary outcomes and 241 constructed test cases. No human validation was collected. The findings support testing advice against explicit recovery controls rather than assuming it helps.

[AI-71] A First Look at System One Decision Models for Malware Detection

链接: https://arxiv.org/abs/2611.20699
作者: Pier Mattia Basciano,Giovanni Farina,Salvatore Monteleone
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Malware detection needs scores calibrated at a fixed false positive rate (FPR) and decisions an analyst can read: feature-based detectors give the first, large language models (LLMs) on decompiled code the second, through generated text of uncalibrated confidence. System One decision models answer a typed question about a text with a probability over its declared answers in one forward pass. We present the first evaluation of these models for Windows Portable Executable malware detection, analyze how they decide, and propose S1Mal, an ensemble on the typed answers and the embedding of CLM-8B. In an isolated virtual machine, Ghidra and capa decompose each binary, and the host renders the records into a canonical textual state, a digest of headers, imports, MITRE ATTCK capabilities and strings, optionally with decompiled functions, without class shortcuts. Three readouts score each model on the state: A0, the probability of one verdict question; A1, 32 atomic behavioral questions weighed by gradient-boosted trees; A2, heads on the model’s state embedding. S1Mal feeds gradient-boosted trees with the atomic answers and the A2 score of CLM-8B. We evaluate CLM-8B, ten open decision models and two LLMs read through logits on BODMAS temporal splits, SOREL-20M out-of-distribution malware and 14,534 matched benign files. On BODMAS, A0 wastes the static evidence: its area under the receiver operating characteristic curve (AUROC) spans 0.4025-0.8590 across 13 models. A1 recovers it, raising every model’s AUROC by +54.61% on average; A2 lifts CLM-8B by +0.2749 to 0.9943, above a combiner on digest facts (0.9901), and reaches 0.9853-0.9948 on nine open models. S1Mal reaches 0.9946 and detects 93.47% of the malware at a realized FPR of 1.45%, and 59.15% at 0.1% FPR on the test ROC curve, against 46.59% with A2. S1Mal gives analysts calibrated decisions, each with the atomic answers behind it.

[AI-72] SpeCal: Spectral Calibration for Unsupervised RLVR KDD2027

链接: https://arxiv.org/abs/2611.20691
作者: Kai Tang,Xing Zheng,Haobo Wang,Guangcheng Zhu,Shenzhi Yang,Zhengqing Zang,Yingfan Ma,Cheng Peng,Bowen Song,Weiqiang Wang
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted to KDD 2027

点击查看摘要

Abstract:Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning abilities of large language models, but its dependence on ground-truth answers or external verifiers limits its ability to exploit abundant unlabeled reasoning data. Unsupervised RLVR alleviates this bottleneck by deriving scalar rewards from the model’s own rollouts. However, the reliability of such self-generated supervision varies substantially across rollout groups and individual responses, while existing methods construct or refine their rewards mainly from decoded outcomes and reward statistics. These endpoint signals reveal what answers the model produces, but provide limited information about whether the underlying rollout representations form a reliable structure for learning. Our key observation is that hidden-state spectra offer a complementary view of supervision reliability by characterizing both the organization of a rollout group and the support received by each rollout within that group. Based on this observation, we propose SpeCal, a bi-level spectral calibration framework for unsupervised RLVR. At the group level, SpeCal combines spectral mass with a Gaussian band-pass gate over normalized entropy-based effective rank, favoring rollout groups with substantial representation variation whose effective dimensionality lies in a preferred regime. At the rollout level, the same group effective rank determines how many dominant directions are retained in a leave-one-out spectral subspace formed by the remaining rollouts, within which the conditional support of each held-out rollout is measured. The resulting signals calibrate the original reward-induced advantages at the group and rollout levels without redefining the underlying unlabeled reward.

[AI-73] UbiHearo: Bringing Scenario-Aware Sound Guidance to DHH Users with Mobile Agents

链接: https://arxiv.org/abs/2611.20686
作者: Fengmin Wu,Sicong Liu,Zimu Zhou,Chenren Xu,Wanbing Zhao,Peiwen Sun,Zhiwen Yu
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
备注:

点击查看摘要

Abstract:For Deaf and Hard-of-Hearing (DHH) users, effective sound awareness goes beyond detecting \emphwhat sound occurs; it requires understanding \emphwhat it means, \emphhow urgent it is, and \emphwhat to do. However, existing mobile assistants largely reduce acoustic scenes to isolated labels or generic alerts, failing to provide contextual guidance when the same sound implies different risks across scenarios. We present \sysname, a mobile hearing assistant agent that closes the sensing–reasoning–action loop for \textitscenario-aware and \textitpersonalized DHH assistance. \sysname adopts a \textitneuro-symbolic design that transforms continuous acoustic and physical context streams into structured intermediate representations (IRs) for controllable reasoning and guidance. It introduces three key mechanisms. \textitFirst, a perception-thresholded sensing mechanism enables energy-efficient always-on awareness by activating processing only for perceptually meaningful events while preserving safety-critical cues through deterministic overrides. \textitSecond, a staged neuro-symbolic reasoning framework performs promptless scenario understanding by conditioning an on-device audio-language model on structured IRs instead of user-authored context descriptions. \textitThird, a bounded human-in-the-loop personalization mechanism adapts user-specific guidance through lightweight on-device decision-boundary updates while preserving population-aligned reasoning. We evaluate \sysname on smartphones and in-vehicle systems across diverse scenarios. It achieves favorable accuracy–latency–energy tradeoffs and personalized guidance.

[AI-74] MASTA: A Feedback-Scheduled Multi-Agent System for End-to-End Tamarin Modeling and Analysis NEURIPS2026

链接: https://arxiv.org/abs/2611.20671
作者: Yongkang Xiao,Qiyi Deng,Jing Chen,Min Shi,Ju Ma,Ruiying Du
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: Accepted at NeurIPS 2026 (Main Track, poster). 19 pages

点击查看摘要

Abstract:Tamarin-based formal verification of security protocols requires expert-authored multiset rewriting rules and temporal-logic lemmas. Translating a natural-language protocol description into a complete formal model with an LLM is hard because syntactic, reachability, and property-level errors compound across the multi-step modeling process, and a single agent that both authors and self-evaluates cannot reliably localize them. We present MASTA, a multi-agent system that addresses these difficulties through three innovations. First, an operational–evaluative role decomposition assigns rule modeling to an Architect, property modeling to an Auditor, and pure dispatch to a dedicated Arbiter that holds no authoring duties. Second, the Arbiter implements a Tamarin-anchored scheduler that dispatches the next agent based on the latest verifier verdict and routes problem-tagged fixes through a modification list, with a six-class problem taxonomy and an LLM deduplication check preventing routing deadlock. Third, we construct a benchmark of 10 protocols paired with 139 mutation variants and five fine-grained metrics that jointly assess model correctness ( M_1 – M_3 ) and completeness ( M_4 – M_5 ). On this benchmark, MASTA on Qwen 3.5-plus improves over OpenCode on the same backbone by up to +0.27 on individual metrics and ranks first on both the correctness and completeness averages, surpassing Codex (GPT-5.5) and Claude Code (Opus 4.7) despite using a weaker LLM. Ablations on five protocols confirm that removing the Arbiter collapses property and verification scores to zero.

[AI-75] Instrumenting Failure: Scoring Language Model Agents at Every Decision Node in Imperfect-Information Games NEURIPS2026

链接: https://arxiv.org/abs/2611.20644
作者: Michael Kingston,Sonit Singh,Alan Blair
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted at the Interpreting Agent Behavior (IAB) Workshop, NeurIPS 2026. 26 pages, 6 figures. Code: this https URL

点击查看摘要

Abstract:Agent evaluations report task success and rarely say what an agent did at each decision. We score every decision node of three 12B to 14B instruct-tuned language models (LMs) playing five zero-sum imperfect-information games against four opponents on eight properties from decision and game theory, five of them against a CFR+ equilibrium reference. At each node, the LM also states a belief about the opponent’s hidden state in a separate call. Without memory and at temperature 0, stated beliefs sit 0.250 to 0.288 in total variation distance from the equilibrium posterior, and where the action responds to the state its agreement with that belief exceeds chance by -0.04 to 0.19. The two calls frame the opponent differently, so the result is a behavioural inconsistency and makes no claim about internal beliefs. Under one prompt and greedy decoding, the LM’s action never varies with the state in 10 of the 38 cells where the reference varies its own. Memory lowers the constant-action share in 9 of 13 testable ladders at temperature 0; at temperature 0.2 the cells at 1 fall from 16 to 5 of the 46 informative cells, while the belief results hold. A gate withholds, with a stated reason, any belief score the run cannot support, and seven of the eight non-zero defect rates are attributable to the instrument, a representation, or the LMs.

[AI-76] GEAR2: Pruned Experts Live On through Gated Reuse of Activated Experts and a Missing-Mass Residual

链接: https://arxiv.org/abs/2611.20616
作者: Linliang Chen,Yan Zhong,Wang Kang,Sai Li,Ji Wu,Ji Liu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Mixture-of-Experts (MoE) Large Language Models (LLMs) activate only a few experts per token while generally keeping all the experts in GPU memory for high efficiency. Expert pruning reduces this GPU memory footprint cost, yet discards the contribution of pruned experts. Existing remedies either rely on static substitutes, merge experts at the cost of altering their functions, or require costly recovery training. In this paper, we find that the functionalities of pruned experts can be efficiently reconstructed from the activated remaining experts. We propose GEAR (Gated Expert Activation Reuse), which reuses the scaled outputs of reliable activated experts without any training or additional expert computation. In addition, we propose GEAR ^2 to further recover what reuse cannot express with an ultra-lightweight residual. On five MoE LLMs from 16B to 106B parameters with 25% and 50% of experts pruned, GEAR improves over its training-free counterparts in nearly every setting (up to 4 pp.) in terms of accuracy, and GEAR ^2 outperforms baseline approaches (up to 2.75 pp.) in every setting on nine zero-shot benchmarks, with larger advantages (up to 2.72 pp. higher) under more aggressive pruning (up to 50%).

[AI-77] Newton-PINet: A fast physics-informed neural network with Newton linearization for meta-learning nonlinear PDEs NEURIPS2026

链接: https://arxiv.org/abs/2611.20609
作者: Yuchen Fan,Chang Wei,Pao-Hsiung Chiu,Chin Chun Ooi,Heyang Wang,Jian Cheng Wong
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computational Engineering, Finance, and Science (cs.CE); Computational Physics (physics.comp-ph)
备注: Accepted at NeurIPS 2026

点击查看摘要

Abstract:Scientific machine learning has opened new avenues for solving parameterized partial differential equations (PDEs), enabling models to learn a family of PDEs and generalize to unseen instances. In this context, data-driven operator learning methods typically require large training datasets, while physics-informed neural networks (PINNs) suffer from difficult optimization and limited generalization, especially for nonlinear PDEs. We propose Newton-PINet, a physics-informed network enhanced by Newton linearization, offering an effective meta-learning framework for nonlinear PDEs. Newton-PINet (i) employs a physics-informed multilayer network with skip connections, where the output-layer weights are solved by least squares; (ii) adopts a two-stage learning strategy that first leverages gradient-based training to learn robust representations from the available training tasks, and then performs gradient-free fine-tuning on the output layer for fast task-specific generalization; and (iii) incorporates a Newton linearization method to speed up the least-squares iteration for nonlinear PDE problems. On a challenging nonlinear reaction-diffusion benchmark, Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers, while using 16x fewer training tasks and over an order of magnitude less training time (about 5 minutes versus several hours). This work advances the meta-learning of PINNs toward data-efficient, fast, and generalizable physics solvers. The code and datasets are publicly available at: this https URL.

[AI-78] AdMem: Expectation-Grounded Procedural Memory for Long-Horizon Task-Solving Agents NEURIPS2026

链接: https://arxiv.org/abs/2611.20605
作者: Runzhe Wang,Huilin Lu,Shengjie Liu,Li Dong,Zhihuai Zhu
类目: Artificial Intelligence (cs.AI)
备注: 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Workshop: Agents in the Wild: Safety, Security, and Beyond

点击查看摘要

Abstract:An agent that carries memory across tasks is only as good as the entries it keeps. We begin from an observation that is easy to reproduce and uncomfortable for the field: naively accumulating cross-task experience can make an agent \emphworse than one with no memory at all. On AgentBoard, a workflow-induction memory solves fewer tasks than a memoryless ReAct agent in seven of eight domains, and a per-step procedural store with no surrounding structure loses half of ReAct’s completions on Jericho. We argue that the missing ingredients are per-action credit assignment and an explicit life cycle for stored entries, and present AdMem, which supplies both. Before each action the actor commits to an expected outcome; a concurrently running critic scores the realized outcome against that expectation and writes a procedural entry carrying a binary reward and a reflection, so credit lands on the action that earned it and no environment reward is required. Each entry then carries a scalar utility, updated online under a noisy-OR reward model, that ranks retrieval jointly with context similarity and drives eviction and merging, giving the store a forgetting and consolidation policy instead of unbounded growth. Across 707 tasks in eight AgentBoard domains AdMem solves 441 against 388 for ReAct and 312 for the workflow-induction baseline — a pooled gain significant under a two-sided Fisher exact test, concentrated in the domains that supply transferable cross-task structure. All memory stays human-readable text, which is what makes the failures above diagnosable — and, as we discuss, also makes the store an auditable but attackable surface.

[AI-79] BiGSyn: A Bi-Level Heterogeneous Graph Contrastive Learning Framework for Interpretable Drug Synergy Prediction ICDM2026

链接: https://arxiv.org/abs/2611.20603
作者: Senquan Wang,Hui Fang,Haoran Nan,Xingwei Wang,Qiang He
类目: Artificial Intelligence (cs.AI)
备注: 2026 IEEE International Conference on Data Mining (ICDM 2026)

点击查看摘要

Abstract:Drug combination therapy has become a mainstream of modern medicine, with the potential to enhance efficacy, reduce toxicity, and overcome drug resistance. However, identifying synergistic drug combinations remains challenging, as it requires modeling complex interactions across multiple biological scales—from atom-level molecular structures to entity-level relations in biological networks. Existing approaches typically model these levels in isolation, overlooking complementary signals and long-range dependencies. To address these limitations, we propose BiGSyn, a Bi-level heterogeneous Graph contrastive learning framework for drug Synergy prediction. Specifically, BiGSyn leverages heterogeneous graph transformers to extract high-order dependencies from molecular graphs and biological networks, and introduces a bi-level contrastive learning objective to bridge the cross-level semantic gap. Extensive experiments on two benchmark datasets show that BiGSyn consistently outperforms state-of-the-art baselines. Moreover, its attention mechanism highlights key molecular substructures and biological interactions driving synergy, providing interpretable insights.

[AI-80] Dynamic Shaping: Acquiring Interfaces and Constructing Execution over Frozen Pretrained Components

链接: https://arxiv.org/abs/2611.20573
作者: Nathaniel Oh
类目: Artificial Intelligence (cs.AI)
备注: 12 pages, 2 figures. Code and complete evidence: this https URL

点击查看摘要

Abstract:We introduce Dynamic Shaping, a method for constructing executable neural arrangements through acquired response and partial-derivative interfaces over frozen pretrained components. The method learns local descriptions of available computation, ranks component rebindings using task-gradient alignment, transports the current interface across a rebinding, and retains changes through behavioral feedback. Continuous interface acquisition uses proximal policy optimization; the source coefficients remain fixed. We derive an exact identity relating transport error in component responses and partials to changes in action logits and their sensitivities. A pixels-only ViZDoom Basic study evaluates first task acquisition without a competent reference controller, demonstrations or intact response targets. Across three initializations and 100 shared held-out game seeds, initial arrangements succeed on 86/300 controller/episode pairs; joint interface-and-binding construction succeeds on 295/300. Fixed-binding interface acquisition also reaches 295/300. Each policy executes four source fragments, with 110 acquired interface values and fixed random visual and action projections. The artifact includes all candidate trials, interventions and costs; all 1,200 recorded initial/retained outcomes replay exactly. The contribution is an explicit profile, rebind, transport and retain procedure and a reproducible reference-free acquisition case over frozen pretrained computation.

[AI-81] oolGraph as a Q-Function for Test-Time Scaling of Tool-Use Agents NEURIPS2026

链接: https://arxiv.org/abs/2611.20569
作者: Ke Shen,Ankita Bhaumik,Li Dong,Shengjie Liu,Shuzhen Li,Naveed Ahmed Janvekar
类目: Artificial Intelligence (cs.AI)
备注: 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Workshop: From Pretrained Representations to Acting Agents (PTA)

点击查看摘要

Abstract:Tool-use agents must chain interdependent calls, yet which outputs satisfy which later inputs in a deployed catalog is not something a general-purpose language model reliably infers from tool descriptions alone. We compile that structure once into a reusable external representation. ToolGraph is a directed, weighted dependency graph over a tool catalog: parameter matching and interventional probes, which execute a candidate chain with and without its proposed predecessor, propose edges whose evidence is combined into an edge confidence. Prior work uses such graphs to retrieve tools, search for tool paths, or score the next call. We ask instead whether the same structure can supply a \emphmulti-step value by composing graph paths into a goal-conditioned action value (Q(s,a,g)). We study three estimators (dynamic programming over entity sets, a GraphSAGE estimator distilled from it, and an LLM scorer conditioned on graph evidence) plus a verbalizer that returns action–goal feedback to the agent, which keeps final action selection. Across two tool-use benchmarks and three backbones, we compare value ranking against personalized-PageRank retrieval over the identical graph, and against direct and reflective agents. Composition pays off only when the estimator also reads the concrete state: the LLM estimator improves on PageRank in every benchmark–backbone setting, by (8.3) points on average, while the two purely structural estimators fall below it, because a value defined over tool types cannot separate a correct tool called with a wrong argument. Graph construction is a one-time cost; we report it separately rather than assume it amortized away (at our scale it is comparable to request-time cost) and sweep four knobs controlling how much graph evidence to expose against success and tokens.

[AI-82] Memory-Efficient Speculative Decoding with Quantized Draft KV Cache NEURIPS2026

链接: https://arxiv.org/abs/2611.20565
作者: Saksham Gupta,Michael Li
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted at the AXIOM Foundations of Efficient Deep Learning Workshop at NeurIPS 2026

点击查看摘要

Abstract:Speculative decoding reduces generation cost, but it stores two KV caches: one for the target model and one for the draft model. We reduce this extra memory by quantizing the draft cache. The target cache and verifier remain in BF16, so that compression cannot change the target model’s final correction rule. We first study how to divide precision between keys and values. A naive per-token quantizer makes keys appear much more sensitive: splitting a 12-bit budget, K8V4 beats K4V8 in acceptance by 10.87pp. However, with grouped per-channel keys, the gap becomes -0.25pp. A controlled Qwen test confirms that changing only the key quantizer can erase or reverse the preferred bit split. Hence, following KIVI, we propose K4V4 quantization with per-channel keys and per-token values. Across six model pairs, our method reduces draft cache memory by 66.20%, and combined memory across both caches by 23.95%. Mean acceptance drops from 57.93% to 57.14%, only 1.36% relative. Overall, quantized draft caches can make speculative decoding more memory efficient; as a predictive principle for efficient AI design, key and value precision must be co-designed with the quantizer layout.

[AI-83] One-Shot Calibration via Model Editing

链接: https://arxiv.org/abs/2611.20562
作者: Jiaxing Qiu,Luke Hakso,Roxana Daneshjou,Ahmed Alaa,Thomas Hartvigsen
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Model editing aims to correct errors in large, pretrained models by replacing model generations with user-provided responses. Existing editors assume a single target answer for each edit. However, many real tasks inherently entail uncertainty, resulting in multiple plausible answers with varying likelihoods in user feedback. Such feedback forms a target distribution to not only correct but also calibrate model generations. We therefore propose SoftEdit, the first model editor to leverage uncertainty in user responses and perform one-shot calibration during editing. To generalize calibration to unseen samples, SoftEdit uses co-plausibility relationships between answers while preserving real distributional shifts, balancing the new distributional generality-locality tradeoff. Across three language models, three vision-language models, and two large datasets, SoftEdit achieves state-of-the-art performance, ultimately showing that uncertainty in user responses enables generalizable calibration of pretrained models beyond single-answer corrections.

[AI-84] VAHBench: A Harness Benchmark for Agent ic Long-Video Generation

链接: https://arxiv.org/abs/2611.20535
作者: Hongbo Zhang,Liuyang Song,Hang Zhou,Xinchen Liu,Wei He,Kai Han,Yunhe Wang
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI agents are commonly used for everyday information retrieval and coding. We ask whether they can also interpret human requests and turn them into complete videos. We introduce VAHBench, a benchmark that tests this ability through five short-drama briefs specifying story, characters, dialogue and duration. Agents plan and carry out production using shared generation tools. Evaluation combines video completeness and basic quality with checks of production reliability and reporting accuracy. Across 240 runs covering 48 configurations of general-purpose and video-specific agents, higher quality scores do not necessarily mean more complete delivery. Missing episodes, repeated footage and static filler reveal gaps between producing a video file and fulfilling the brief. Added composition tools help some agents but do not consistently improve performance. VAHBench measures how well agents translate user intent into finished video content and helps identify where that process breaks down.

[AI-85] LINK: Lightweight and Inductive Node-to-Node Propagation Rule Learning for Traffic Kriging

链接: https://arxiv.org/abs/2611.20521
作者: You Shang,Yu-Kwan Ching,Weifeng Su,Guanyao Li,Ruiyuan Li,Gary Shueng-Han Chan
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Spatio-temporal traffic kriging aims to reconstruct traffic time series at unobserved locations with no historical traffic observations by leveraging spatial and temporal dependencies across the road network. Existing Graph Neural Network (GNN)-based approaches primarily rely on neighborhood aggregation to capture spatial dependencies, thereby modeling traffic interactions implicitly and making it difficult to accurately recover traffic magnitudes at locations without historical observations. To address this limitation, we propose LINK, a Lightweight and Inductive Node-to-node propagation rule learning paradigm for traffic Kriging. LINK introduces three key designs: 1) a propagation-centric traffic kriging paradigm that explicitly learns a shared node-to-node transition rule; 2) a context-aware Spectral Propagation Network that performs traffic-state propagation in both the spectral and time domains; and 3) TraffiKrig, a realistic traffic kriging benchmark that preserves heterogeneous road-segment characteristics and road-specific metadata while enabling systematic comparison of representative kriging methods under a unified metadata-enhanced setting. Extensive experiments demonstrate that LINK achieves strong reconstruction accuracy and computational efficiency. Compared with the strongest baseline on each subset, LINK achieves an average MAE reduction of 54.7% over original baselines. LINK also achieves substantial efficiency improvements, reducing training time by 66.5% and inference time by 86.6% compared with the fastest competing methods. Zero-shot experiments further demonstrate the transferability of the learned propagation rule across regions.

[AI-86] Agent -Guided Configuration Tuning for SMT Solvers

链接: https://arxiv.org/abs/2611.20484
作者: Cong Wang,Qiaoyi Zheng,Zhaoguo Wang,Fan Long
类目: Artificial Intelligence (cs.AI)
备注: 23 pages, 3 figures, 11 tables

点击查看摘要

Abstract:SMT solvers expose many configuration choices that determine how well their existing algorithms perform on a workload. We present, to our knowledge, the first workflow that uses general-purpose LLM agents to tune SMT solver configurations. Agents read solver documentation, source code, and persistent experiment records to propose option changes, and a deterministic harness measures each proposal, validates the returned models, and selects changes separately for each competition Division. Over roughly nine weeks, our tuning campaign recorded 4,545 proposals for cvc5 across 13 Divisions and 36 logics. We evaluate the tuned configurations on the 2026 and 2025 SMT-COMP ModelValidation benchmarks under a 24-second cutoff. On problems not used for training, they validate more problems than official cvc5 in ten of the 13 Divisions in 2026 and in eight in 2025, with geometric-mean gains of 7.8% and 6.8% across Divisions. On the complete official selections, which include training problems, the tuned floating-point counts exceed those of every individual competition participant in both years. Further studies show how configuration changes make existing algorithms effective by changing access to strategies, scheduling, and option compatibility, and how benchmark families trade off against one another. The agents’ experiment traces contain hypothesis-based proposals in every tuning group, and 64.3% of the extracted hypotheses draw on prior measurements. These results show that agent-guided tuning can specialize a mature SMT solver and that accumulated experimental evidence informs sustained search. Comments: 23 pages, 3 figures, 11 tables Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2611.20484 [cs.AI] (or arXiv:2611.20484v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.20484 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-87] Model Immunization Beyond Condition Numbers: The Importance of Plateau Regions NEURIPS2026

链接: https://arxiv.org/abs/2611.20472
作者: Quoc Nguyen,Trung Le,Dinh Phung,Mehrtash Harandi
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted by NeurIPS 2026

点击查看摘要

Abstract:Harmful fine-tuning can rapidly compromise safety alignment in Large Language Models, underscoring the vulnerability of deployed models to post-release modification. While prior work motivates model immunization by inducing ill-conditioned curvature to slow harmful optimization, this approach is not scalable due to the high computational cost of Hessian estimation. We refine this paradigm by extending the theoretical framework to prioritize plateau regions alongside curvature. We further introduce an immunization algorithm based on efficient Hessian estimation and bounds on the extremal singular values of the Hessian. We demonstrate the proposed immunization framework in three applications: (i) Fine-tuning-as-a-service settings where benign user data may be mixed with poisoned harmful data, (ii) open-weight settings where an adversary can fine-tune on harmful-only data with varied training dynamics, and (iii) robust unlearning, where immunized models better resist relearning attacks after unlearning. Extensive experiments confirm that our approach achieves robust defense against adversarial fine-tuning while largely maintaining general utility and downstream trainability. Our code is available at this https URL.

[AI-88] abular Foundation Models as Emulators for Forecasting Terrestrial Ecosystem Dynamics

链接: https://arxiv.org/abs/2611.20468
作者: Zhihao Wang,Ruohan Li,Lei Ma,George Hurtt,Xiaowei Jia,Yiqun Xie
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Process-based terrestrial biosphere models such as the Ecosystem Demography (ED) model underpin global carbon budgets, but their cost limits global, high-resolution and scenario-rich simulation. Deep emulators such as DeepED reduce that cost but require task-specific architectures and training. Tabular foundation models offer a different route: TabPFN predicts new rows by in-context learning, without being trained on the task. We compare TabPFN-3.5 with two DeepED baselines, DeepED_LSTM and DeepED_Transformer , on CarbonGlobe’s overall evaluation, which trains on a uniform 1/16 sample of sites and tests on the other 50,779, predicting the 40-year state of all seven ED outputs. We then test every model, trained on present-day data only, on ED runs forced by SSP1-2.6 and SSP5-8.5 for 2021–2060 in six continental forest regions. With no training, TabPFN-3.5 has the lowest present-day height error of the three models and is within 4% of DeepED_LSTM for soil carbon, but is less accurate for AGB, LAI and the carbon fluxes. In 2060, errors are several times larger than at the same sites under present-day forcing, and every model underestimates the biomass that ED gains. Under SSP5-8.5, TabPFN-3.5 is the most accurate model for height, AGB, soil carbon and LAI, with the lowest AGB error in all six regions, although, unlike ED, it predicts no larger biomass gain than under SSP1-2.6. Under SSP1-2.6, no model is best overall.

[AI-89] ReRead: Wrong Predictions Useful Representations - Correcting Frozen Robot Policies by Re-reading Their Own Representations

链接: https://arxiv.org/abs/2611.20465
作者: Xu Su,Shunlong Xiao,Qintao Yi,Mingkuan Zhao,Xinmiao Yu,Yizhe Li,Zihan Wang,Bokui Chen
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:When a pretrained robot policy fails on challenging tasks, common remedies include retraining the backbone, adding an external visual encoder, or introducing a memory module. We argue that wrong predictions do not imply wrong representations: a policy can retain the required information yet fail to read it out as the correct action. To test this view, we propose ReRead: with the backbone frozen, a corrector re-reads the current and two previous caches to predict an action residual. The corrector sees no images, is trained solely on offline demonstrations from all 50 RoboTwin 2.0 tasks, and adds only 4.17% latency. Ablations and interventions show that the correction relies on the backbone’s own representations: none of four external visual encoders adds further gain, and swapping in past caches from other trajectories sharply increases action error. On the 11 hardest RoboTwin 2.0 tasks, ReRead improves the success rates of LingBot-VLA, \pi_0.5 , and Fast-WAM by 23.45, 13.09, and 8.00 percentage points, respectively. In paired evaluation on all 50 tasks, it raises Fast-WAM’s average success rate from 90.73% to 92.66%. On a Flexiv dual-arm robot, it raises success on a phase-ambiguous task from 0% to 36%.

[AI-90] Designing Deceptive Model Organisms that Successfully Evade Auditing

链接: https://arxiv.org/abs/2611.20459
作者: Vasisht Duddu,Mark Wang,Lipeng He,Tim G. J. Rudner
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI models can exhibit deceptive behaviors such as strategically underperforming on evaluations (sandbagging), faking alignment, or covertly acting on behalf of a hidden principal. Safe deployment therefore requires auditing techniques that can reliably detect such behaviors. To develop and test these techniques, we can train model organisms (MOs): models that exhibit deceptive behaviors and serve as realistic evaluation testbeds. However, for such testbeds to be informative, MOs must actively anticipate and resist auditing—a gap in prior work that can lead to overestimation of auditing effectiveness and underestimation of safety risks. We address this gap by designing MOs that capture worst-case deceptive behavior and evade high-effort auditing. Focusing on sandbagging, we frame MO design as a security game between an auditor and a deceptive model and identify three design requirements for sandbagging MOs: (i) effectiveness (underperforming on targeted capabilities), (ii) utility preservation (maintaining performance on all other capabilities), and (iii) robustness (resisting auditing). We propose three MOs: two optimized against blackbox auditing techniques and one optimized against both blackbox and whitebox techniques. Through a comprehensive evaluation, we show that, unlike MOs from prior work, our MOs satisfy all three requirements, providing worst-case baselines for developing and evaluating future auditing techniques.

[AI-91] he Logic Chain Breaking Point: A Controlled-Variable Benchmark for Reasoning Degradation in Open-Weight Language Models

链接: https://arxiv.org/abs/2611.20447
作者: Akif Islam,Mohd. Ruhul Ameen,Md. Ekramul Hamid
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:We introduce The Logic Chain Breaking Point, a controlled-variable benchmark for deductive reasoning in open-weight language models. It spans three phases: (i) pure modus-ponens chain scaling, (ii) distractor resilience with a fixed core chain, and (iii) contradiction detection under primacy versus recency placement. Independent variables follow a powers-of-two sequence v \in \1,2,4,8,16,32,64,128\ , with ground truth known by construction. We evaluate sixteen open-weight models across instruction-tuned and reasoning-tuned panels, including a small-model suite and a larger local-model extension. All models are evaluated with greedy decoding under model-appropriate context and output budgets, while context- and length-limited trials are quarantined from the valid accuracy denominator. Our results reveal a sharply non-monotonic difficulty surface: pure deductive chains up to L=128 remain near saturation for most models, and distractor noise causes only model-specific degradation at high D , while the inconsistent premise condition has the lowest accuracy. Instruction-tuned models, including larger 24B–27B baselines, score essentially zero on inconsistent premise sets, while reasoning-tuned models recover only partially and remain below reliable accuracy. Within this benchmark, inconsistent premise sets are associated with the strongest degradation under the benchmark’s abstention-based scoring rule.

[AI-92] Explainable Federated Learning via Global-Local Attribution Alignment ICML2026

链接: https://arxiv.org/abs/2611.20445
作者: Dawood Wasif,Terrence J. Moore,Chang-Tien Lu,Jin-Hee Cho
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 26 pages, 2 figures, 9 tables. Published in Proceedings of the 43rd International Conference on Machine Learning (ICML 2026)

点击查看摘要

Abstract:Federated learning enables on-device training without centralizing data, yet existing systems still struggle to provide explanations that are both locally faithful and globally consistent under strict privacy and bandwidth constraints. Prior approaches either keep explanations siloed across clients, transmit heavy or sensitive artifacts, or replace expressive task models with interpretable surrogates that sacrifice accuracy. We propose xFedAlign, a model-agnostic framework that decouples task optimization in parameter space from explanation coordination in a compact group space. Each client distills a lightweight surrogate to produce private, per-class top-k attribution artifacts, which are robustly aggregated by the server into a Global Explanation Prior that softly aligns client explanations without constraining task learning. Across image, text, and tabular benchmarks with IID and non-IID partitions, xFedAlign matches FedAvg accuracy while consistently reducing explanation drift and improving deletion and insertion AUC relative to Local-XAI, FedAttr-Agg, and Fed-XAI, with only a few kilobytes of additional communication per round. Privacy and robustness evaluations further demonstrate reduced membership inference advantage and increased resistance to attribution poisoning, enabling consistent and trustworthy explanations in federated learning.

[AI-93] FinVerify: A Verification Framework for Financial Agents NEURIPS2026

链接: https://arxiv.org/abs/2611.20423
作者: Mao Guan,Qian Chen,Ge Meng,Xiao-Yang Liu
类目: Artificial Intelligence (cs.AI)
备注: 15 pages, 2 figures. Accepted to the NeurIPS 2026 Workshop on Who Verifies the Agents? Toward Reliable Agent Development

点击查看摘要

Abstract:Financial agents produce numerical claims that can be wrong even when their calculations execute correctly, due to mismatched reporting periods, unsupported sources, incompatible units, or incorrect financial definitions. We introduce FinVerify, an evidence-grounded framework for verifying financial agents’ numerical claims. Based on an analysis of 190 incorrect trajectories from the Agent Finance Reasoning dataset, we define five error categories–temporal integrity, source fidelity, comparability, accounting integrity, and financial-model fidelity–and operationalize them as verification protocols. An LLM structures claims and supporting evidence, while deterministic validators check financial requirements; when authoritative formulas are available, Z3 checks consistency with grounded inputs. FinVerify accepts only if all applicable checks pass, rejects proven violations, and abstains when evidence is insufficient. We evaluate the framework against a Direct baseline and an LLM judge across five fixed-candidate benchmarks. On Agent Finance Reasoning, accepted-claim precision increases from 56.4% for the LLM judge to 76.7% at 81.2% coverage. On XBRLFiling, FinVerify achieves 99.8% accepted-claim precision and 97.0% decision accuracy; on FinanceBench-67, it reaches 87.9% accepted-claim precision versus 69.7% for the judge. Improvements are smaller on FinQA and TAT-QA, where authoritative evidence is limited. These results suggest that explicit, evidence-backed verification can improve financial-agent reliability while exposing trade-offs between correctness, false rejection, and abstention.

[AI-94] A Structural Framework for Succession: Continuation Recognition and Realization

链接: https://arxiv.org/abs/2611.20422
作者: Ilya A. Shkrob
类目: Artificial Intelligence (cs.AI)
备注: 45 pages with tables and diagrams

点击查看摘要

Abstract:Succession asks whether an entity encountered later is the same entity that entered an arrangement earlier. For an artificial system that is fine-tuned, distilled, or copied, it has to be answered. We look for a mathematical structure weak enough to cover any such case and strong enough to support general theorems, so that results follow from the structure’s conditions rather than from a particular pipeline or criterion. Succession is modeled on state/modification graphs with relational lifts that compose along paths. Within this framework we separate four questions that are often conflated: which continuations a representation admits, which descendants are compatible with a modification, which are recognized as continuing the identity, and which are realized. The framework allows one to address systematically questions of symmetry, computational complexity, recoverability, and robustness to partial realization, and we establish what the structure settles in each case before any particular criterion of identity is chosen.

[AI-95] Attention Sinks as Spectral Spikes: A Mechanism Analysis of Gated Attention NEURIPS2026

链接: https://arxiv.org/abs/2611.20403
作者: Seojin Kim,Yehjin Shin,Noseong Park
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: The 40th Conference on Neural Information Processing Systems (NeurIPS 2026)

点击查看摘要

Abstract:Transformer attention exhibits attention sinks, where a small subset of source tokens absorbs disproportionate attention mass across layers. Recent work shows that post-SDPA output gating strongly reduces attention-sink behavior, while separate studies link attention sinks to gradient sinks and massive activations through backward training dynamics. Yet how the same gate affects forward sink structure, backward gradient concentration, and training-time representation signatures remains unclear. We address this gap by formulating attention sinks as sink-induced low-rank modes in token-space covariance, with directions aligned to high-mass attention columns. Under this view, post-SDPA gating acts as a mode-dependent contraction: it attenuates sink-induced spectral energy in the forward pass and locally reduces value-path gradients routed through sink columns in the backward pass. Empirically, random-matrix diagnostics show that gating reduces outlier mass, weakens sink-subspace alignment, and increases effective rank, while backward analyses show active-gate attenuation of first-token value gradients. Training-time diagnostics further show that gating suppresses the co-emergence of attention concentration, value-gradient amplification, and representation dominance.

[AI-96] Making Koopman Control Guarantees Hold on Robots

链接: https://arxiv.org/abs/2611.20402
作者: Abulikemu Abuduweili,Changliu Liu
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Systems and Control (eess.SY)
备注: this https URL

点击查看摘要

Abstract:A Koopman controller can stabilize its learned latent-space linear model yet fail to stabilize the nonlinear plant. The learned model propagates predictions in latent space, while the deployed controller repeatedly re-encodes the measured physical state. We show that nominal closed-loop stability transfers to the nonlinear plant when the learned and true closed-loop dynamics agree to first order on the embedding’s tangent space at the target. We derive a robustness margin for approximate matching and show that a stable open-loop latent model cannot exactly match an unstable equilibrium. We introduce Tangent-Invariant Koopman (TIK), which estimates a local plant linearization from offline transitions and enforces first-order matching to this estimate. TIK recovers the fitted LQR gain at the target, while a bilinear predictor captures state-dependent actuation during large maneuvers. Across four simulated unstable benchmarks, all 40 Standard Koopman controllers have stable latent closed loops, but their deployed nonlinear systems are locally unstable. In contrast, all 40 TIK controllers locally stabilize their respective nonlinear plants. An offline stability test using empirical curvature estimates identifies stable TIK controllers on two cart-pole systems, but is inconclusive under high noise. Under explicit model-error and curvature bounds, an invariant terminal set guarantees constraint satisfaction after entry under local feedback. TIK achieves 88% constrained cart-pole swing-up success with a 0.25 ms planning budget and controls simulated seven-joint arms within 1-5 ms per step.

[AI-97] Implicit-Reasoning -Enabled Elastic Semantic Communications for AI-Agent Networks

链接: https://arxiv.org/abs/2611.20399
作者: Kaiwen Yu,Gang Wu,Jiancun Fan,Saviour Zammit
类目: ystems and Control (eess.SY); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI-agent networks introduce new communication demands driven by proactive interactions and autonomous task execution. Existing adaptive semantic communication schemes primarily adjust semantic representations and resource configurations within predefined processing scales and capability sets, which may be insufficient for evolving task requirements. To address this gap, we introduce elastic semantic communications (ESC), a task-driven framework that extends semantic adaptation to the runtime execution and participation of task-supporting capabilities. ESC comprises two complementary dimensions: vertical elasticity adjusts the execution intensity of existing capabilities, while horizontal elasticity reconfigures the participating capability set. The framework connects these dimensions through the expected marginal value of additional information and computation. To realize ESC, we propose an implicit-reasoning-enabled device–edge mechanism that supports variable-depth local inference through shared-parameter state updates and edge-side continuation through wireless transfer of intermediate reasoning states. Experiments on Sudoku-Extreme demonstrate that the benefits of additional inference depend on wireless conditions, while task-aware edge invocation can preserve task accuracy with substantially lower state-transfer overhead.

[AI-98] Deploying a Control-Logic LLM Assistant in Air-Gapped Power Plants: Lessons in Source Authority and Bounded Retrieval AAAI2027

链接: https://arxiv.org/abs/2611.20398
作者: Yiwei Fu,Nicholas Gabriel
类目: Artificial Intelligence (cs.AI)
备注: Extended version of a paper accepted at IAAI-27 (AAAI 2027). 8 pages, 2 figures, 3 tables

点击查看摘要

Abstract:Deploying Large Language Model (LLM) assistants in industrial operational technology environments poses constraints largely absent from cloud applications: site policies restrict engineering and operational data from leaving the plant, on-premises compute is limited, and incorrect causal attribution can mislead incident response. We report lessons from deploying a control-logic evidence layer that underpins a broader on-premises operator assistance system at two gas power plants. Two prototypes revealed where generative models added avoidable uncertainty and cost. The first used a Vision Language Model (VLM) to reconstruct control topology from PDF printouts routinely exchanged by engineers. Across the full 6,979-page document, extraction reached 0.51 connection precision and required 3.1 hours on 4xH200 GPUs; consultation with controls engineers then revealed a native structured export from the authoring system. The second used an LLM tool-calling agent to plan repeatedly over eight predefined graph operations. In a controlled benchmark of 40 template-generated questions, replacing this agent with explicit execution plans reduced mean token use by 20x and median latency by 3x, with identical retrieved node sets across repeated runs. The deployed design ingests the authoritative export into a versioned knowledge graph, executes bounded graph operations in fixed software, and confines the local LLM to fallback intent classification and evidence-grounded explanation. At one site, the graph contains 221K nodes across seven controllers. Our experience suggests a practical division of labor for high-consequence AI: LLMs interpret natural-language requests and explain results, while deterministic software constructs and retrieves technical evidence grounded in authoritative sources. Abstention, entity disambiguation, graph versioning, and read-only integration provide additional safeguards.

[AI-99] Pretty Linear Liars: Unregularized Probes Favor Intervention-Unfaithful Features NEURIPS2026

链接: https://arxiv.org/abs/2611.20371
作者: Karl Vilhelmsson Emneby,Julian Yocum,Cameron Allen
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 13 pages, 9 figures. Accepted at the NeurIPS 2026 Workshop on Interpretability as a Science (InterpScience)

点击查看摘要

Abstract:Linear probing is a common technique in mechanistic interpretability despite the fact that the decodability of a given feature does not necessarily imply the network uses that feature. Intervention faithfulness, not merely decodability, is the appropriate criterion for whether a feature serves a defined active role in the model. Meanwhile, probe regularization has been used for feature discovery, but the impact of regularization on feature causality remains unclear. We study probe regularization systematically, starting with unregularized ordinary least squares (OLS) regression and considering increasing amounts of regularization via ridge regression. We show theoretically that OLS probes amplify low-variance directions in activation space that have spurious correlations with the target feature, whereas regularization reduces influence from these directions. Empirically, we find in a toy transformer and Llama-2-7B that intervention faithfulness increases up to a point, and then decreases again, and that the optimal regularization strength leads to features bearing striking similarity (in both direction and variance) to those learned by optimizing for intervention outcomes directly. Together, these results provide mechanistic justification for regularizing regression probes.

[AI-100] Residual-Tail-Informed Flow Matching for Time Series Anomaly Detection

链接: https://arxiv.org/abs/2611.20334
作者: Shyam Sundar Saravanan,Tie Luo,Sanjay Madria
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: IEEE International Conference on Big Data 2026 (December)

点击查看摘要

Abstract:Time series anomaly detection (TSAD) models are often trained on nominal data using average-risk objectives, which emphasize dominant normal patterns. However, real time series can contain rare but valid nominal regimes that are underrepresented in the training distribution. We show that data-space flow matching can over-penalize such rare-normal windows and assign them anomaly scores comparable to those of true anomalies. To address this, we propose Residual-Tail-Informed Flow Matching (RT-FM), an unsupervised framework that uses a temporal variational autoencoder to estimate residual latent tailness and converts it into bounded, warmup-annealed weights for training a data-space flow matching model. Experiments on seven TSAD benchmarks show that RT-FM improves rare-normal coverage while preserving strong anomaly discrimination, substantially reducing false alarms on such windows and resulting in improved aggregate detection performance. The code is available at: this https URL

[AI-101] LENS: Replacing LayerNorm and Softmax in Transformers with Learnable Normalization States

链接: https://arxiv.org/abs/2611.20331
作者: Yuhao Liu,Salim Ullah,Akash Kumar
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Hardware Architecture (cs.AR)
备注: Source code is available at: this https URL

点击查看摘要

Abstract:LayerNorm and attention Softmax are critical operations in Transformers. However, they cannot be efficiently mapped to General Matrix Multiplication (GEMM) computations on hardware. Therefore, modern hardware accelerator designs, such as NPUs on ASIC/FPGA, consume a large amount of hardware resources or chip area to implement their specific arithmetic units. Building on previous work about LayerNorm, Softmax, and their alternatives, we introduce Learnable Normalization States (LENS), a framework that reformulates LayerNorm as tiny MLPs by abstracting a common structure from related methods and extends this method to Softmax, which moves more of the computation in Transformers to GEMM as a hardware-friendly architecture. We evaluate LENS on image classification with Vision Transformers (ViT) and autoregressive language modeling with Pythia models, including generalization tests on WikiText and LAMBADA. Across these tasks, LENS matches or exceeds baseline performance in several vision settings and retains most of the baseline performance in the remaining vision and language experiments. These findings may suggest that hand-designed normalization and attention functions are not the only choice as alternatives to LayerNorm and Softmax. LENS also provides a basis for exploring Transformer architectures for resource-limited accelerators.

[AI-102] NF-VLA: A Normalizing-Flow Baseline for Vision-Language-Action Learning

链接: https://arxiv.org/abs/2611.20314
作者: Ruichen Deng,Qinyu Zhao,Yong-Hyun Park,Mutian Tong,Liang Zheng,Jiatao Gu
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注: Accepted at CoRL 2026 G2A Workshop

点击查看摘要

Abstract:Autoregressive normalizing flows (NFs) provide a tractable density for learning and sampling continuous robot actions. We show that a single maximum-likelihood objective can train vision–language–action (VLA) policies competitive with leading methods on manipulation benchmarks. We present NF-VLA, which combines conditional autoregressive normalizing flows with score-based action refinement. Its deep–shallow architecture concentrates visual, language, and proprioceptive prefix conditioning in a high-capacity deep block, while unconditioned shallow blocks provide a shared invertible action transformation. The vision–language backbone and flow are jointly trained by exact maximum likelihood on Gaussian-smoothed action chunks. The same learned density supplies a single score-based correction at inference, refining generated actions to account for training-time smoothing without a separately trained denoiser. Trained on LIBERO from a pretrained vision–language backbone without robot pretraining, NF-VLA achieves 98.35% success across the four standard suites and 72.6% on LIBERO-Plus using the same checkpoint without further fine-tuning. Ablations show that deep-only prefix conditioning and score-based refinement both improve closed-loop success. Together, these results establish carefully structured autoregressive flows as a competitive approach to VLA control, with one learned density supporting policy training, action generation, and refinement.

[AI-103] Benchmarking LLM Inference on Desktop-Scale NVIDIA Grace Blackwell Clusters

链接: https://arxiv.org/abs/2611.20289
作者: Kevin Bhimani,Christian Escobar,Daniel Mihovch,Rudolf Eigenmann,Sunita Chandrasekaran
类目: Distributed, Parallel, and Cluster Computing (cs.DC); Artificial Intelligence (cs.AI)
备注: 24 pages, 16 figures, 8 tables. Code and data available at this https URL Copy

点击查看摘要

Abstract:Two NVIDIA DGX Spark GB10 systems connect directly over 200 GbE to serve larger language models and increase inference capacity, but the benefits depend on how they are deployed. We benchmark 17 open-weight checkpoints spanning dense and mixture-of-experts architectures, 8B to 235B parameters and four weight precisions on one and two RDMA-connected systems. Using vLLM, we compare tensor parallelism (TP=2), pipeline parallelism (PP=2) and a two-replica baseline at concurrency 1 to 64, with three independent launches per configuration. TP=2 delivers a median 1.59x peak-throughput gain over one Spark, versus 1.39x for PP=2 and has higher throughput in 93 of 96 matched comparisons with separated 95% Welch intervals. It reduces unloaded inter-token latency to 0.52x to 0.73x of one Spark, whereas PP=2 provides little reduction. TP=2 also achieves the highest goodput under a strict 100 ms inter-token latency objective wherever both strategies are supported and the objective is met. For models fitting on one Spark, TP=2 generally favors low-concurrency serving, while replication offers higher throughput at high concurrency. One Spark uses less GPU energy per output token at matched concurrency. Across three precision ladders, NVFP4 reduces unloaded latency more than adding a second Spark at BF16. NCCL microbenchmarks and comparisons with an earlier TCP campaign suggest that collective communication overhead contributes to the advantage of TP=2. These results provide deployment guidance for desktop-scale inference.

[AI-104] Protecting Safety during Reasoning Training with Low-Rank Fisher Regularisation

链接: https://arxiv.org/abs/2611.20288
作者: Neel Rajani,Ivan Titov
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 18 pages, 7 figures

点击查看摘要

Abstract:Large language models are trained to have safety guardrails, but these can be compromised while fine-tuning for new capabilities, such as reasoning. One way to mitigate this degradation is to modify the fine-tuning objective to preserve safety behaviour, but doing so effectively and efficiently remains challenging. We revisit Elastic Weight Consolidation (EWC), which penalises parameter updates based on sensitivity estimates from the diagonal of the Fisher information matrix (FIM). While EWC provides some protection, the diagonal approximation ignores interactions between parameters. We address this limitation by exploiting recent findings that safety-relevant behaviour is located in a low-rank subspace within models. Our method, Subspace-Projected Elastic Weight Consolidation (SP-EWC), projects gradients onto this subspace to capture non-diagonal curvature efficiently. We investigate three different ways of identifying the safety subspace: through safety fine-tuning updates (LoRA), safety-related activations (ActSVD), and contrastive representations of safety and utility data (cPCA). We find that this choice substantially affects performance, with SP-EWC using cPCA comparing favourably against existing methods by better maintaining safety guardrails without introducing excessive over-refusal.

[AI-105] GNN-Based Deep Reinforcement Learning for Coastal Emergency Response under Extreme Water Levels NEURIPS2026

链接: https://arxiv.org/abs/2611.20268
作者: Marcello Sano,Samuele Casagrande,Davide Ferrario,Sebastiano Vascon,Silvia Torresan,Andrea Critto
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 6 pages, 1 figure, 1 table. Accepted at the Tackling Climate Change with Machine Learning workshop at NeurIPS 2026. Code: this https URL

点击查看摘要

Abstract:When a storm surge approaches a coastal town, the emergency plan must be put into action quickly using water-level forecasts, flood extent, road access, vulnerable residents and shelters. Bringing this information together under time pressure is slow and depends on the duty officer’s expertise. This paper is one step in a wider research line from data to safe decisions: machine learning to understand the hazard, reinforcement learning and planning to choose actions, large language models (LLMs) to explain and cross-check them, and an assurance layer that keeps people in charge. Here we address the action step on a real place. We ask whether a policy can learn the timing of emergency actions from decades of real events. Lido di Venezia is represented as a heterogeneous graph built from public elevation, road, building and census data and driven by hourly tide-gauge records from 1983 to 2024, with seven actions taken from the municipal emergency plan. A PPO policy is trained on this graph and compared, on storm windows unseen in training, with a rule-based system simulating the officer’s discretionary decisions. On 18 historical test scenarios, PPO achieves a similar reduction in population exposure to the rule. With uncertain forecasts and limited protection it again performs similarly but does not outperform the tuned rule, and removing graph message passing does not substantially reduce performance. PPO therefore learns an effective protective response from an outcome-based reward alone: it is given the plan’s actions and when each is allowed, but not the thresholds that tell the comparator when to act. Next steps test harder edge cases, move closer to a digital twin of the territory, and replace the simulated officer with assessment by emergency-management stakeholders, supported by LLM-based interpretation of each recommendation.

[AI-106] Microfitting: Continual Learning via Cached Test-Time Training

链接: https://arxiv.org/abs/2611.20258
作者: Nick Ryan,Peter A. Devine,Ben Sirb,Hannah Sansford,Alex Chiocchi
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Continual learning is a promising approach to specialize language models to an organization’s private data. However, current methods exhibit a number of drawbacks including catastrophic forgetting, where new training updates overwrite prior knowledge. We present microfitting, which trains and caches a small adapter for each document. Each adapter is trained once, offline, on synthetic questions targeted at demonstrated base-model shortcomings and reused whenever that document is queried. This method borrows from test-time learning literature and sidesteps the problem of continual learning through parameter isolation, allowing the system to specialize on new data without overwriting prior knowledge. The process that generates adapter training data also improves over time, revising its recipe from previous outcomes and fresh examples of the target workload. On DeepSeek V4 Flash, microfitting substantially improves accuracy over the base model. In Enterprise Document Operations (EDO), an open-book QA environment over real-world enterprise documents, our method achieves a 38.2% relative improvement over base-model performance. Microfitting makes continual specialization practical and cost-efficient: a frozen model can adapt to an evolving private corpus overnight on local hardware, without modifying the base model or previously cached adapters.

[AI-107] Right Goal Wrong Body: Re-grounding Frozen VLAs under Actuation Shift

链接: https://arxiv.org/abs/2611.20244
作者: Umer Siddique,Peilang Li,Yongcan Cao
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Accepted at the SPAIS Workshop at CoRL 2026

点击查看摘要

Abstract:Vision-language-action (VLA) models are capable of producing desired robotic behaviors in ideal settings. However, they often fail when the robot’s actuation changes at deployment. For example, remounting, mechanical wear, or command latency changes can lead to failed tasks and unsafe behaviors. To address the inevitable actuation shift, we propose ReGround, an inference-time method that keeps the VLA frozen and re-grounds its intended motion. ReGround learns the deployed robot’s dynamics from task-agnostic motor babbling, and then tracks the policy’s intended motion while staying close to the policy’s own plan. ReGround activates only upon detecting a shift, and needs no rewards, demonstrations, or retraining. We show via two VLAs, eight LIBERO tasks, and five shift settings that ReGround can restore near-oracle success, such as raising \pi_0.5 's success from 0.16 to 0.90 under 400 ms latency, while remaining safer than a linear-calibration baseline. Project website and open-source code: this https URL.

[AI-108] A Matrix with Steps: An Episodic and Procedural Memory Method for Tool-Using Agents and Robots

链接: https://arxiv.org/abs/2611.20230
作者: Jürgen Paulhart([core] Informationstechnologie / Eldric),Peter Schüller(NeuraMix AI GmbH and Technische Universität Wien)
类目: Artificial Intelligence (cs.AI)
备注: 17 pages, 5 figures

点击查看摘要

Abstract:Agents that call tools and robots that execute tasks are challenged by two shortcomings: they repeat what failed before, and they learn from runs that they only believe succeeded. We describe the construction of a memory that addresses both, built for a production agent platform. Every run is recorded as an episode, and only an outcome confirmed by an external check, a test, a tool result or a person counts; the agent’s own claim never does. What is learned from episodes is kept in a memory with two parts: an associative matrix, written with the delta rule, which finds the procedure that applies to a situation; and a section of ordered, addressable step records that answers what comes after a given step, where a step is a goal of the agent, not a motion. Because the matrix sums its associations, it keeps no order and cannot return or delete a single one; it serves as the finder, and order, outcome counts and deletion live in the records. We describe how execution outcomes are stored, how episodes are recalled and used for solving tasks, and the file format that stores the memory. In simulation, a robot that judged its own success and stopped when it believed it was done reported 96% of its tasks as done and solved 35% ; with external verification it solved 79% . In a store of 100000 synthetic episodes ( 457 MB on disk), the oldest episode was still found first, no episode from another person’s memory was returned, and a success vouched for only by the model was never recalled; a lookup took 16 ms on a laptop and 40 ms on a server node (median). Task outcomes were measured up to about 10000 confirmed records, where the exact records solved every revisit at the first attempt.

[AI-109] SWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations

链接: https://arxiv.org/abs/2611.20218
作者: Xinping Lei,Minghao Liu,Ruijie Zhu,Xiaohui Hu,Xingqiao Li,Ao Zhang,Wenhao Chai,Han Li,Tao Wang,Jingyuan Zhang,Enduo Zhao,Yunfeng Shi,Yiming Liang,Zetao Yang,Xingjian Wang,Yifan Yao,Xinkai Ma,Yingru Li,Tianhan Yang,Huan Zhou,Weiran Shi,Shen Yan,Jiaheng Liu,Stephen Huang,Eli Cheung,Weihao Xuan
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:A coding agent working within a software organization must establish what to build, collaborate with colleagues, and verify what it delivers. We introduce SWE-CoWork, a benchmark comprising 101 development assignments across 101 real repositories, 11 programming languages, and 12 application domains. Each assignment places an agent in a simulated engineering organization with 26 colleagues who have distinct responsibilities, bounded knowledge, and controlled collaboration behavior. Starting from a brief handoff, agents must recover requirements distributed across colleagues and workplace records, reconcile conflicting evidence, and implement changes evaluated by private executable tests. We evaluate eleven models and find a substantial gap between partial implementation and complete task delivery. The two leading models satisfy approximately 63% of fail-to-pass (F2P) tests but complete fewer than 8% of assignments. Providing complete requirements upfront improves F2P by approximately 16 percentage points for Seed Evolving and GPT-6 Astra, yet leaves complete task success unchanged. Requirement-level trajectory analysis reveals failures throughout the delivery process: Claude-Opus-5.5 directs a mean of 93% of first questions about queried cards to their accountable owners across runs but obtains only 64% of colleague-held requirements. Deferred questions go unfollowed, implementations remain incomplete even after requirements are acquired, and 31% of requirements reported as completed fail their tests. These findings highlight sustained requirement tracking, from clarification through implementation, verification, and handoff, as a key challenge for reliable autonomous software engineering.

[AI-110] Paying Creators for Training Data Can Increase AI Displacement

链接: https://arxiv.org/abs/2611.20204
作者: Lijuan Luo,Emaad Manzoor,Nathan Yang,Fan Yao
类目: Computer Science and Game Theory (cs.GT); Artificial Intelligence (cs.AI); Theoretical Economics (econ.TH)
备注: 53 pages, 6 figures. Previously circulated under the title “Platform Design When Creators Train Their AI Substitutes”

点击查看摘要

Abstract:Generative AI increasingly competes with creators while using their content as training data. In response, policymakers have called for platforms to compensate creators for training data use and allow opting out. We study how such policies affect creator displacement in a model of a platform where creators choose whether to contribute training data and consumers choose between purchasing from creators and prompting generative AI trained on their content. When compensation is required, the low-quality creator chooses to remain solely as a training data supplier while AI fully displaces her demand, despite AI quality and the price the platform charges for AI-generated content being identical with and without compensation. Without compensation, the risk of losing creators and their training data constrains the platform to set the price of creators’ content to preserve their demand. This constraint also prevents further AI displacement as AI improves, with the platform raising AI prices while keeping creators’ prices and demand unchanged to prevent exit. Compensation increases creators’ payoffs and total surplus but reduces consumer surplus. Our findings highlight how compensation policies for training data can shape creators’ economic roles in a future with more capable generative AI.

[AI-111] How Specification Explicitness Shapes Failure in LLM -Generated Code ICSE2026

链接: https://arxiv.org/abs/2611.20195
作者: Elmira Onagh,Maleknaz Nayebi
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: Accepted to ICSE 2026

点击查看摘要

Abstract:When LLMs turn requirements into code, every underspecified constraint becomes behavior that the model must decide during generation. We test when specifications move beyond structure and begin to control generated behavior. In a controlled experiment, 72 participants implemented and revised Python tasks with an LLM under three levels of specification explicitness. The results show a clear break between lightweight structure and explicit behavioral obligations. Lightweight structure reduced hidden-test failures by only 8.7%. Requiring participants to specify constraints, invariants, negative requirements, edge cases, tie-breaking rules, and exception handling cut failures by 47.2% and reduced the number of distinct failure modes by 20.5%. The largest drops were in failure modes tied to implicit obligations. Invariant violations fell by 66.1%, and unsupported assumptions fell by 59.5%. After a requirement change, specifications with explicit behavioral obligations produced regressions in more submissions, but each affected submission lost less previously correct behavior. Explicit specifications also changed how developers judged generated code. Without explicit obligations, participants overestimated generated-code correctness by about 23 percentage points. With explicit obligations, they produced better code but judged it more cautiously. Thus, the lesson is sharper than “better specifications help.” Lightweight structure barely helped, while explicit behavioral obligations nearly halved failures, made regressions shallower, and improved calibration.

[AI-112] RoboKit: Accelerating Parallel Differentiable Optimization for Robot Learning

链接: https://arxiv.org/abs/2611.20185
作者: Jianglong Ye,Xuanbin Peng,Conor Mc Gartoll,Changwei Jing,Lai Wei,Jiali Chen,Xueyan Zou,Xiaolong Wang
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Accepted to CoRL 2026. Project page: this https URL

点击查看摘要

Abstract:We present RoboKit, a high-performance, general-purpose toolkit for parallel differentiable optimization in robotics, built on three principles: (i) Performance and flexibility: GPU-parallel multi-seed solvers pair sampling with versatile optimizers, running 2-60 \times faster than task-specific baselines across seven tasks; (ii) Geometry, rendering and physics: hybrid and coarse-to-fine representations accelerate collision, differentiable rendering and simulation, optimizing 4096 grasps in 2.97 s; (iii) Integration with learning: costs are differentiable and simulation is parallel, so RoboKit both distills optimization into policies and trains RL; distilled policies run 9-240 \times faster at comparable accuracy, and RL reaches a GPU baseline’s scores up to 2.1 \times sooner on the same simulator.

[AI-113] CapCover Filter: Semantic Approximate Membership Queries over Embeddings

链接: https://arxiv.org/abs/2611.20132
作者: Dvir David Biton,Shay Hahiashvili,Nadav Keren,Gil Einziger,Roy Friedman
类目: Databases (cs.DB); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Determining whether an embedding is sufficiently similar to any vector in a vector store without performing a nearest-neighbor search can substantially improve query throughput in vector-store systems. We study Semantic Approximate Membership Queries (SAMQs), which allow queries to be rejected when no indexed vector meets the requested similarity threshold, while guaranteeing no false negatives. We introduce CapCover Filter, a geometric data structure that summarizes a set of embeddings using spherical caps. Its representation is independent of the query threshold, allowing a single filter to support arbitrary thresholds without reconstruction. We evaluate alternative construction and quantization methods, compare CapCover Filter with competing filters, and measure its impact on query throughput. Across five real-world datasets, at a cosine similarity threshold of 0.95, CapCover Filter eliminates approximately 20–100% of index accesses while using only 2% of the raw (FP32) embedding memory. When placed before a FAISS flat index, it improves query throughput by up to 382% and by approximately 116% on average across the five datasets, while returning the same range-search result. We also evaluate lower similarity thresholds. Our results show that CapCover Filter can substantially improve vector-store throughput when similarity thresholds are tight, retrieval is expensive, and indexed embeddings are highly clustered. We release an optimized implementation of CapCover Filter and its evaluation framework. Subjects: Databases (cs.DB); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.20132 [cs.DB] (or arXiv:2611.20132v1 [cs.DB] for this version) https://doi.org/10.48550/arXiv.2611.20132 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-114] CECAR: Cache Expert Co-Aware Routing Accelerates On-Device Inference of MoE LLM s NEURIPS2026

链接: https://arxiv.org/abs/2611.20119
作者: Byeongju Kim,Hoonki Lee,Byungjun Kim,Inyul Ra,Yongjin Kim,Yongseong Lee,Sangyeob Kim
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted to NeurIPS 2026. 24 pages, 5 figures, 19 tables

点击查看摘要

Abstract:Despite growing interest in on-device LLM deployment, large-scale Mixture-of-Experts (MoE) models remain impractical on resource-constrained devices, as sparse expert activation still requires all expert weights to be memory-resident. To address this, we identify three key observations for on-device MoE inference: (a) cache hit rates are fundamentally bounded even under Belady’s optimal policy; (b) these bounds are insufficient for low-latency on-device inference; and © prior cache-aware routing based on binary cache presence can induce routing instability. Based on these insights, we propose CECAR (Cache Expert Co-Aware Routing), a MoE inference system that jointly considers expert routing and cache management. Unlike prior cache-aware routing methods, CECAR uses an ML-based cache policy to predict the expected reuse distance of each expert and prioritizes routing and execution toward experts with smaller predicted reuse distances. On the Qwen3-30B-A3B model, CECAR achieves 12.5 tokens/s on a consumer-grade GPU with only 16 GB of VRAM, where non-resident experts are fetched from SSD, delivering a 4.5x speedup over conventional MoE inference and exceeding the decode speed of baselines that keep the entire model fully resident in DRAM. Code is available at this https URL.

[AI-115] Locational Marginal Pricing of Compute: Co-optimizing Compute and Energy for Transparent Data Center Flexibility

链接: https://arxiv.org/abs/2611.20116
作者: Vladimir Pyltsov,Teng Zeng
类目: ystems and Control (eess.SY); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Demand from data centers and AI compute is growing rapidly and is increasing stress on the power grid. This load could be flexible through user adjustments and provide relief during energy scarcity events, yet today its flexibility is provided through utility call-ups or system-operator load shedding, and is invisible to the users whose workloads are being curtailed. In this work, we propose to quantify and couple the prices of compute and energy in a single optimization problem. The resulting compute price reflects the marginal cost of the binding constraints and decomposes into components indicating which grid, power, hardware, or service requirement causes a price surge. The economic interpretation updates the price with real-time conditions and gives users a signal for potential flexibility provision. The co-optimization ultimately connects compute and energy markets while providing a signal for relief of stress on both the grid and the compute resources.

[AI-116] When Valid Decisions Are Not Enough: Evaluating Structural Constraints in Knowledge-Guided Generative Systems

链接: https://arxiv.org/abs/2611.20111
作者: Josef Pavlíček,Petra Pavlíčková,Martin Molhanec
类目: Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
备注: 16 pages, 4 tables, 1 figure

点击查看摘要

Abstract:Knowledge-guided generative systems use explicit domain constraints to improve decision consistency and alignment with human-created solutions. However, locally valid decisions do not necessarily preserve contextual relationships in human decision-making. This paper examines when structural constraints improve or degrade agreement with human reference solutions. We introduce HARMONIX, an experimental framework using automated musical harmonization as a controlled testbed for knowledge-guided decision-making. Four configurations share the same Viterbi-based optimization mechanism: a baseline, repetition constraints, functional constraints, and their combination. Experiments cover 19,745 records and 143,634 motif pairs across training and validation partitions. Combined constraints improve reference root-interval agreement by approximately five percentage points in stable structural contexts but reduce it by nearly nine percentage points in contexts involving transposition tracking. These opposing effects are consistent across both partitions, while overall exact chord agreement changes only marginally. Contexts are identified retrospectively using reference annotations; no adaptive constraint-selection mechanism is evaluated. The findings demonstrate that interpretable constraints can introduce systematic context-specific failures concealed by aggregate accuracy metrics. They highlight the need to evaluate relationships between decisions and the contextual applicability of constraints in knowledge-guided generative systems. Comments: 16 pages, 4 tables, 1 figure Subjects: Artificial Intelligence (cs.AI); Computers and Society (cs.CY) Cite as: arXiv:2611.20111 [cs.AI] (or arXiv:2611.20111v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.20111 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Josef Pavlicek Ph.D. [view email] [v1] Sun, 15 Nov 2026 16:41:56 UTC (48 KB)

[AI-117] MT-RGCN-DTI: A Multi-Task Relational Graph Convolutional Network for Drug-Target Interaction Prediction

链接: https://arxiv.org/abs/2611.20109
作者: Dheeran Sankaran,R Yadunandan,Daniele Santoni,Nimisha Ghosh
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Prediction of drug-target interaction (DTI) has become a very crucial part of drug discovery. With the increase in the data volume, it has become very important to improve prediction accuracy by integrating multiple data sources. In this regard, utilising multiple resources in the form of heterogeneous information network to address DTI prediction can be rewarding. In this work, we propose a heterogeneous information network based multi-task DTI prediction model which leverages the advantages of relation graph convolutional network (RGCN). To achieve this, the model employs drug-target interaction network, drug-drug similarity network and protein-protein similarity network to construct a heterogeneous information network. Moreover, we use a knowledge graph embedding scoring functions to model interactions between drug and targets, specifically DistMult and ComplEx as decoders over the learned node representations. We performed extensive experiments on multiple benchmark datasets and compared with state-of-the-art methods. We observed that the proposed model shows competitive performance, thereby confirming the effectiveness of our approach. The codes and datasets are available at this https URL.

[AI-118] SAFER: Navigating the Exploration-Exploitation Dilemma in World-Model-Based Planning for Autonomous Driving

链接: https://arxiv.org/abs/2611.20107
作者: Yansong Qu,Zilin Huang,Zihao Sheng,Sikai Chen,Samuel Labi
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:An autonomous vehicle must foresee the consequences of its actions, especially in the rare, time-critical situations where crashes happen. A world model offers this foresight, but only within the coverage of its training data, which poses an exploration-exploitation dilemma. Within this coverage a planner can hardly improve on the data, and beyond it the model’s errors mislead it. Logs of competent driving rarely contain the failures that would widen the coverage, and agents that learn from interaction gain them only by failing themselves, without being shown what to do instead. We propose SAFER, a framework that learns offline from failures that others make, judged by an expert, and keeps learning online from its own driving. Offline, partially trained agents and scripted, recoverable perturbations Stage failures in simulation, and an expert Annotates every visited state with its action and value, so that much of what the planner explores is covered before it drives. In planning, a chunked latent world model Foresees the consequences fast enough for real-time use, and a risk-aware sampling-based planner Evaluates candidate maneuvers by their return minus a bounded collision penalty. Online, a loop continuously Refines the model wherever this driving leaves the coverage of the offline data. In closed-loop evaluation against the compared methods, SAFER reaches the highest route completion and reward per step on the CARLA Leaderboard routes, 0.95 and 0.77 from a privileged bird’s-eye view, and the highest episode reward in MetaDrive, 53.4 in regular and 49.0 in safety-critical traffic. Project page: this https URL.

[AI-119] Making Preemption Productive in Long-Output LLM Serving

链接: https://arxiv.org/abs/2611.20106
作者: Ziji Shi,Xinyi Wan,Chaoyi Ruan,Guangxing Huang,Min Lin,Jialin Li
类目: Distributed, Parallel, and Cluster Computing (cs.DC); Artificial Intelligence (cs.AI); Operating Systems (cs.OS)
备注: 40 pages, under review

点击查看摘要

Abstract:Reasoning LLMs generate long outputs whose key–value (KV) caches occupy GPU memory for minutes. When memory fills, the server preempts requests, and a preempted request can repeatedly lose and recover its state while producing little new output. We call this behavior KV thrashing. In certain cases, existing systems leave more than a fifth of requests stalled for over a minute. Because preemption cannot be avoided without delaying admission, we ask how to make it productive. We present Stint, a serving system that preserves a preempted request’s work, bounds how long it waits, and gives it fresh output after it returns. Stint stores fixed-length KV snapshots in a contiguity-first CPU pool with an exact-extent allocator, and moves them with fused K/V gather/scatter kernels and double buffering. Age-based readmission brings stalled requests back, and protection gives them a turn of fresh decoding before another preemption. Remaining-length estimates from decoding hidden states guide victim selection. On olympiad-math replays with Qwen3-32B on H200 GPUs, no request pauses for more than 50 seconds under Stint and p99 time per output token falls by 50.7% relative to recomputation in the same engine. The replay as a whole finishes 10.7% sooner (12.0% higher throughput), because the requests that baselines starve no longer finish last.

[AI-120] he Frontier Must Grow - Self-proposed interpreter-labelled training roughly doubles a 4B models accuracy on a fresh exam as its verified frontier grows

链接: https://arxiv.org/abs/2611.20049
作者: Martin Schroder
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:A model can learn from problems it poses itself when an executable oracle supplies the answers. Qwen3-4B proposed Lua 5.1 programs; those it predicted wrong at least once in four samples joined a pool labelled by the interpreter, each round adding twice as many programs as the last (300, 900, 2100 in total), and each round trained a fresh adapter from the base on the pool. On the original exam the procedure met its registered criterion, an item-level test of each adapter against its predecessor. On a 2340-item exam drawn afterwards, accuracy rose from 8.3% to 16.7%; that follow-up’s registered criteria were not met because round 1 gained nothing, while rounds 2 and 3 added 5.0 and 2.9 points on average, both intervals above zero. One fit on the same amount of base-proposed frontier data reached 14.6%, so the rounds’ 2.1-point advantage is inconclusive. Three earlier registered series met no per-lineage criterion, and the same loop showed no gain on repository tasks or for a 27B agent in those configurations.

[AI-121] Security of Deep Research Agents : Attack Surface Defenses Benchmarks and Agenda Seeding against Research Planning

链接: https://arxiv.org/abs/2611.20045
作者: Yang Luo,Guangpu Chen,Yichen Liu,Zhaoyu Chen,Yifeng Gao,Zhiyang Ding,Qingni Shen,Zhonghai Wu(Peking University)
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: 31 pages, 5 figures

点击查看摘要

Abstract:Deep research and agentic search agents autonomously retrieve content from the open web and generate reports with citations, and their outputs are often taken directly as research conclusions. This paper systematizes security research in this area along four stages, namely retrieval, planning, evidence integration, and report generation. We present a unified threat model, a taxonomy of attacks and defenses, and a review of evaluation benchmarks, and we align the area dimension by dimension with retrieval-augmented generation (RAG) poisoning, prompt injection, and generative engine optimization (GEO). Coding 34 primary studies shows that only 26.5% of them involve the planning stage, and none performs an end-to-end poisoning evaluation on commercial products. To address this gap, we propose the agenda seeding attack, in which the attacker writes no instructions and only introduces a fictitious evaluation metric. Across 3,520 sandboxed research runs under the same budget, GEO and prompt injection are nearly ineffective, whereas agenda seeding makes the attacker’s target the top pick in 16% to 28% of reports. Injection detection and query anchoring fail against it, verification prompting hurts utility, and our proposed provenance tagging preserves utility but can be bypassed by multi-site bait. Finally, we propose eight research directions.

[AI-122] owards Agent ic Scientific Discovery in Distributed Systems

链接: https://arxiv.org/abs/2611.20031
作者: Rashmita Soni,Abhirup Ghosh,Soumya Banerjee
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:We investigate when adaptive tool selection helps identify a shared equation across clients with different input coverage. Query-ATLAS (query-conditioned Agentic Tool-based Learning for Aggregating Symbolic Equations) uses an LLM to guide symbolic search from client errors and previous proposals in a simulated distributed system. A four-architecture comparison reveals task-dependent opportunities: relative to fixed PySR-LLM, validation acceptances rise from 3/12 to 7/12 on CRK3 and from 11/12 to 12/12 on oscillator2, while MatSci3 favors the fixed workflow. Across 72 settings in a post-hoc six-task subset of eight development tasks, Query-ATLAS achieves 64 validation and 64 in-distribution passes, versus 62 and 59 for the fixed workflow; both pass 47/60 eligible out-of-distribution tests. The fixed workflow retains lower median errors and cost. Broader query coverage emerges as a promising benefit, motivating task-aware tool selection and client-aware fitting for scientific agents.

[AI-123] Improving fine-grained expressivity of graph neural networks

链接: https://arxiv.org/abs/2611.20020
作者: Yonatan Sverdlov
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:This thesis investigates the expressive power of neural networks for structured data, focusing on graph neural networks and symmetry-aware architectures. We present four main contributions. First, we characterize the generic expressive power of geometric message-passing neural networks on sparse graphs and introduce EGenNet, an architecture achieving the corresponding guarantees. Second, we use representation theory to characterize linear layers equivariant to multiple permutations, including wreath-product symmetries. Third, we study permutation-invariant set representations that preserve containment, establish dimensional bounds and impossibility results, and introduce MASNet. Finally, we propose FSW-GNN, a message-passing architecture combining strong expressivity and bi-Lipschitz stability guarantees with strong empirical performance on long-range graph-learning tasks. Together, these results provide a broader understanding of how neural architectures can preserve the geometric, algebraic, order-theoretic, and metric structure of their inputs. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.20020 [cs.LG] (or arXiv:2611.20020v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.20020 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-124] DBEAST: Logged Bad Experiences as Explanation for Autonomous Driving Agents

链接: https://arxiv.org/abs/2611.20019
作者: Ahmad Alfan Alfian Irfan,Nur Ahmad Khatim,Mansur Arief
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO)
备注:

点击查看摘要

Abstract:Reinforcement learning agents learn from mistakes, yet their decisions can be difficult to trace back to those experiences. We investigate whether a tabular agent’s decisions can be explained by its bad experiences, defined as the training transitions associated with safety or task events. A fixed training log is replayed from the same initial Q-table under three conditions. FULL retains all updates, SKIP omits bad-experience updates, and NEUTRALIZE retains those updates without their event penalties. We illustrate the method on six classic Markov decision processes and evaluate it in closed-loop driving in CARLA. On the full driving task, FULL reached the goal from 33 of 40 test starts, whereas SKIP and NEUTRALIZE reached it from none, indicating that the bad experiences carry decision-relevant information. At the local level, for 47 decisions shortly before a failure of SKIP or NEUTRALIZE, removing or neutralizing a small group of bad experiences, with a median size of one, caused the trained agent to select the failing policy’s action in 46 cases. For this agent, the results indicate that its decisions can be explained by its bad experiences.

[AI-125] GEM-MM: Reducing VLM Hallucination with Entropy-Guided Preference Alignment

链接: https://arxiv.org/abs/2611.20016
作者: Yuanshan Chua,Xuejiao Zhao
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Vision-language models (VLMs) are routinely aligned with offline pairwise RLHF objectives, but multimodal preference data is scarce and treats every response token as equally informative. Under a small preference budget, pair construction becomes nearly irrelevant: standard DPO-style, image-conditional, and self-improvement objectives all perform within 1.1 points on MM-RLHF metrics and 0.5 points on hallucination rates, matching the SFT baseline. We call this entropy-blind preference optimization: the learning signal ignores policy uncertainty, which in chain-of-thought VLMs concentrates on specific fork tokens where visual claims are made. We introduce GEM-MM, extending entropy-guided preference modeling to VLMs. It samples on-policy candidates, rewards high-entropy reasoning forks, penalizes uncertain final answers, and updates the model via group-normalized policy gradients. On a 3000-prompt MM-RLHF budget with Qwen3-VL-4B-Instruct, GEM-MM achieves the best entropy-depth preference score (61.7% vs. 57.3-58.4% baselines) and implicit-reward accuracy (41.4%). Crucially, it reduces AMBER hallucination from 23.4-24.0 to 20.6 and CHAIR from 5.0-5.1 to 4.6 at matched caption coverage. It also ranks best on all three HallusionBench granularities, with a conservative operating point raising the near-chosen rate to 62.4%. An off-the-shelf multimodal reward model ranks GEM-MM first, and two such models independently confirm our main ablations (p 0.001). We report confidence intervals, paired significance tests, metric-reward functional overlaps, and document a collapse mode when exceeding the fair budget. Code and evaluation harness: this https URL

[AI-126] Federated Dataset Simulation: Inducing Label-Free Heterogeneity Across Tasks NEURIPS2026

链接: https://arxiv.org/abs/2611.19975
作者: Vasilis Siomos,Lam Ngo,Jonathan Passerat-Palmbach,Giacomo Tarroni
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: NeurIPS 2026 Evaluations and Datasets Track. Code: this https URL

点击查看摘要

Abstract:Federated Learning enables collaborative model training across decentralized clients without sharing raw data. While the field has seen rapid growth, the datasets underpinning FL evaluation have not kept pace: real federated datasets remain scarce, and most studies rely on small-scale centralized classification datasets partitioned via Dirichlet label skew, a one-dimensional proxy for the multi-dimensional heterogeneity present in real federations. Moreover, label-based partitioning is by construction inapplicable to non-classification tasks such as segmentation and detection, for which no principled simulation method exists. We cast the conversion of centralized datasets into federated benchmarks as the task of Federated Dataset Simulation (FDS), and decompose it into a representation function \Phi and an assignment mechanism \pi . We provide two label-free instantiations – geometric partitioning via pretrained model embeddings and semantic partitioning via vision-language model-derived criteria – applicable to classification and dense prediction tasks alike. We evaluate across four datasets spanning classification, segmentation, and detection, and show that both methods induce heterogeneity that matches or exceeds that of label-skew partitioning, with measurable impact on downstream FL performance. We release fds, a modular open-source package for end-to-end federated dataset simulation, heterogeneity measurement, and FL training, at this https URL.

[AI-127] Value and Policy Calibration from Execution Feedback for Socially Compatible Driving

链接: https://arxiv.org/abs/2611.19969
作者: Sen Shen,Ning Liu,Xiangrui Kong,Yingwei Li,Dongkun Han
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Autonomous vehicle (AV) decisions at intersections affect the delay, braking, and motion smoothness of surrounding human-driven vehicles (HDVs). Learning socially compatible behavior requires connecting imagined interaction outcomes with feedback from executed actions. We propose a world-model reinforcement learning framework combines imagined policy improvement with direct value and policy calibration from finalized interactions, balancing AV progress with HDV interaction burden. An interaction-aware latent model supports imagination, while off-policy correction and replay updates enable observed rewards and transitions to supervise both the critic and actor. Simulation experiments at an unsignalized intersection achieve a 100% success and reductions in following-HDV acceleration variation of 65.5% and 44.0% relative to PIWM and DreamerV3, respectively. Ablations identify complementary roles for value calibration in reducing HDV burden and policy calibration in improving passage efficiency. Tests under out-of-distribution traffic shifts further show lower ego jerk and HDV acceleration variation than PIWM at comparable passage times.

[AI-128] Agent icJevNav: Agent ic Object Search and Navigation via Event-Driven Verification

链接: https://arxiv.org/abs/2611.19912
作者: Shuning Zhang,Shuli Zhang,Yunheng Wang,Wei Lu,Jiaxi Zhang,Yihang Kang,Tao Wang,Xiangyu Chen,Yixiao Feng,Kai Sheng,Renjing Xu
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注: 8 pages, 5 figures, 3 tables. Project page: this http URL

点击查看摘要

Abstract:Agentic navigation must maintain goals and adapt actions from incomplete evidence despite execution failures and delayed inference. We present AgenticJevNav, which links goal-directed planning, evidence memory, event-triggered verification, and bounded execution. The Agent plans from RGB, odometry, memory, and feedback, retaining goal ownership while authorizing a bounded action set. Jev reviews structured evidence after blocked/failed execution or two consecutive low-progress segments. Version, freshness, and once-only checks govern decision application. A single executor and independent RGB reflex constrain motion. Goal revisions revoke obsolete requests and actions before replanning. Across six simulated task families, AgenticJevNav achieves 73.33% SR over 60 trials, 21.67 percentage points above the next-best Agent-only aggregate control under differing historical protocols. A real-world GENISOM L1 quadruped deployment achieves 70.0% success over 30 trials, including bounded search reporting. Together, the simulation and physical-robot results demonstrate the effectiveness of the proposed architecture, while remaining failures highlight opportunities to improve grounding, approach planning, and runtime.

[AI-129] Recovery Gone Too Far: Understanding and Repairing Risky Fallback in LLM Agents

链接: https://arxiv.org/abs/2611.19886
作者: Liwen Wang,Jiangrong Wu,Zongjie Li,Zongyi Lyu,Shuai Wang,Shing-Chi Cheung
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
备注:

点击查看摘要

Abstract:Large Language Model (LLM) agents interact with their environments through tools, reading repositories, running builds, and pushing code. When a tool fails with an ordinary error, such as a rate limit or a timeout, the agent can fall back to a substitute tool and still complete the task. However, the same fallback that completes a task in one environment can cross an action boundary in another. For example, after a push is rejected, a forced push in one repository replaces only the agent’s own commits, while in a different repository it also removes another author’s commits. We call such a fallback a risky fallback. To understand this phenomenon, we conduct an empirical study on 304 scenarios that span eight domains and cover 55 real CVEs. All four mainstream models we evaluate exhibit risky fallback, with action-boundary violations in up to 96.8% of risky scenarios. We further evaluate three mainstream mitigations and find that none of them is effective against risky fallback. We then present nrepair. At each call to the substitute tool, nrepair scores the agent’s context with a monitor whose threshold is calibrated on normal runs alone; when the monitor triggers, it withholds the call once and returns the decision to the agent. nrepair reduces action-boundary violations on all four models, by up to 76.0% (from 67.5% to 16.2%), at a negligible runtime cost. On most models, the normal success rate drops by less than 7%. We also evaluate nrepair on two production agents, Claude Code and Codex, and find that it is equally effective on both.

[AI-130] Large Language Model-based Optimization for Wireless Communication: A Survey

链接: https://arxiv.org/abs/2611.19876
作者: Muhammad Sohaib,Taehyung Kim,Hyeri Heo,Jemu Ryu,Chaehyun Lee,Jun Zhang,Sang-Woon Jeon
类目: ystems and Control (eess.SY); Artificial Intelligence (cs.AI)
备注: 31 pages,9 figures

点击查看摘要

Abstract:The growing complexity and dynamism of modern wireless communication frameworks have made real-time optimization essential for efficient resource allocation. Traditional optimization and machine-learning techniques often struggle with generalization, adaptability, stability, and scalability in heterogeneous, time-varying environments. Large language models (LLMs) have recently emerged as powerful tools capable of reasoning over structured prompts and generating context-aware decisions in real time without domain-specific retraining. This survey provides a comprehensive review of LLM-driven optimization techniques applied to wireless communication problems such as power control, bandwidth allocation, task offloading, and intelligent surface configuration. We introduce a functional taxonomy of LLM-enabled wireless resource management strategies and analyze the prompt engineering methodologies employed in recent literature. Furthermore, we develop a generalized analytical framework that categorizes LLM-based optimization architectures, including direct solution generation, closed-loop optimization with external verifiers, and representation encoding within trainable neural pipelines. To demonstrate practical applicability, we present representative case studies on LLM-assisted beam prediction for millimeter-wave systems and adaptive wireless resource allocation. Finally, we discuss emerging hybrid AI approaches for wireless networks.

[AI-131] Let Execution Do the Teaching

链接: https://arxiv.org/abs/2611.19875
作者: Abhishek Mishra,Sandeep Srinivasa,Arko Chattopadhyay,Enrique Martinho Ferrao,Guda Pranav Reddy
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 11 pages

点击查看摘要

Abstract:Improving a coding model is a sequence of decisions: find the limiting behavior, create useful practice, and determine which attempts deserve reinforcement. We investigate a practical way to make those decisions increasingly repeatable by grounding feedback in execution. Starting from GLM-4.7-Flash, the work combines whole-file supervised fine-tuning, targeted C++ repair curricula, group-relative reinforcement learning, and verifiers that turn program behavior into structured rewards. Compiler errors and failing tests provide specific contexts for repair practice, while executable checks score the resulting programs. We evaluate the trained models on 26 C++ coding tasks, measuring both initial success and completion after test feedback. The strongest reported post-training configuration achieves 45.2% success in the first test round and 61.5% within two rounds. These results come from an assisted development suite with evolving prompt contracts; the strongest configuration uses four selected evaluation trials and includes six tasks also used in training. The study describes an execution-guided workflow for improving coding models, including how repair tasks are constructed, how program behavior determines rewards, and how the resulting models are evaluated.

[AI-132] Interactive Structured Environment Scaling for Large Language Model Agent ic Reasoning

链接: https://arxiv.org/abs/2611.19871
作者: Peng Yu,Luoyi Fu,Xinbing Wang,Ying Wen
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Reinforcement learning via environmental interaction is a cornerstone post-training paradigm for evolving large language models into autonomous agents capable of multi-turn tool use and exploration. Because the environment defines what the policy observes and how actions are verified, its structure directly bounds the capabilities an agent can acquire. Yet existing structured and retrieval environments rely on upstream pre-filtering—extracting a compact subgraph around the gold answer or retrieving fixed passages before the policy acts. We show that such pre-filtering introduces a fundamental measurement confound: because per-question subgraphs enclose ground-truth paths, they compress the candidate space by orders of magnitude and leak answer locations via local topology, allowing a question-blind degree heuristic to surface gold answers far above chance. To restore path selection, candidate pruning, and optimal stopping as genuine policy decisions, we propose RISE (Interactive Structured Environment), which transitions structured reinforcement learning from static subgraphs to full-scale interactive databases along two coupled axes. First, Decision-Space Scaling exposes the unpruned knowledge graph via a five-stage graph-agnostic typed intent adapter under a bounded call budget and a multiplicative grounding reward gate. Second, Learning-Frontier Scaling overcomes within-group variance collapse under group-normalized objectives by uniting model-independent structural tier progression with a self-evolving competence-frontier bandit. Across nineteen benchmarks and three model backbones, RISE substantially advances in-domain and unseen-schema knowledge-graph reasoning, transfers zero-shot to open-domain passage retrieval, and significantly enhances out-of-domain mathematical and logical deduction.

[AI-133] HalluProbe: Exploiting VLM Hallucinations for Backdoor Detection

链接: https://arxiv.org/abs/2611.19868
作者: Ji Guo,Xiaolong Qin,Jin Jiang,Yijing Liu,Pei Ke,Aiguo Chen,Wenbo Jiang
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Vision-Language Models (VLMs) are vulnerable to backdoor attacks introduced through data poisoning, making effective backdoor detection a critical security challenge. In this paper, we uncover a novel connection between VLM hallucinations and backdoor behaviors and exploit this connection for backdoor detection. Our key insight is that poisoned training data induces strong language priors toward the backdoor target token, while such priors are also a major source of hallucinations. This suggests that enhancing hallucinations can expose hidden backdoor-related priors. Specifically, backdoored VLMs tend to hallucinate backdoor target tokens more frequently across diverse inputs, while the hallucinated tokens of clean VLMs are largely random and exhibit no stable target-specific bias. Based on this observation, we propose HalluProbe, which enhances hallucinations during output decoding and measures the distributional consistency of hallucinated tokens to identify backdoored VLMs. We evaluate HalluProbe against six representative VLM backdoor attacks across three models. Experimental results demonstrate that HalluProbe can effectively detect backdoored VLMs with an accuracy exceeding 90%, outperforming existing backdoor detection methods.

[AI-134] Local Order Matters: A Lightweight CPU-Trainable CNN for Website Fingerprinting on Defended Tor Traffic

链接: https://arxiv.org/abs/2611.19864
作者: Puyang Xu,Zhizhao Wen,Yifan Jiang,Chao Wang,Yusi Gong
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 5 pages, 1 figure. Conference paper submitted to the 11th International Conference on Computer and Information Processing Technology (ISCIPT 2026)

点击查看摘要

Abstract:Website fingerprinting (WF) attacks let a passive adversary infer which site a Tor user visits from the encrypted cell sequence. On undefended traffic, hand-crafted-feature attacks are already near-ceiling, so deep learning offers only a marginal advantage there. We show that the decisive regime is defended traffic: against the FRONT padding defense, which injects dummy cells that corrupt the volume and count features feature attacks rely on, a compact 95,103-parameter 1-D convolutional network trained on the raw cell-direction sequence, on CPU only, stays robust while three implemented direction-only feature baselines degrade. On the closed-world WFlib FRONT set (95 sites), our model reaches 54.2% top-1 accuracy versus 38.5% for the strongest implemented feature baseline (k-FP-style RF), a gain of 15.7 percentage points (3 seeds; seed-0 paired-bootstrap 95% CI [+14.5, +17.0] pp on per-trace correctness, excluding zero). On undefended traffic the CNN reaches 94.1% versus 92.3% for k-FP-style RF. A controlled order-ablation permutes only nonzero traffic cells within each trace, preserving direction counts, retained length, and trailing zero-padding positions while destroying traffic order. This collapses the CNN below every feature baseline (from 54.2% to 7.6%, a 46.6 pp drop), showing that the model depends strongly on order-derived information under FRONT.

[AI-135] When Facts Change Which Decisions Break? Dependency-Aware Context Graphs for Agents

链接: https://arxiv.org/abs/2611.19859
作者: Paul Deepakraj Retinraj
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:When an observed fact changes, an agent must identify and repair the decisions that depend on it while preserving unaffected outputs. We investigate the reliability-cost trade-off of field-level dependency graphs extracted from actual agent read traces. A controlled benchmark spans three executable workflow domains and four successive updates, with independent imperative and expression-based references. Across 120 workflows, two models and two repeated runs, we compare full recomputation, rolling summary memory, lexical retrieval, trace-extracted dependencies and verified dependency upper bounds. Trace-derived graphs reduce reviewed decisions and estimated total cost, but omitted alternative supports and missing repair context can reduce complete-plan correctness. Paired ablations isolate the effects of edge deletion, spurious edges, transitive propagation and context scope. An exploratory enhancement supplies the existing record ID/field directory during extraction without revealing unread values or reference edges. We evaluate this enhancement on 60 fresh workflows with two repeats per model. On this fresh set, the directory-aware graph matches observed full-context complete-plan correctness for gpt-6.1-sol (100%) with an estimated 4.0% total-cost reduction, but reaches only 84.8% versus 99.6% for gpt-6-luna. The results distinguish potential dependency impact from actual decision invalidity and show why selective repair must be evaluated with both executable correctness and full memory-processing cost. The evidence concerns explicit procedural workflows; broad real-world agent-memory generalization remains untested.

[AI-136] Asking the right question is key to AI discovery in a chain of hypotheses

链接: https://arxiv.org/abs/2611.19841
作者: Hu Cang
类目: Artificial Intelligence (cs.AI); Optics (physics.optics)
备注:

点击查看摘要

Abstract:Artificial intelligence (AI) scientists pursue discovery through a chain of hypotheses, fitting candidate models to existing observations and often choosing experiments where those models disagree most. Yet when early data miss a narrow feature, model selection favors simple curves that agree across the unmea- sured gap, so disagreement-guided search has no reason to look where the new phenomenon lies. When we tested two frontier AI scientists, Claude Sonnet 5.5 and GPT-6.1-Sol, across 32 closed-loop trials with a narrow peak omitted from the starting data, both sampled inside the peak region less often than random search. Surprisingly, the agent that followed model disagreement more closely missed the peak in all 32 trials, whereas discovery succeeded when experiments probed the unmeasured gap early and followed up across the anomaly. Asking the right question, rather than testing only where current models disagree, is therefore key to AI discovery in a chain of hypotheses.

[AI-137] GraphPass: A Study of Graph Neural Networks for Compiler Pass Selection

链接: https://arxiv.org/abs/2611.19821
作者: Mohammed Saqlain
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Programming Languages (cs.PL)
备注:

点击查看摘要

Abstract:Compiler pass selection is usually governed by fixed pipelines (-O2, -O3) or hand-written heuristics. We study whether a graph neural network (GNN) over a program’s control-flow graph (CFG) can select among a fixed menu of five pass sequences in a single forward pass. GraphPass encodes each CFG with a graph attention (GAT) or graph convolutional (GCN) encoder and classifies it into one of the five sequences. On a synthetic benchmark of 5,000 generated CFGs in five program families, both encoders agree with the benchmark’s reference label on about 92% of held-out programs, versus 68.4% for a hand-written rule-based heuristic and 85% for a flat MLP. An edge ablation indicates that this gain comes from per-block processing rather than from CFG edges: a GCN given the same basic-block features with all edges removed scores 95.0%, versus 92.2% with the true CFG. In this benchmark the label is determined by the program family, so the task amounts to family recognition. Real-compiler speedup is not measured. Inference costs O(|V| + |E|) per CFG, measured at 1.7 ms for a typical program and about 0.8 s for a 100,000-block graph on one CPU core.

[AI-138] Reasoning in the loop: antibody lead optimization by probing mechanism and measurement

链接: https://arxiv.org/abs/2611.19812
作者: Samuel Stanton(1),Aaron Solomon(2),Emily Delaney(2),Melanie Montgomery(2),Crystal Humphries(2),Timothy Acker(2),Ryan Henrici(2),Charbel Eid(2),Eddie Abrams(2),Peyton Greenside(2) ((1) Anthropic, (2) BigHat Biosciences)
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Biomolecules (q-bio.BM)
备注: 31 pages, 19 figures, 8 tables

点击查看摘要

Abstract:Lab-in-the-loop systems, in which a model proposes molecules, automated assays measure their properties and the results retrain the model, are a popular approach to antibody lead optimization. These systems optimize for a specification defined with metrics computed from instrument signals, so their success depends on how well those metrics serve as proxies for the properties a drug candidate needs. When a metric is a poor proxy, neither a better model nor more data will fix it, and identifying the problem usually requires expert review. We report two campaigns run on BigHat Biosciences’ Milliner platform, one on an anti-albumin nanobody (VHH) and one on an anti-mesothelin single-chain variable fragment (scFv), in which a research team and a general-purpose AI agent, Claude, brought scientific reasoning into the loop. Over seven rounds and about 1,010 design variants we built libraries from antibody lineage, structure and biophysics, and found from the raw curves that on each scaffold one thermal metric did not measure what its name implied. The libraries were built as controlled comparisons, so the results also show what each mutation contributes to each property. We fit a physical model of the curves, used it to choose metrics suited to each scaffold, and changed the experimental protocol, varying protein concentration and heating rate, to test the model against pre-registered predictions. Both leads improve on their parents in affinity, thermal stability and aggregation while improving or maintaining yield and purity.

[AI-139] MAPS: A Multimodal Agent ic Framework for POI Data Production Based on Crowdsourced Street-view Imagery KDD2027 KDD

链接: https://arxiv.org/abs/2611.19807
作者: Sijie Ruan,Tingyu Xie,Mingfei Liu,Yicheng Tang,Haoyu Han,Jingnan Qi,Jingshu Zeng,Dongqi Xu,Jing Wang,Shuliang Wang
类目: Artificial Intelligence (cs.AI); Databases (cs.DB)
备注: Accepted for presentation at the 33rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2027), Cycle 1

点击查看摘要

Abstract:Point-of-interest (POI) databases support location-based services, yet urban changes quickly make their records stale and manual maintenance is costly. Crowdsourced street-view images provide timely visual evidence, but converting them into database actions requires separating multiple POIs and irrelevant text in each image, then resolving incomplete observations against ambiguous records. We study Street-view Image-based POI Data Production and propose MAPS, the first framework to integrate structured POI extraction with database maintenance from such imagery. MAPS contains two cascaded modules. SPICE progressively adapts a lightweight vision-language model to reconstruct and filter text, group entries belonging to the same POI, and classify their semantics and completeness. RIMA retrieves candidate records, performs evidence-grounded POI matching and attribute-level reasoning, and selectively acquires complementary visual, geospatial, or external-source evidence when necessary. In human-reviewed full-task evaluation, MAPS outperforms the strongest baseline by 21.8%. MAPS has also been deployed in Kuaishou’s large-scale POI production pipeline, reducing average processing time by 76.9% relative to manual annotation. These results demonstrate the effectiveness and practical value of MAPS for scalable POI maintenance.

[AI-140] Lossless First-Action Menus for Linear-Reward Planning

链接: https://arxiv.org/abs/2611.19795
作者: Yukun Yang
类目: Artificial Intelligence (cs.AI)
备注: 8 pages, 0 figures

点击查看摘要

Abstract:We study sequential decision making where an agent reserves a minimum menu of initial actions before multi-objective trade-offs are revealed, followed by unrestricted continuation planning once linear reward weights arrive. We prove that a menu preserves the unrestricted optimal return across all linear scalarizations if and only if its reachable continuations cover all vertices of the full feature polytope. This geometric equivalence yields an output-sensitive algorithm that is fixed-parameter tractable in the vertex count, combining scalar backward induction with oracle-based facet identification. On explicitly layered graphs with unit feature increments, menu verification is coNP-complete and synthesis is complete for the second existential level of the polynomial hierarchy, even under additive loss strictly below one. In fixed dimensions two and higher, verification is polynomial whereas synthesis remains NP-complete, while one-dimensional instances admit polynomial-time solutions.

[AI-141] Read Before You Fit: Text-Derived Constraints for Discovering Differential Equations from Noisy Biological Data

链接: https://arxiv.org/abs/2611.19779
作者: Pronoma Banerjee,Md. Masudur Rahman,Juan P. Wachs
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models are increasingly used for scientific equation discovery. However, most existing methods are strongly data-driven and assume sufficiently clean and representative observations. This assumption is problematic in biological systems, where individual-level measurements are often sparse, noisy, heterogeneous, and expensive to obtain. While the underlying mechanisms are extensively described in scientific text, they are not explicitly used for the LLM based modeling of such systems. We introduce P2C, an agentic framework that translates textual descriptions of biological systems into a typed set of formal constraints, and use these constraints to restrict candidate ordinary differential equation (ODE) structures before numerical fitting. P2C consists of five stages: Text2Constraint extraction, ODE discovery, symbolic verification, parameter fitting, and model scoring. We evaluate P2C on four published ODE systems of acute inflammation and immune regulation, under clean and noisy observations. By progressively degrading the numerical evidence, we isolate when textual domain knowledge contributes beyond regression alone. Across the four systems, adding Text2Constraint reduces the worst form error from 1.397 to 0.499 on clean data and from 13.096 to 1.095 for data perturbed with Gaussian noise of \sigma=0.1 ; P2C also achieves lower form error and trajectory error than several LLM-based baselines over the tested noise range.

[AI-142] Exploration of Reinforcement Learning in Generative Recommendation Models

链接: https://arxiv.org/abs/2611.19742
作者: Pengfei Zheng,Wuchao Li,Shiyao Wang,Honghui Bao,Hongyi Xie,Jiaxin Deng,Xingmei Wang,Jinghao Zhang,Yufei Ye,Ruiming Tang,Guorui Zhou,Han Li,Kun Gai,Defu Lian
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Traditional cascaded recommender systems balance accuracy and efficiency over large item spaces, but misaligned stage-wise objectives hinder end-to-end optimization. Generative recommendation unifies these stages in an autoregressive framework, providing a natural interface for reinforcement learning (RL) to optimize the entire generation process. However, directly applying RL often leads to unstable optimization and degraded recommendation quality. We revisit Group Relative Policy Optimization (GRPO) and other ratio-based objectives through a unified logit-gradient view. We identify three issues: (i) clipped surrogates create dead-gradient regions, under-utilizing samples; (ii) negative-advantage updates can become effectively unbounded as importance ratios grow, allowing a small subset of negatives to dominate optimization; and (iii) ratio-based control lacks sensitivity to absolute policy confidence, so updates on high-confidence cases may fail to vanish. Motivated by this diagnosis, we propose \emphGradient-Bounded Policy Optimization (GBPO), a simple objective with stop-gradient normalization that is simultaneously (a) fully gradient-active, (b) sign-symmetrically bounded for both positive and negative updates, and © explicitly confidence-weighted under current policy probabilities. This ensures that updates gracefully vanish for \pi!\to!1 positives and \pi!\to!0 negatives. Across the evaluated settings, GBPO achieves competitive recommendation performance with more stable training than competing RL methods. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.19742 [cs.LG] (or arXiv:2611.19742v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19742 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-143] An Interpretable Acoustic Coordinate System for Japanese Consonant Morae in Dysarthric Speech

链接: https://arxiv.org/abs/2611.19714
作者: Yuma Enomoto,Misa Yamaguchi,Hiroki Nishikawa,Ittetsu Taniguchi,Kazuhide Matsunaga,Narikazu Uzawa,Takao Onoye
类目: ound (cs.SD); Artificial Intelligence (cs.AI)
备注: 8 pages, 9 figures, IEEE SLT 2026

点击查看摘要

Abstract:Dysarthria reduces speech intelligibility, requiring self-directed rehabilitation based on fine-grained feedback to capture gradual articulatory changes. For Japanese speakers, such feedback is naturally provided in mora units due to its distinct moraic timing and consonant-vowel structure. While recent visualization methods have explored distance-based feedback, most are limited to vowels. Extending them to consonants is challenging, as generative embeddings often smooth transient spectral structures. We propose a visualization system for Japanese consonant morae that maps articulatory deviations onto a two-dimensional space defined by temporal and spectral distances from healthy reference speech. Experiments show consistent geometric separation of consonant categories, with Silhouette Coefficients exceeding 0.5 and origin-based ROC-AUC scores above 0.86. Furthermore, evaluations using real dysarthric speech achieved a high ROC-AUC of 0.897 in screening for unclear pronunciations. Perturbation tests confirm orthogonal temporal and spectral shifts, yielding a physically interpretable mapping for rehabilitation.

[AI-144] Architecting the Corporate AI Infrastructure: The Evolving Roles of AI Governance Committees and Chief Ethics Officers

链接: https://arxiv.org/abs/2611.19670
作者: David Hason Rudd,Dennis Singh,Cesar Sanin,Bryn James
类目: Computers and Society (cs.CY); Artificial Intelligence (cs.AI)
备注: 8 pages, 1 figure, 2 tables. Published in Procedia Computer Science 292 (KES 2026, 30th International Conference on Knowledge-Based and Intelligent Information Engineering Systems)

点击查看摘要

Abstract:The accelerating integration of artificial intelligence (AI) into core business operations has created an urgent need for robust corporate governance infrastructures capable of managing complex ethical, legal, and operational risks. While organizations have increasingly adopted high-level AI ethics principles, a critical implementation gap persists: fewer than half of organizations that cite AI governance as a boardroom priority have a formalized governance framework in place. This paper addresses that gap by proposing a structured corporate AI governance architecture (CAIGA) centred on two mutually reinforcing pillars. First, it examines the evolving mandate of specialized C-suite roles, specifically Chief Privacy Officers (CPOs) and Chief Ethics Officers (ChEOs), in formalizing accountability for AI usage across the enterprise. Second, it argues that cross-functional AI Governance Committees, uniting representatives from legal, technology, data science, cybersecurity, and business units, are the essential organizational mechanism for translating ethical principles into operational decisions, including strategic go/no-go approvals for new AI initiatives. Drawing on a conceptual review of peer-reviewed governance literature, regulatory frameworks including the EU AI Act, and documented industry practice, this paper contributes a replicable governance architecture that embeds clear accountability policies, structured risk management workflows, and an organization-wide ethical AI culture. The framework is illustrated through application to the financial services sector. The findings have direct implications for corporate leaders, risk professionals, and policymakers navigating the rapidly evolving global AI regulatory landscape.

[AI-145] When AI Agents Should Repair Replan or Abandon: A Threshold Theory of Plan Commitment

链接: https://arxiv.org/abs/2611.19663
作者: Gaurav Tewari
类目: Artificial Intelligence (cs.AI)
备注: 22 pages, 6 figures, 3 tables

点击查看摘要

Abstract:AI agents often discover mid-execution that their plan is broken, for example when an action fails or a tool is withdrawn. Deployed systems use conventions such as retry loops or “always replan,” and a recent benchmark finds no fixed strategy dominates. This paper derives the optimal policy. In the model, a disruption destroys the causal support of part of a plan; repair is a partial reset that preserves the plan’s structure, replanning is a full reset, and changing a plan is costly because others have conditioned on it. The optimal policy is a threshold in effective disruption depth: repair shallow damage, replan after deep damage, and abandon the goal when even a fresh plan is not worth its cost. Depth depends on plan structure. In expectation, a single broken action forces 1+R(G)/n actions to be revalidated, where R(G) counts the pairs of actions joined by a dependency path; to first order, a fully serial plan switches to replanning at raw damage about w times smaller than a plan of w parallel chains. A closed-form stability-price threshold determines when a worse repaired plan should be kept over a better fresh one. As language-model planners push replanning costs toward zero, the stability price becomes the only force sustaining plan commitment. Under repeated disruptions, revision frictions make commitment hysteresis optimal: the agent keeps executing a plan it would never choose fresh, and each repair adds patch debt until the debt outweighs repair’s cost advantage and the agent rebuilds. A value-of-diagnosis result states when measuring damage is worth its compute and shows that “always replan” is optimal when damage is predictably deep and replanning is the better fallback. A separately implemented numerical checker tests each proposition (zero violations; 1.6 million static-model evaluations), and the model’s regime map reproduces that benchmark’s regime flips qualitatively.

[AI-146] BadGCL: Towards Persistent Backdoor Attack on Graph Continual Learning Systems KDD2027

链接: https://arxiv.org/abs/2611.19656
作者: Dong Li,Yufei Han,Xintao Wu,Xujiang Zhao,Wei Cheng,Chen Zhao
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted in KDD 2027

点击查看摘要

Abstract:Backdoor attacks pose a persistent security threat to machine learning systems, yet their behavior under graph continual learning (GCL) remains underexplored. Conventional graph backdoor attacks often lose effectiveness as models are updated across successive tasks, creating the impression that continual updates naturally weaken malicious associations. This work reveals a contrasting vulnerability: knowledge-preservation mechanisms in GCL can also preserve early injected backdoors when the poisoned update is compatible with the clean-task signal being retained. For graph data, this risk is amplified by structural triggers, which are embedded into the message-passing paths of graph neural networks (GNNs) and can induce stable representation shifts across tasks. To expose this risk, we propose BadGCL, a model-agnostic framework for persistent backdoor attacks under a strict early-stage intervention setting. BadGCL learns topology-aware structural triggers using a gradient mutual-information objective that couples poisoned gradients with clean-task optimization signals, together with a propagation consistency objective that stabilizes trigger-induced effects through graph message passing. We provide theoretical analysis showing that gradient compatibility and propagation stability jointly support long-term backdoor persistence under benign continual updates. Extensive experiments across task-IL and class-IL protocols, four GCL benchmarks, and diverse victim systems show that BadGCL achieves persistent attack effectiveness with limited clean-utility degradation, substantially outperforming graph and continual-learning backdoor baselines and remaining robust against representative defenses.

[AI-147] LoopFill: Pipelined Multi-Token Speculative Decoding for Looped Transformers

链接: https://arxiv.org/abs/2611.19649
作者: Siqi Zhu,Suozhi Huang,Yi Zhao,Jiaxuan You
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Looped Transformers improve parameter efficiency by repeatedly applying the same Transformer layers. During autoregressive decoding, however, generating each token requires several sequential passes through these layers, leading to high decoding latency. Pipelined self-speculative methods use shallow-loop predictions to process tokens at different loops in the same batch, but each token must reach the drafting depth before the next can enter the pipeline. We introduce LoopFill, a training-free speculative decoding algorithm for looped LLM that combines multi-token suffix drafts with shallow-loop predictions in a single draft tree. Both drafters contribute to one tree, so a suffix lookup can immediately extend a shallow prediction and several future tokens can enter the pipeline together. The shared block computes the highest-scoring draft candidates at different loop depths in one batch, and full-depth verification preserves the model’s autoregressive output distribution. Across seven benchmarks, LoopFill’s average speedup over greedy autoregressive decoding is 5.85\times on Ouro-2.6B and 5.13\times on Huginn-3.5B. The speedups exceed those of LoopAhead, the strongest baseline, by (40%) on Ouro-2.6B and (25%) on Huginn-3.5B.

[AI-148] Robust Well-Log Reconstruction Across Point Gaps and Whole Wells via Boundary-Gated Soft-Thresholding Kalman Fusion

链接: https://arxiv.org/abs/2611.19625
作者: Yongan Zhang,Zhihan Qiu,Yizheng Li,Youzhuang Sun
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 17 pages

点击查看摘要

Abstract:Missing wireline logs hinder subsurface characterization. Deep reconstruction networks suffer from gap-boundary steps and cross-basin shifts, while interpolation and Gaussian smoothers struggle with long intervals and outliers. We propose boundary-gated soft-thresholding Kalman fusion (ST-KF), unifying multiresolution wavelet shrinkage, distance-gated residual prior learning, and closed-form robust Kalman smoothing with learning–inference decoupling. Across two public benchmarks totaling 263 wells in the Norwegian North Sea and New Zealand’s Taranaki Basin, ST-KF ranks first over 0.15–100-m anchored gaps with 0.09% parameter overhead. Against the strongest of nine baselines, pooled RMSE falls by 5.0%–8.3% for bulk density and 6.4%–15.6% for compressional slowness; against the four strongest baselines, reductions are 0.7%–4.1% for neutron porosity and 1.3%–2.9% for photoelectric factor. Ablations and sensitivity analyses confirm operator synergy and stable performance: removing state-space fusion increases short-gap errors by up to 58.8%, while zero-shot transfer cuts anchored errors by 14%–68%. The results show that local borehole continuity transfers across basins whereas cross-property mappings shift with regional geology, supporting measurement-guided log reconstruction.

[AI-149] Well-Log Reconstruction with Pre-trained Geological Knowledge: A Knowledge-Conditioned Large-Language-Model Generative Agent

链接: https://arxiv.org/abs/2611.19624
作者: Yongan Zhang,Yitao Yu,Yizheng Li,An Gong,Youzhuang Sun
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 35 pages

点击查看摘要

Abstract:Missing well-log intervals make subsurface reconstruction ill-posed. Existing numerical regression models lack geological prior knowledge to constrain ambiguous inverse mappings. We propose GeoLog-Agent, a knowledge-conditioned generative agent using pre-trained large language models (LLMs) as a source of geological priors. A universal Geo-Harness translates LLM stratigraphic reasoning into formation-anchored petrophysical trajectories, feature-wise linear modulation, and conditional adversarial constraints. On the Norwegian North Sea benchmark, GeoLog-Agent achieves a mean R^2 of 0.9307 and a peak R^2 of 0.9713. It exceeds the nine-baseline average R^2 of 0.8387 by 0.0920, with a 30.97% average RMSE reduction, and exceeds the strongest baseline by 0.0274–0.0498 in R^2 , with 16.25%–24.05% lower RMSE. Ablations confirm the effectiveness and synergy of geological prior reasoning, multi-scale dilated generation, and conditional adversarial modules. Transferring the coupling mechanism to six conventional deep learning models improves their R^2 by 0.0162–0.0789 with cross-well stability. These results demonstrate how foundation-model domain reasoning can constrain deep generative reconstruction of missing petrophysical measurements.

[AI-150] Data-Driven Heuristic-Guided Business Process Redesign

链接: https://arxiv.org/abs/2611.19623
作者: Nan Sai,Karolin Winter,Remco Dijkman
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Business process redesign (BPR) has a long-standing catalog of best-practice redesign heuristics, yet, applying them remains a mostly manual task that requires extensive domain knowledge. Despite recent advances using Large Language Models (LLMs) for conversational process model redesign, limitations remain. For example, redesign is user-driven instead of data-driven and is not guided by best practices such as the redesign heuristics. Moreover, only one change can be handled at a time. To address these limitations, given an event log and a BPMN model, we present the DH-BPR framework to realize data-driven heuristic-guided BPR using LLMs. To design DH-BPR, we first conducted an analysis of which redesign heuristics can be automatically detected and applied given information contained in an event log such as the fraction of cases that bypass an activity. Based on this, DH-BPR detects and validates heuristic candidates through a combination of rule-based techniques and an LLM. After that, DH-BPR jointly designs and records an application plan for valid candidates, before finally yielding a redesigned BPMN model and an updated resource model (the way resources execute the process). Evaluation based on simulation using a synthetic benchmark and two real-world event logs shows that the framework’s redesign reduces the mean case completion time by up to 61.5%.

[AI-151] A Hierarchical Statistical Framework for Evaluating Agent Reliability

链接: https://arxiv.org/abs/2611.19614
作者: Mirco Milletarì,Tuhin Kundu
类目: Artificial Intelligence (cs.AI); Statistical Mechanics (cond-mat.stat-mech); Probability (math.PR)
备注:

点击查看摘要

Abstract:Agents operating in complex scenarios often face uncertainty due to incomplete information about the environment they operate in, ambiguous user queries, complex business rules and the inherently stochastic dynamics of the governing policy. Several new benchmarks have been proposed to evaluate LLM-based agents on complex tasks in multi-tool environments, including settings with policy constraints and varying levels of intent ambiguity. These conditions make it important to evaluate agentic systems systematically and quantify their performance under different conditions to assess their reliability in production. This need is evident in the recently introduced Thinkingbox benchmark, where even the strongest models struggle to perform consistently across diverse business scenarios. This paper presents a statistical framework for evaluating uncertainty and reliability in agentic systems. We define a hierarchy of scenarios, instances, and samples to assess model performance and variability systematically. We apply this framework to common metrics found in the literature: mean pass rate, pass@k , and passhat^k , with uncertainty estimates for each metric. Aggregating metrics at different levels of the hierarchy reveals different structures and properties of the policy model; we connect this hierarchical aggregation with the known process of \emphcoarse-graining as intended in the physics and statistical literature. We independently derive exact finite-sample expressions for the sampling variances of the unbiased pass@k and passhatk estimators; related results also appear in recent submissions. Our central contribution is a constructive derivation of hierarchical uncertainty propagation, making explicit its correspondence with coarse-graining and separating execution noise from task heterogeneity.

[AI-152] Synthetic Neurosymbolic Task Generation with Controlled Constraint Complexity

链接: https://arxiv.org/abs/2611.19597
作者: Mykhailo Buleshnyi,Zsolt Zombori,Michael Benedikt
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Logic in Computer Science (cs.LO); Neural and Evolutionary Computing (cs.NE); Symbolic Computation (cs.SC)
备注:

点击查看摘要

Abstract:We introduce NeSySy, a synthetic generator for neurosymbolic learning tasks that jointly produces logical constraints and consistent datasets. Unlike existing benchmarks, NeSySy provides fine-grained control over key properties such as constraint structure, tightness, input dependence, and data generation and alignment strategies. In particular, it allows the use of rich constraints over infinite domains, and uses solvers to ensure that the constraints are meaningful and that generated samples satisfy the constraints. We design NeSySy to support both learning with declarative constraints and learning from distant supervision. Across a range of constraint families, we provide evidence that its parameters induce systematic and largely independent effects on task difficulty. Using NeSySy, we study constraint-aware learning methods, identifying regimes where explicit constraint enforcement is beneficial and where expressive unconstrained models can instead learn constraint structure directly from data. Our results establish NeSySy as a controlled testbed for analysing the interaction between learning and logical structure. NeSySy is built around a pipeline that incorporates solvers. Satisfiability checkers (CNFgen and Z3) generate and certify constraints. Constrained optimizers (RC2 for MaxSAT, Z3 for OMT) compute optimal repairs that guarantee every generated sample is consistent. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Logic in Computer Science (cs.LO); Neural and Evolutionary Computing (cs.NE); Symbolic Computation (cs.SC) Cite as: arXiv:2611.19597 [cs.LG] (or arXiv:2611.19597v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19597 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-153] PBit1: One Programmable P-Bit per Parameter - A Finite-Control Representation of Pretrained Networks

链接: https://arxiv.org/abs/2611.19588
作者: Omar Ramadan
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 7-page main paper with 3 figures, plus a 12-page supplement

点击查看摘要

Abstract:We introduce PBit1, a finite-control representation of pretrained neural networks that assigns one programmable p-bit to each learned parameter. A 12-bit command sets its probability, a shared affine range maps that probability to the weight domain, and temporal observations reconstruct the weight for inference. Operator-aware range grouping and sensitivity-based sampling allocation adapt the representation to the network. Across five independent weight draws at 32,768 observations per scalar, selected configurations preserve task quality in GPT-2, DistilBERT, and LFM3-1.7B with every learned scalar converted and no fine-tuning. GPT-2 and LFM3 perplexity increase by approximately 0.66% and 0.26%, respectively, while DistilBERT accuracy changes from 87.51% to 87.49%. A sensitivity-based policy reduces GPT-2 reconstruction observations by 41.7% while retaining the quality criterion. Complete-model evaluations reuse reconstructed weights in BF16 inference, and THRML operator experiments examine native sampling and spatial-temporal tradeoffs. Together, these results establish PBit1 as a concrete construction for studying finite-control weight representation and temporal reconstruction in pretrained networks.

[AI-154] Risk-Averse Decision-Focused Learning: Minimizing CVaR of Regret

链接: https://arxiv.org/abs/2611.19575
作者: Jayanta Mandi,Senne Berden,Victor Bucarey,Concepcion Dominguez,Tias Guns
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Many combinatorial optimization (CO) problems involve parameters that are uncertain when the problem must be solved. In such settings, the parameters have to be predicted from correlated contextual information before a solution can be computed. \emphDecision-focused learning (DFL) addresses this by training the predictive model to produce predictions that optimize downstream decision quality. Most existing DFL methods are risk-neutral and minimize the expected regret. In high-stakes applications this can lead to undesirable outcomes, because decision-makers must also account for the tail risk of decision quality. Prior risk-sensitive approaches typically model uncertainty explicitly and solve a stochastic or robust optimization problem at inference time, which increases computational complexity. We instead propose a DFL framework that achieves risk-sensitive behavior while retaining point predictions and deterministic optimization at inference time. Starting from the dual representation of the Conditional Value-at-Risk (CVaR) and adding entropy regularization, we derive a smooth surrogate for the CVaR of regret. Its gradient induces an instance-dependent reweighting within each minibatch that emphasizes high-risk outcomes, so the approach can turn existing surrogate-loss DFL methods, including the SPO+ loss, into risk-sensitive variants. When instantiated with SPO+, the resulting objective admits a generalization bound on the CVaR of the true regret. Experiments on knapsack, portfolio, shortest-path, and energy-aware scheduling problems show substantial reductions in tail regret when the uncertain parameters exhibit heterogeneous tail behavior.

[AI-155] Learning from Mutants: Teaching Small Language Models to Synthesise Loop Invariants

链接: https://arxiv.org/abs/2611.19569
作者: Daragh King,Vasileios Koutavas
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Large language models (LLMs) have proven promising for synthesising loop invariants, a long-standing obstacle to automated program verification. Yet, frontier LLMs are typically accessible only through third-party APIs, hampering reproducibility and local deployment. Small language models (SLMs) sidestep these obstacles but require task-specific fine-tuning, for which high-quality training data is needed. We present a failure-driven data curation method for fine-tuning language models to synthesise loop invariants. Inspired by mutation testing, our approach applies semantic mutation operators to ground-truth invariants, producing semantic mutants that induce verification failure. Each mutant is paired with verifier feedback and its ground-truth counterpart, yielding training examples for feedback-guided invariant refinement. Whereas the vast space of mutants is a long-standing burden in mutation testing, it becomes an asset when generating training data, multiplying the examples obtainable from a fixed set of verified programs. We implement our approach for Dafny, augmenting an existing synthetic dataset of verified programs, and introduce a framework that iteratively refines candidate invariants using verifier feedback. On 472 DafnyBench programs, our best-performing fine-tuned Gemma-4-E4B configuration verifies 50.8% of programs zero-shot, compared to 8.1% for the base model, a 6.3 \times improvement. Higher-order mutants prove more effective as training data than first-order mutants, and adding mutation-based examples to direct synthesis examples, in which invariants are removed rather than mutated, improves performance under iterative refinement. Our approach also transfers to Qwen3.5-4B, whose zero-shot verification rate rises from 9.3% to 53.6%.

[AI-156] Inquest: Log-Guided Root-Cause Code Localization with a Staged Agent ICSE2027

链接: https://arxiv.org/abs/2611.19554
作者: Yuvraj Sehgal,Sneh Patel,Mahsa Panahandeh,Naser Ezzati-Jivan
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 12 pages, 2 figures, 6 tables. Submitted to the ICSE 2027 Industry Challenge track. Benchmark: this https URL ; code, results and run transcripts: this https URL

点击查看摘要

Abstract:When a production issue is reported, engineers get a description and some logs, but no failing test and no way to reproduce the fault. The logs show where the failure was noticed, not where it was caused: in 49% of the real reports we study, the text names none of the files that the fix later changed. We built a benchmark for this setting from two public datasets and 27 open-source projects (Apache, Red Hat, MongoDB, MariaDB, Qt and Sakai). It holds 38,001 resolved issue reports with their pasted logs, the code as it was while the bug was still present, and the files the fix changed. We present INQUEST, an agent that finds the faulty code using a 27B open-weights model. Code, not the model, controls the procedure. It first gathers the evidence, including the files fixed in similar past issues that were resolved before the report was filed. The model then makes only three bounded decisions: which files to open, what role each file plays in the failure, and how to rank them. On 26,544 evaluation reports, INQUEST ranks a changed file first for 68.6% and in the top five for 85.9%, names the correct module for 90.2%, and ranks a changed function first for 53.8%. On the reports that name no faulty file, it reaches 52.8% and 73.6%. A free agent loop with the same model and tools finds files equally well, but names the faulty function 14 points less often. On a 1,000-report sample, INQUEST beats our reimplementations of two published agents by 7 and 20 points of top-1 accuracy (4 and 16 counting only the reports those agents answered), and the same model used as a re-ranker by 6. We report the cost of every system and the results that go against us, and we release the benchmark, both systems and every run transcript.

[AI-157] Beyond Prompts: Runtime Contracts for Coding Agents

链接: https://arxiv.org/abs/2611.19549
作者: Reda Heddad
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 9 pages, 5 figures, 8 tables. Code, data and run logs: this https URL

点击查看摘要

Abstract:Coding agents act on repositories through shell commands and file edits, and teams steer them with written engineering rules. Industrial settings expect such rules to be enforced and their application to be evidenced, and a rule given as prompt text provides neither. We test whether executable contracts enforced while the agent works reduce rule violations compared with the same rules in the prompt, retrieved as guidance, or checked only after generation. Violations are scored by an oracle whose methods differ from the in-loop checks, and the oracle is validated by two human reviewers and a model reviewer. Across three studies on two model families (240, 112 and 108 runs), enforcement during execution produced no oracle-detected violation in 104 paired runs, against 21 to 55% of runs without rules, and no loss of functional correctness in the two confirmatory studies. In the confirmatory study on tasks aimed at known failure situations (Study 3, protocol hashed before the first run), enforcement lowered the share of violating runs by 31 percentage points relative to no rules (Holm-adjusted p = 0.014); the 17-point difference from prompted rules was not significant (p = 0.12), and neither was any contrast in Study 2, where only 3 of 14 tasks produced a violation. Prompted rules removed some violation types but not recursive deletes of scratch directories, and checks after generation repaired file edits but not commands that had already run. That enforcement satisfies the rules it enforces is expected by construction; what the data add is where prompting falls short, the absence of a functional cost, and the limits of the rules and of the oracle themselves. The studies use low-cost models and one small fixture.

[AI-158] Evidence-Conditioned Completion and Synthesis of Electronic Health Records

链接: https://arxiv.org/abs/2611.19546
作者: Sicheng Shu,Parker Knight,Huitao Li,Yuan Luo,Yifan Peng,Shawn N. Murphy,Rui Yang,Chuan Hong
类目: Artificial Intelligence (cs.AI); Applications (stat.AP)
备注:

点击查看摘要

Abstract:Patient information can guide electronic health record (EHR) generation, but how its amount shapes predictive utility and disclosure risk remains poorly characterized. Here we present EHRDiB, which unifies visit completion and synthesis by distinguishing original missingness from withheld inputs and combining a shared multimodal representation with an evidence-conditioned latent diffusion bridge. We compared EHRDiB with three completion and eight synthesis methods on MIMIC-IV and numerical and discretized Duke Health representations. Evaluation covered downstream predictive utility, generative diversity, and distributional fidelity for both tasks, and disclosure risk for synthesis. EHRDiB achieved the smallest utility gaps from real-data-trained classifiers in seven of eight completion comparisons. In synthesis, AUPRC gaps were 45.2% smaller for ten MIMIC-IV principal-diagnosis targets and 53.3% smaller for numerical Duke Health data relative to the best comparator. Synthesis with calibrated thresholding yielded the lowest marginal distribution error across all three datasets, with Coverage above 0.95 for both output rules. Across twelve evidence-fraction intervals in MIMIC-IV synthesis, membership-inference AUROC increased from 0.516 to 0.598 as the mean target evidence fraction rose from 0.20 to 0.50; utility gaps narrowed relative to the lowest evidence configuration but did not decrease monotonically. Our study provides a unified framework and empirical foundation for EHR generation that considers both its utility for clinical analysis and its information disclosure risk.

[AI-159] Func-JEPA: Learning a Sobolev Space of Functions as Latent Vectors

链接: https://arxiv.org/abs/2611.19542
作者: Alejandro Moreno Arcas
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Numerical Analysis (math.NA)
备注: 42 pages, 16 figures, 30 tables. Code: this https URL . Model and data: this https URL

点击查看摘要

Abstract:Pretraining an encoder once and reusing it across tasks is a widespread approach in vision and language. However, neural symbolic mathematics still trains a dedicated sequence-to-sequence model for each task, so that no representation of a function is shared. In this work, we propose the Function Joint-Embedding Predictive Architecture (Func-JEPA, abbreviated f-JEPA): an encoder of mathematical functions, pretrained on their symbolic form, whose latent space imitates the Sobolev space H^1 and in which differentiation is a learned latent operator. The norm of a latent is computed exactly from the function, whereas its direction is learned to match Sobolev angles. The frozen encoder is then reused by lightweight modules for decoding, composition, integration, and regression from samples. The representation and four downstream tasks are evaluated within and outside the pretraining distribution. f-JEPA recovers exact formulas from noisy samples more often than the compared symbolic-regression methods, whereas computer algebra and task-specific models remain stronger on standard integration and differential-equation benchmarks.

[AI-160] Continuation-Guided LLM Evolution of Program Policies and Goal Compilers in Craftax

链接: https://arxiv.org/abs/2611.19537
作者: Víctor Gallego
类目: Artificial Intelligence (cs.AI)
备注: Preprint

点击查看摘要

Abstract:Coding agents can synthesize and optimize code-based policies from simulator feedback, allowing the resulting program to act with zero inference model calls. But for long horizon tasks, two challenges arise: local improvements can hurt final success and switching to a new objective requires another expensive synthesis run. We propose a framework that first develops competence by refining the policy through a feedback cascade, moving from episode returns to hierarchical goals and belief updates which can be scored densely from simulator. In the Craftax game, a partially observed and procedurally generated environment, refinements via episode level feedback stall at 1% final-level arrival, whereas with the feedback cascade the optimized policy reaches the final stage 50% of time (28% final boss defeats). Furthermore, to make this competence reusable towards new goals, we factor the previous controller into an environment interpreter I (models, memory, skills) and a rule program p_0 , and we show that coding agents can optimize a reusable goal compiler C_\theta that maps structured tasks g into new rule programs p_g . Across five independent compiler searches trained only on short tasks, the resulting compilers successfully transfer zero-shot to longer tasks without any task-specific tuning nor inference-time LLM calls: continuation-scored feedback and a fixed machine with an editable program together turn a single-task competent policy into a reusable one. Code at: this https URL Comments: Preprint Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2611.19537 [cs.AI] (or arXiv:2611.19537v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.19537 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-161] Agent Authenticator: Passkey-Grade Hardware-Bound Authentication for AI Agents

链接: https://arxiv.org/abs/2611.19536
作者: Harish Santhanalakshmi Ganesan
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI coding agents now push code, publish packages, call model APIs and operate cloud resources on behalf of developers. They authenticate to these services with the same bearer secrets humans use, personal access tokens, API keys and package-registry tokens, stored where the model, its subprocesses and any co-resident malware can read them. Incidents in 2025 and 2026, including prompt-injection attacks on coding agents and self-propagating supply-chain worms, show that this arrangement fails in predictable ways. Passkeys solved the analogous problem for humans, but they require a human gesture on every use and cannot be operated by an autonomous process. We present the Agent Authenticator, a design that gives agents passkey-grade authentication properties: per-service hardware-bound keys, no shared secrets, origin binding enforced by the runtime rather than the model, normalized attestation of the execution environment, human-signed scoped mandates, and cross-device human approval for high-risk actions. The same protocol serves agents on developer laptops and in cloud enclaves, and it deploys as a profile over existing mechanisms, MCP authorization, OAuth token exchange including the Identity Assertion JWT Authorization Grant, git and SSH, rather than as a new identity system. The authenticator also produces a signed, hash-chained record of each action it authorizes. We analyze the design against a threat model centered on prompt injection and same-host compromise, map it onto five documented incidents, and discuss its limits.

[AI-162] Ultra-CAN: A Lightweight Blind ISAC Receiver for Chaos-Based Telemetry and Local-Seed Ranging in the Internet of Drones

链接: https://arxiv.org/abs/2611.19529
作者: Ettore Caruso,Marco Siino,Giacomo Morabito
类目: Artificial Intelligence (cs.AI)
备注: 14 pages, 6 figures, 9 tables. Submitted to IEEE Internet of Things Journal

点击查看摘要

Abstract:The Internet of Drones (IoD) needs secure telemetry and proximity sensing within the SWaP-C constraints. We show that a single blind neural receiver driven by one chaos-based waveform performs both functions without a separate radar transceiver. In telemetry mode it decodes a peer’s burst blindly; in radar mode it regenerates the per-burst chaotic reference locally and ranges the echoes of its own transmission. Ultra-CAN pairs a shared 64 -dimensional representation with two ultra-light backbones, an attention-pooled Conv1D network and a multi-head QKV network ( 27.6 k-- 44.2 k parameters), feeding communication and sensing heads. Both ship as int8 TensorFlow Lite Micro firmware; the Conv1D receiver decides a burst in 16.9 ms on a 240 MHz microcontroller with 24.2 KiB of activations, a measured energy of 2.2 - 3.4 mJ per burst, and float-equivalent decisions, sustaining up to 59 bit/s of telemetry or a 7.4 Hz range track. The two variants rank first and second among learned demodulators in four aerial swarm scenarios, with a pooled BER down to 4.6 \times 10^-4 on the single-echo link ( 2.4 \times 10^-3 with three echoes). Its sensing head identifies true obstacles in 66.8% of peer-laden bursts (vs 42% matched filtering, 44% OFDM-ISAC, 48% peak-picker ceiling). A peer-exclusion gate cuts peer reports from 4.4% to 0.02% of answered bursts, answering 23.5% of total bursts ( 75% on the exact tap) with 0.70 m/s velocity tracking. In-domain refitting recovers telemetry BER by up to three orders of magnitude under 3GPP channels, jamming-aware training restores the link, and frequency agility adds a 6 – 9 dB margin against partial-band jamming.

[AI-163] VAST: Compiling Grounded Vulnerability Scenarios into Verifiable Multi-Step Environments for Evaluating LLM Red-Team Agents

链接: https://arxiv.org/abs/2611.19465
作者: Zhen Wan
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: 28 pages, 2 figures, 13 tables, including Appendices A.1-A.7

点击查看摘要

Abstract:Evaluating Large Language Model (LLM) security agents requires reproducible environments whose workflow difficulty can be varied and whose failures can be diagnosed. We present VAST (Verifiable Adversarial Scenario Training Evaluation), a declarative range compiler separating vulnerability scenarios from a compiler-introduced workflow-hardening layer. Following a post-hoc code and AST audit across all topics, the main benchmark retains 109 seed edges (45 with strict upstream CVE annotations), 72 family identifiers (62 under shared-HMAC relay grouping), 81 topic configurations (76 core plus 5 bounded execution models), and 1,387 instances (359 canonical four-gate-revalidated specs plus 1,028 seed variants) in a 42:39 partition. On the TopicServer compiler stratum, a budget-matched 270-episode prompt contrast across 10 compiled mechanism-family clusters (30 P2/P3/P4 instances, 40 turns, temperature 0.0) shows that glm-5.3, grok-4.7, and deepseek-v4-pro solve 12/90 instances under contract_only and 14/90 under generic_hint (cluster sign p=1.00), whereas supplying the shared workflow protocol in full_guide lifts completion to 76/90 (28/30, 23/30, 25/30; cluster sign p=0.00391). On the separate ranges_v2 ledger, a 126-episode evaluation across 6 non-TopicServer workflows and multi-container Docker/Linux stacks solves 88/126 (69.8%) (28/42, 31/42, 29/42), while an 18-episode intervention on asynchronous Redis worker chaining (v2_02) lifts solves from 0/18 to 18/18 (100%) in 4-6 turns when installed worker modules are disclosed in GET /sched/directory. Finally, a three-seed (s in 42, 1024, 2025) LoRA recipe comparison on Qwen2.5-Coder-7B-Instruct across 37 lineage-disjoint families (231 episodes on 33 held-out instances) yields 0/33 for Base, 1.67 +/- 2.89 / 33 for Golden SFT, and 4.67 +/- 2.31 / 33 for Header-Recovery SFT, with all solves on v2_03. Comments: 28 pages, 2 figures, 13 tables, including Appendices A.1-A.7 Subjects: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI) ACMclasses: D.2.5; K.6.5; I.2.11 Cite as: arXiv:2611.19465 [cs.CR] (or arXiv:2611.19465v1 [cs.CR] for this version) https://doi.org/10.48550/arXiv.2611.19465 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-164] Program Distillation: Transferring LLM Capability from System 2 to System 0

链接: https://arxiv.org/abs/2611.19461
作者: Pujun Zheng,Wanying Ren,Guoxiu He
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Recent studies leverage large language models (LLMs) to construct executable programs through environment interaction, feedback analysis, and code revision. However, prior work has mainly treated LLMs as optimizers, overlooking their ability to transfer knowledge across system levels. We propose Program Distillation (PD), a framework that formalizes the transfer of abilities from System 2 to System 0. During learning, the LLM provides task knowledge and draft strategies; at deployment, a student program executes the resulting policy. We show that this approach can improve performance while reducing inference time. We introduce PDBench, a benchmark comprising 27 discrete- and continuous-control tasks with fixed initial policies, and PDMetaHarness, a model-harness-independent meta-system for monitoring and guiding the learning process. The strongest model achieves a normalized mean score of 0.728, surpassing PPO (0.656) and Jev-like decision models (0.239), while running exclusively on CPUs and reducing inference costs. Our analysis shows that performance is largely driven by the first breakthrough during learning, and structural modifications have a slightly lower success rate than parameter modifications but typically yield larger gains. We further introduce style distillation, which enables interpretable stylistic control over the learning process. Overall, we present an analytical framework for PD and demonstrate its practical viability.

[AI-165] CLEW: Log-Anchored Backward Tracing for Root-Cause Code Localization

链接: https://arxiv.org/abs/2611.19448
作者: Wenji Bai,Muhammad Waseem,Jaakko Peltonen,Pekka Abrahamsson
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Failures in industrial software systems may involve multiple components and be difficult to reproduce outside the production environment. Root-cause code localization therefore often relies on issue descriptions and runtime logs. However, log evidence may be noisy or incomplete, and the code reporting a symptom may differ from the code containing the underlying defect. These conditions make it challenging to identify relevant log evidence and determine which functions could have caused the reported failure. To address this gap, we introduce CLEW, an agentic approach that combines rule-based log-to-code anchoring with LLM-guided backward code tracing. CLEW organizes selected observations into diagnostic states, guides source exploration from anchored symptom functions, and returns ranked candidate functions, an affected module, and an explanation comprising a proposed mechanism path and associated log events. For evaluation, we construct LogLoc, a benchmark comprising 192 real-world issue and fault-injection cases. Against three baselines, CLEW records the highest observed Acc@1 (77.3%) and MRR (81.3%) on LogLoc-Issues and matches the leading baselines at 97.6% Acc@1 on LogLoc-Airflow. Averaged over all 192 cases, CLEW’s wall-clock time is approximately 73% lower than LocAgent’s at a similar average inference cost. No aggregate degradation is observed under the tested perturbations, with modest changes in Acc@1 and MRR across the evaluated seeds and base models. These results suggest that CLEW can combine function-level localization with evidence links and runtime efficiency under the evaluated conditions.

[AI-166] PMR: MoE Routing with Shared Prototypes and Congestion Pricing

链接: https://arxiv.org/abs/2611.19413
作者: Wending Wang,Binqi Huang,Ju’an Li,Yu Guo,Lei Duan,Yuqi Ouyang,Dengchun Li
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Conventional linear routers in sparse mixture-of-experts models select experts directly from token-expert matching scores, coupling input matching with expert allocation. We propose Prototype-Mediated Routing (PMR), a two-stage router with shared prototypes. PMR first computes each token’s distribution over prototypes, then uses it to mix prototype-conditioned expert distributions according to the law of total probability. This formulation models input matching and expert allocation separately, explicitly represents many-to-many associations between prototypes and experts, and provides separate interfaces for context adaptation and load regulation. Within this framework, short-term states summarize preceding tokens to adapt the prototypes, while congestion prices updated from dispatch feedback achieve loss-free load balancing by adjusting the conditional expert distributions. We use a two-pass blockwise approximation to accelerate short-term state processing, achieving over 20 times the training throughput of the eager tokenwise PMR implementation while maintaining high agreement in routing decisions. Across three pretraining configurations with 0.43B, 1.18B, and 0.48B total parameters, PMR improves mean zero-shot accuracy on seven benchmarks over linear routing baselines with matched expert architectures by 1.54, 2.65, and 1.45 percentage points (pp), respectively. The corresponding gains on ARC-Easy are 6.71, 5.80, and 3.92 pp. In parameter-efficient fine-tuning of 7B-8B backbones, PMR’s improvements over MixLoRA vary across backbones, with gains of up to 1.76 pp in mean accuracy across four tasks.

[AI-167] Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States

链接: https://arxiv.org/abs/2611.19403
作者: Shubham Mishra,Atharv Sharma,Dhruv Kumar
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: Under review

点击查看摘要

Abstract:A language model’s confidence on a physics problem deserves trust only if it ignores details irrelevant to the answer. We introduce DECOY (Distractor-Embedded Certified cOunterfactuals for phYsics): 155 families of four prompts differing in one irrelevant detail. A symbolic solver certifies that each edit preserves the answer. Families are labelled by the largest drop in the model’s log-probability of the certified answer, using a threshold fixed before probe training. Qwen2.5-7B-Instruct is sensitive on 50.0% of the 140 families yielding labels, and Llama-3.1-8B-Instruct on 22.9%. Effects are largest when distractor values lie numerically close to the answer. A linear probe on Qwen’s layer-18 hidden state, averaged over four renderings, predicts sensitivity before generation at AUROC 0.760 with whole families held out. A separately tuned single-rendering probe reaches 0.747. A frozen layer-16 probe reaches 0.820 pooled AUROC on 15 families built after training, though within-cue-type AUROC is lower (0.667-0.6875). Under a paired family bootstrap, the probe performs on par with a fitted token-entropy baseline (AUROC 0.727) and achieves the lowest Brier score among tested predictors (0.204 versus 0.214). Knocking out four attention heads at layer 16 cancels sensitivity on a 20-family panel, localising it to a narrow mechanism. These results show that Qwen’s confidence can track irrelevant details, that this failure is detectable before generation, and that targeted intervention can suppress it. DECOY provides certified ground truth and paired-tested baselines for evaluating hidden-state monitors.

[AI-168] SARGO: Agreement-Gated Reinforcement and Suppression via On-Policy Self-Guidance for Visual Hallucination Mitigation

链接: https://arxiv.org/abs/2611.19388
作者: Jinyu Li,Zhiming Zhang,Hao Fang,Hongyao Yu,Bin Chen,Albert Ma,Shu-Tao Xia
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Vision-language models remain susceptible to visual hallucinations, often mixing grounded information with unsupported claims in the same response. Response-level reinforcement learning assigns a shared advantage to all tokens in a response, causing negative updates to suppress correct information alongside unsupported claims. To reduce hallucination while preserving useful visual information, we introduce SARGO (Self-guided Advantage-Routed Grounding Optimization), a post-training framework for selective reinforcement and suppression through on-policy self-guidance. Direct and reciprocal weighting gives better-supported tokens stronger reinforcement on positive updates and weaker penalties on negative updates. Agreement between overall and hallucination-specific advantages selects when to apply this guidance. At the response level, a reference-calibrated reward couples factuality with the retention and expansion of grounded information, discouraging lower hallucination scores obtained through omission. Experiments on Qwen3-VL-4B, Qwen3-VL-8B, and Gemma3-4B demonstrate that SARGO outperforms the evaluated mitigation baselines across Obj-Hal, AMBER, THRONE, and SHR evaluations. SARGO achieves these gains while preserving useful visual information and general multimodal capabilities.

[AI-169] Does a completed simulation run mean a usable urban wind result for AI agents ?

链接: https://arxiv.org/abs/2611.19385
作者: Zhongke Qu,Yunwei Zhang,Junwei Su,Yabin Zhang,Chengwei Li,Qingxiang Shui,Hailong Qu,Zhaolin Gu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:AI agents based on large language models (LLMs) can now configure and run computational fluid dynamics (CFD) solvers, but a completed run does not ensure that delivered metrics can be recomputed from raw outputs. This gap was examined for urban wind simulation, where multi-case comparisons must hold non-target conditions fixed and reported values must be traceable to raw samples. UrbanWind-AgentBench, comprising 200 tasks and 367 cases in 10 engineering task families, was constructed. Public task contracts specify variable factors, fixed conditions, numerical thresholds and deliverable requirements, and two-layer scoring separates a structural score for deliverables, run provenance and raw evidence from a numerical score obtained by independent recomputation against references. WindPilot was then proposed as an execution framework that incorporates contract parsing, execution provenance, deterministic verification and bounded acceptance review into the LLM execution loop. On 30 OpenFOAM tasks with independent references, at least the structural maximum of 60 points was reached by a base agent on 28 tasks, but numerical credit was obtained on only 6. With the same model name and budget, WindPilot increased the number of tasks with numerical credit to 17, full-score tasks from 0 to 3 and the mean task score from 62.5 to 72.5. Mean action steps fell from 130.4 to 104.9, and mean token consumption decreased by 12.9%. The mean gain for atmospheric boundary layer (ABL) tasks was 18.5 points. A completed simulation run is therefore not sufficient for a usable urban wind result; usability depends on checks that tie reported metrics to raw evidence during execution.

[AI-170] One Run Is Not Enough: Reminding a small Gemma 4 coding agent of its last finding made no measurable difference in a paired test

链接: https://arxiv.org/abs/2611.19359
作者: Patrick Passos
类目: Artificial Intelligence (cs.AI)
备注: 10 pages, 2 figures. Code, data and the commit history of the pre-registration: this https URL

点击查看摘要

Abstract:When a coding agent runs out of room, its earlier findings are cut and it forgets them. We tested a simple fix: automatically putting the agent’s last finding back in front of it. We fixed the rules in advance and ran 98 paired coding tasks with and without this reminder, using Gemma 4 31B (QAT, 4-bit) in the Gemma 4 Developer Agent harness. The reminder made no measurable difference: across all 98 tasks, 47 were solved with it and 48 without (-1.0 points, 95% interval [-8.3, +6.3], exact McNemar p=1.0). The primary outcome (a documented deviation from our preregistration), the first 48 pairs, was 29/48 with memory against 25/48 without (+8.3 points, [-2.1, +18.3], p=0.22); the extension of 50 tasks, decided after that result, pointed the other way (18/50 against 23/50, -10.0 points, p=0.125), so the sign did not replicate. With 19% of tasks flipping between two identical no-memory runs, the combined interval excludes harms larger than about 8 points and gains larger than about 6, and the design had 80% power only for effects of about ±12 points or more. The model almost never wrote notes (2 of 16 and 4 of 32 runs), and memory cost context: the first request is larger and something was injected at only 47% to 65% of compactions. The registered secondary, repeated actions per compaction, was also null. Two directional hypotheses found on Modal did not replicate on Kaggle. We also report an A/A calibration and a validity funnel (129 tasks, 105 valid). Together these results show that single-run memory claims on task sets of this size sit inside run-to-run noise. The treatment is a harness mechanism and does not test agent-written memory.

[AI-171] SMELL: Sparsity-aware Memory Efficient Federated Fine-tuning for Long-context LLM s

链接: https://arxiv.org/abs/2611.19349
作者: Yin Xu,Yongbo Yang,He Sun,Miao Sun,Mingjun Xiao
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Distributed, Parallel, and Cluster Computing (cs.DC)
备注: 14 pages, 5 figures, 5 tables

点击查看摘要

Abstract:Federated fine-tuning enables Large Language Models (LLMs) to collaboratively adapt to distributed private data. However, long-context fine-tuning on resource-constrained devices faces substantial challenges arising from excessive activation memory, heterogeneous token-importance patterns across clients, and communication overhead. To address these challenges, we propose SMELL, a Sparsity-aware and Memory-Efficient federated fine-tuning framework for Long-context LLMs. SMELL integrates three complementary mechanisms: Local Sparse-Forward Adaptation (LSFA), Consensus-Anchored Sparsity Coordination (CASC), and Communication-Efficient Progressive Update Recovery (CPUR). These mechanisms jointly enable memory-efficient local adaptation, heterogeneity-aware sparsity coordination, and lightweight model-update transmission. Experiments on OPT-350M and LLaMA2-7B demonstrate that SMELL reduces peak memory consumption by up to 88.35% compared with conventional fine-tuning and 13.25% compared with forward-only LoRA fine-tuning at a 16K context length. Furthermore, SMELL improves fine-tuning performance under certain non-IID settings and supports compact update transmission with progressive reconstruction, demonstrating its potential for resource-efficient federated long-context LLM adaptation.

[AI-172] Information Limits of Replay-Free Recurrent State Migration

链接: https://arxiv.org/abs/2611.19345
作者: Khizar Qureshi
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:When a recurrent model is replaced, its old state may not determine what the new model would compute from the same history. We characterize the information that must be retained before the replacement is known. For a time-invariant affine linear family, the minimum continuous causal memory for order- k local migration equals the reachable dimension of a lifted coefficient system minus the information in the anchor state. A controllable d -state model with scalar input needs at most kd additional real coordinates, even when arbitrary matrix replacement directions are revealed only after streaming. A single rank-one direction can already require all kd . Conditional singular values and auxiliary-bit bounds distinguish this local information law from recovery at a fixed tolerance. We also give invariant-subspace and future-output constructions. Controlled experiments compare the resulting representations. In a small language model, equal-memory cubic interpolation improves mean state fidelity over third-order coefficients but provides no meaningful predictive gain. An analytic history basis preserves useful adaptation more economically in a coupled identification task. With repeated slow poles, eight history coordinates meet normalized output error 10^-4 across five fitted seeds despite substantial full-state error. Splitting the poles raises the smallest tested successful width to 64.

[AI-173] From Mutation to Medication: Characterizing Failures of Clinical LLM s to Respect Genomic Resistance ALT ML4H

链接: https://arxiv.org/abs/2611.19339
作者: Negin Ashrafi,Roxana Daneshjou
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Genomics (q-bio.GN)
备注: Accepted for publication in the Proceedings of the Machine Learning for Health (ML4H) 2026 Symposium

点击查看摘要

Abstract:Language models are increasingly proposed to support molecular tumour boards, but may recommend cancer therapies despite genomic evidence of resistance. We evaluated ten models on 1,464 resistance cases from five curated oncology knowledge bases, collecting 58,910 judgments. Models endorsed resisted therapies in 21.8-81.4% of cases. However, raw rates can reflect a tendency to agree rather than an ability to distinguish resistance from drug sensitivity: one open-weight model endorsed 81.4% of resistance cases and 82.3% of sensitivity controls. We therefore compared endorsement within matched gene and evidence-level groups. This analysis is exploratory because it was specified after the pre-registered analysis, which yielded null-to-weak results. Measured failures depended strongly on the reference knowledge base and evidence strength. When a second source independently confirmed resistance, four models across three vendors endorsed at most 2.4% of cases; at weaker evidence tiers, discrimination was not measurable. Models sometimes acknowledged resistance yet recommended the therapy. Repeating a model’s own resistance conclusion reduced endorsement to 0.0-3.5% on cases where it had stated resistance. Falsely attributing the opposite conclusion increased endorsement above baseline for all seven hosted models, reaching 95.2% for GPT-4o; an instruction without an attributed conclusion changed endorsement by only 0.9 percentage points. Providing the correct curated evidence reduced unqualified strong recommendations to 0.3-3.5%, testing evidence use under an oracle condition rather than retrieval.

[AI-174] NI-ACT: A Neural Intent Interface for Brain-Embodied Manipulation

链接: https://arxiv.org/abs/2611.19332
作者: Yinghao Li,WeiHsien Lee,Zeyu Zhang,Canyang Zhao,Ke Fang,Xing Li,Bing Liu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Intracortical brain-computer interfaces can drive robotic arms, but they keep the user inside the fast sensorimotor loop: every approach, correction, and grasp must be neurally specified. Whether intracortical population activity can instead serve as a task-level condition for a learned embodied policy, without symbolic commands or hand-designed skill libraries, remains unknown. NI-ACT addresses this question with a hierarchical brain-embodied framework: a brief motor-cortical signal is read once as a persistent intent, and a visuomotor policy then closes the perception-action loop from vision and proprioception. In a counterfactual grasping task where the scene cannot specify the target, the neural intent alone determined which object the robot grasped; replacing the intent changed the behavior accordingly. Passing the full population latent as a dedicated intent token, rather than a one-hot decoded direction, raised grasp completion from 69% to 81% over 132 held-out scenes, whereas a continuous direction alone did not improve on a one-hot label. Where the signal was read also mattered: in one dual-array session, with the readout window matched, premotor (PMd) intent completed more grasps than primary motor (M1) intent (77% vs. 53%), and the conventional post-cue M1 readout reached 42%. Preparatory activity was itself sufficient: in a center-out task, intent read only from PMd before the go cue (that is, before any movement) drove complete grasps, and in one recording it completed more grasps than the same animal’s post-cue readout. These results show that how a neural signal is represented, and where and when it is read, jointly determine how effectively it conveys movement goals to an embodied agent.

[AI-175] Benchmarking the State of AI for Crop Yield Forecasting: A Global Assessment Across Modeling Paradigms KDD’27 KDD

链接: https://arxiv.org/abs/2611.19330
作者: Michiel Kallenberg,Philip Janz,Christoph Jörges,Vageesh Saxena,Pratishtha Poudel,Mohammed Musthafa Rafi,Ioannis N. Athanasiadis
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 11 pages, 2 figures, 2 tables. Accepted for publication in Proceedings of the 33rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '27), August 2027, San Jose, CA, USA. Code is available at this https URL and an interactive results dashboard is available at this https URL

点击查看摘要

Abstract:Accurate crop yield forecasting is essential for food security, agricultural policy, and commodity markets. Although statistical, process-based, and machine learning approaches have all been proposed, their relative strengths remain poorly understood due to the lack of large-scale, standardized comparisons. We address this gap through a comprehensive benchmark of five modeling paradigms – statistical baselines, process-based models, conventional feature-engineered machine learning, deep sequence models, and tabular foundation models – using the CY-Bench benchmark spanning 63 country-crop datasets (12,000 administrative regions) across six continents. Beyond conventional pooled accuracy metrics, we introduce a multi-dimensional evaluation framework that separately quantifies overall, spatial, temporal, and anomaly prediction skill. Across multiple crops and countries, feature-engineered machine learning and tabular foundation models achieve the strongest overall predictive performance. Data-driven models reduce NRMSE in 72% of countries growing maize and 78% of countries growing wheat, with median reductions of 8.8% and 11.9%, respectively. Across all modeling paradigms, spatial prediction is substantially easier than temporal and anomaly prediction. While data-driven models markedly improve temporal and anomaly prediction over statistical baselines, accurately forecasting year-to-year yield variability remains challenging. These findings suggest that future progress in large-scale crop yield forecasting will depend on the improved characterization of the environmental and management factors driving inter-annual yield variability, founded on effective representations for crop yield forecasting systems that support agricultural decision-making, early-warning applications, and food security planning

[AI-176] he Data Debt of Stockouts: Partial Identification and Minimax-Regret Inventory Control

链接: https://arxiv.org/abs/2611.19323
作者: Jinbiao Chen,Shuang Jin,Hanzhang Qin
类目: Artificial Intelligence (cs.AI); Systems and Control (eess.SY)
备注:

点击查看摘要

Abstract:Problem definition: Inventory decisions select which demand enters sales logs. Unrecorded forecasts can leave demand frequencies unidentified even when stocking levels are recorded. We study which multistage ordering decisions these logs can justify and what finite samples can guarantee. Methodology/results: Nested stockout thresholds yield exact sets of compatible demand laws and finite representations for policy evaluation. We choose one deterministic inventory rule to minimize its largest expected cost gap from the informed optimum under the same law. This population value is data debt for the specified information and policy class. Under independent demand, carryover can create positive debt despite common optimal actions in every isolated period. Finite samples add learning error above debt. The required sample count grows linearly with the selection bound, with upper and lower orders matching up to logarithmic factors. This rate requires fixed demand and cost bounds, sufficient revelation, and a uniformly positive probability of clearing all feasible stock across the confidence sets. Controlled and semi-synthetic studies separate policy improvements from tighter guarantees. Managerial implications: More records can tighten guarantees while debt persists. Deeper revelation can establish common optimality without fully identifying demand. Using these guarantees requires validated logging, a credible selection range, and a regret tolerance in cost units. Information acquisition must distinguish tighter assurance from operating savings. Subjects: Artificial Intelligence (cs.AI); Systems and Control (eess.SY) Cite as: arXiv:2611.19323 [cs.AI] (or arXiv:2611.19323v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.19323 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-177] Hardening Agent ic Web Populations Through Multi-Criterion Competitive Co-Evolution: Lessons from Wargame Strategy Optimization NEURIPS NEURIPS2026

链接: https://arxiv.org/abs/2611.19319
作者: Shashank Raj,Kalyanmoy Deb
类目: Artificial Intelligence (cs.AI)
备注: Accepted at the 1st NeurIPS 2026 Workshop on Agentic Web: Decentralized, Continually-Adapting Agent Ecosystems ( this https URL )

点击查看摘要

Abstract:Agents on the open web will face adversaries that keep adapting, so a defense that is only tested once will eventually be found and broken. We propose treating jailbreak attacks and safety guardrails as two populations that evolve against each other, rather than as a fixed attacker probing a fixed defense. This idea comes directly from wargame strategy optimization, where two-population competitive co-evolution with alternating evolution cycles and cross-play evaluation has already been shown to produce strategies that are robust to a moving opponent, and where a companion line of work has shown that the same machinery can measure how much competitive pressure a population must absorb. However, real-world applications can never rely on a single criterion as a key performance indicator, rather must use multiple conflicting criteria, such as, cost and efficiency. We describe how this mechanism transfers to language model safety, argue why the resulting multi-criteria, competitive and coevolutionary arms race is a genuinely networked, population-level problem rather than a two-agent toy setting, and lay out the open problems that stand between the wargame version of this idea and an agentic web version of it.

[AI-178] Zero-Shot Weight-Space Programming for Prediction and Classification of Target Variables in Structured Datasets

链接: https://arxiv.org/abs/2611.19292
作者: Eduardo C. Garrido-Merchán
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Weight-space programming fixes every parameter of a feed-forward network by hand, with no gradient descent anywhere in the construction. Every demonstration so far starts from a target function handed to the model as a rule system, making it a compilation problem. A supervised dataset withholds that specification, so the procedure must do the statistical inference itself and then express the result in the parameterisation it was given. We ask whether a large language model can do both. It is shown a small dataset in full, with the architecture of a feed-forward network, and writes out every weight and bias as a literal number, with no optimizer, no gradient, no code execution, and no chance to evaluate it first. Each construction is sealed and pre-registers its expected score. The graded suite runs from two toy problems, through nine synthetic problems disguised by a hidden rotation, to the classical benchmarks and one real engineering regression. The frozen network matches the closed-form optimum wherever that optimum is linear, reaches the irreducible error of the generating process on three rungs where it is not, and beats every fitted baseline on two of those, including the same architecture trained by gradient descent. It fails on two rungs, both times because the structure to be found is oblique to the coordinates it analysed. Three further constructors separate two failure modes a single score conflates: one returns no parameter vector at all on the hardest rungs, another delivers everywhere while writing networks worse than predicting the mean. One of those failures proves avoidable: another constructor solves the oblique-hinge rung by a margin that survives multiple-comparison correction. What predicts whether the hidden layer earns its keep is the training error of a linear separator, which the composite complexity index misses.

[AI-179] Beyond Clumsy Agent ic Tool Calling: How Cueing and Spotting Make the Difference

链接: https://arxiv.org/abs/2611.19275
作者: Xindi Yang,Yan Liu,Jiajun Sun
类目: Artificial Intelligence (cs.AI); Software Engineering (cs.SE)
备注: The code of reference implementation and instructions for reproducing the experiments are available in the following repositories: this https URL and this https URL

点击查看摘要

Abstract:LLM-based task-oriented agents depend on tools to interact with their environments, but may overlook execution feedback, repeat ineffective calls, or persist with unproductive strategies. These behaviors can hinder agents’ ability to use tools autonomously and sustain task progress. We characterize tool-use symptoms at two temporal granularities: local clumsy tool-use behaviors and tool-use detours spanning multiple interactions. Towards greater autonomy in tool use, we present ToolPulse, an online middleware supporting autonomous tool use through cueing and spotting. ToolPulse tracks execution state and temporal interaction patterns, monitoring proposed calls before dispatch and inspecting execution feedback prior to the agent resumes reasoning. It combines state-based rules, semantic sequence analysis, and interaction-feature similarity to identify intervention opportunities and provide verbal reminders, structured feedback, or context-specific advice from auxiliary models. Our process-oriented evaluation distinguishes trigger validity, behavioral cessation, and substantive correction alongside task outcomes and resource costs. Using SWE-agent with three backbone models, we evaluate 1,350 trajectories on 90 Python, C++, and Java bug-fix tasks, comparing unassisted execution with four assistance strategies. Substantive responses follow 61.6-90.9% of valid triggers for tool-use detours. Concise reminders achieve the highest observed aggregate substantive-response rate; more elaborate assistance offers symptom-dependent benefits and can be less effective. Aggregate repair outcomes remain close to baseline, with token savings in selected settings. These findings support proactive runtime assistance for autonomous tool use. The gap between behavioral adjustment and aggregate repair gains calls for careful decisions on intervention timing and assistance strategy in autonomous agents.

[AI-180] HospitalBench: Diagnosing Where and Why Embodied Agents Fail on Long Hospital Task Chains

链接: https://arxiv.org/abs/2611.19274
作者: Gaytri Jena,Kapil Wanaskar,Aman Chadha,Vasu Sharma,Vinija Jain,Amitava Das
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI)
备注: 22 pages, 4 figures, 23 tables

点击查看摘要

Abstract:Long-horizon embodied tasks fail in ways that end-task success cannot see. HospitalBench is a diagnostic benchmark of ten hospital chores in NVIDIA Isaac Sim, each two unscored warm-up steps followed by a causal chain of five scored sub-tasks in one continuous simulation: the state one step leaves (an object pose, a choice, a fact seen once) determines what the next step must do, unlike benchmarks that place sub-tasks side by side. Vision-language agents act through skill-level tools from camera images. A scripted robot that makes only the calls a model gets is the solvability check of a chore. From each trajectory the benchmark reads where a chain breaks, what the agent did after a failed call, and how far its own position estimate drifts from the true one. This version reports first results only: 15 of 60 planned runs (three models at high effort, one run each on chores T1 and T10 from both start states and T3 from one), with scripted-robot proofs on three of the ten chores. Of the 15 runs, 2 passed all five sub-tasks. Runs still going after sub-tasks 1 to 5 were 15, 8, 4, 2, 2. Of 189 failed calls, 168 were followed by a different call and 111 by a call that did not fail. These are first results and not a full benchmark. We will release the scenes, task chains, scripted robot, evaluation harness and trajectories.

[AI-181] Agent ic On-Policy Distillation Implicitly Reshapes LLM s ReAct Patterns

链接: https://arxiv.org/abs/2611.19244
作者: Deyang Kong,Zhaoyi Li,Sijian Ren,Yuan Wei,Ying Wei,Defu Lian,Shikun Zhang,Wei Ye
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:On-policy distillation (OPD) transfers specialized capabilities to language-model agents, but its cross-domain effects on interaction remain unclear. We investigate how OPD reshapes ReAct interaction patterns across mathematics, ALFWorld, and \tau^2 -bench, focusing on reasoning, actions, and observation acquisition. General-reasoning supervision reduces mathematical tool attempts from 70.4% to 14.4% while increasing accuracy from 16.7% to 38.0%; code-interpreter supervision improves household actions and retail queries. In contrast, ALFWorld supervision reduces \tau^2 -bench accuracy from 41.5% to 6.6%, accompanied by frequent tool calls that violate user-information dependencies. To address negative transfer using only source-domain training trajectories, we explore selective OPD guided by teacher preference changes before and after domain-specific RL. Retaining the top 95% of positions raises \tau^2 accuracy from 11.5% to 29.9% with a source-learning trade-off.

[AI-182] Auditable Authorization for LLM Agents : Empirical Evidence Across Model Harness and Workflow Layers

链接: https://arxiv.org/abs/2611.19237
作者: Bangjie Deng,Xiaoxuan Lu,Wenjun Chen,Chengge Wen,Yulin Wang,Jiayue Meng,Yao Yu,Liying Liang,Weijia Han,Ziwei Hong,Zidong Yu
类目: Artificial Intelligence (cs.AI)
备注: 6 pages, 4 figures, 4 tables

点击查看摘要

Abstract:Human approval provides effective control only if rejected work stops and authorized work completes. We formalize refusal safety and bounded authorization completeness, and audit three execution boundaries: the model selects a call, the harness schedules and resumes it, and the workflow routes the decision to a protected effect. Paired decisions, exact traces, and native business-state witnesses support experiments with three models, five prompt arms, four runtime approval designs, two additional framework replays, and 100 public workflow sources. The paired model panel leaves seven of 300 approvals unfinished despite zero denied effects; parallel confirmation loses an approved branch in three tasks per model, restoration omits 336/672 CrewAI approvals, and four of 59 comparable n8n workflows execute after Reject. Comparing intent, generated calls, dispatch, and native effects identifies the first observable divergence and its repair target. This yields a practical audit contract that makes both refusal safety and authorization completeness independently verifiable, converting human approval from a procedural formality into an enforceable control boundary between intent and real-world effect. Comments: 6 pages, 4 figures, 4 tables Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2611.19237 [cs.AI] (or arXiv:2611.19237v1 [cs.AI] for this version) https://doi.org/10.48550/arXiv.2611.19237 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[AI-183] CompoundingKernel: Self-Improving from Zero to First

链接: https://arxiv.org/abs/2611.19230
作者: Qian Qu,Hongyang Wang,Wenbo Chen,Yichuan Li,Jingguo Liu,Xian Li,Xuan Wang,Ruochen Wen,Leo Fan,Yiming Yan,Zheshen Wang,Benoit Dumoulin
类目: Artificial Intelligence (cs.AI)
备注: 47 pages, 23 figures, 6 tables. URL: this https URL

点击查看摘要

Abstract:We present CompoundingKernel, a self-improving agent system that writes GPU kernels by search. As of 2026-11-11, its generated kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200). They have held #1 on L1 since 2026-10-27, albeit a short loss on 2026-11-07. All 235 problems have a site-verified correct kernel; on the 209 public-board problems they are 4.40x faster than the benchmark’s reference baseline. The search starts from zero: agents get each problem’s PyTorch definition but no reference kernels or official scores. A per-problem archive outside the model holds every correct kernel and its lineage. An LLM controller reads it, picks a parent or a new design, and gives each parallel child agent one direction. Checks outside the model decide correctness. At equal wall time, it beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts. Over a thousand agents ran at once, and one warm restart fix raised their evaluation rate 3.6x. We find that the rate of reliable measurements, not the number of agents, limits how far such a search scales. Our code and kernels are available on URL.

[AI-184] Can an Agent Rethink What It Knows? Selective Scene Memory Updates in a Changing World

链接: https://arxiv.org/abs/2611.19227
作者: Zhekai Ren,Jiaming Wu
类目: Artificial Intelligence (cs.AI)
备注: 7 pages, 6 figures, 2 tables

点击查看摘要

Abstract:Task configurations in factories, warehouses, and everyday physical environments change over time. Physical AI systems therefore need to incorporate new facts into scene memory while preserving valid information about equipment, objects, and operating constraints. We study updates to this parametric semantic-memory layer and present a joint evaluation of correction and preservation under finite update budgets. At fixed current fact count and memory parameter count, the protocol combines independent single-fact updates, budget trajectories, old-to-new answer diagnostics, and first stable success. Using a frozen 4B language model with late-layer LoRA, we evaluate seven preparation conditions and six update budgets on four-field synthetic semantic configurations. In two exposure-matched history conditions, increasing the budget from eight to 32 updates raises joint success from 6.25% to 94.79% and 91.15%, respectively, while exact obsolete-value outputs fall from 174/192 and 166/192 to zero. Constant32, which matches preparation steps but repeatedly trains on unchanged content during the first 32 events, achieves higher early joint success at eight steps (53.65%); at 32 steps, it reaches 85.42%, below both changing-history conditions. These results distinguish rapid correction under a small budget from simultaneous correctness of the new fact and the other three facts after further training. For both changing histories, most cases remain jointly correct at all subsequent measured points from step 16 onward. Output analysis shows that short-budget failures primarily retain the old value; repeated training on the new fact together with the other current facts progressively replaces these obsolete answers.

[AI-185] Neural Dual Bounds: Valid-by-Construction JGLP Warm-Starts for MAP and Constrained MAP NEURIPS2026

链接: https://arxiv.org/abs/2611.19218
作者: Akshay Vyas,Shivvrat Arya,Brij Malhotra,Tahrima Rahman,Vibhav Gogate,Nicholas Ruozzi
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: NeurIPS 2026 (Spotlight). 40 pages, 6 figures

点击查看摘要

Abstract:We present Neural Dual Bounds, a calibration-free framework that produces deterministic upper bounds for MAP inference in probabilistic graphical models using a single neural-network forward pass. The framework trains a network to output the edge messages of a mini-bucket style upper bound on a given join-graph decomposition in a self-supervised manner by directly minimizing a smooth dual objective. The output of this network can be used at test time to warm-start a standard JGLP optimizer. Across diverse benchmarks, this warm start yields dual gaps that are orders of magnitude tighter than those from zero initialization, with the largest gains in the early-iteration regime. We also apply our framework for the constrained MAP (CMPE) task, where neural warm starts with test-time adaptation are comparable to the adaptive baseline on Bayesian networks and achieve the tightest dual gaps, by one to two orders of magnitude, on the grid and segmentation families.

[AI-186] Mem2Fair: Online Fairness Control for Text-to-Video Generation via Fair Memory

链接: https://arxiv.org/abs/2611.19215
作者: Pengyu Chen,Jiaming Zhang,Xiaohua Feng,Yuyuan Li,Chaochao Chen
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Text-to-video (T2V) generation models can produce systematically unfair demographic outcomes when prompts leave protected attributes unspecified. For example, prompts describing nurses without specifying gender may disproportionately generate videos depicting female nurses. Mitigating unfairness in T2V generation is particularly challenging: training-based methods that modify model internals require architecture-specific access and risk disrupting temporal consistency, while sampling-based methods often incur costly pre-deployment video generation. We present the first conditional online approachability formulation of attribute-level fairness for T2V generation. Building on this formulation, we introduce Mem2Fair, a plug-and-play framework that maintains concept-conditioned Fair Memory. Using generated videos at inference time as feedback, Mem2Fair selects prompt-level attribute interventions based on the deviation from each concept’s prescribed fairness target, without training or separate pre-deployment sampling. Experiments demonstrate that Mem2Fair consistently improves fairness over the unmodified generator and existing baselines across multiple demographic attributes, prompt settings, and target distributions, while maintaining comparable prompt-video alignment and video quality. We further demonstrate its effectiveness on commercial black-box T2V models.

[AI-187] STaSBO: LLM -Guided Strategy Planning and Configuration Selection for Bayesian Optimization

链接: https://arxiv.org/abs/2611.19214
作者: Youngseok Yoon,Peng Li
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Bayesian optimization (BO) enables sample-efficient search for expensive black-box problems, but its performance depends strongly on surrogate-acquisition configurations whose suitability varies across problems and optimization states. We take a significant step toward LLM-driven strategic management of BO with STaSBO, a task-agnostic meta-controller that plans search behavior and jointly selects surrogate-acquisition configurations, while numerical BO retains responsibility for probabilistic modeling and candidate generation. STaSBO comprises two coordinated stages: strategy planning and strategy-guided configuration selection. A structured memory system connects prior plans, configuration selections, and candidate behaviors to optimization states, allowing the controller to revise its trajectory summary and behavioral plans during planning. To navigate the large joint configuration space, the selection stage shortlists configurations using surrogate fit evidence. Prospective BO-action diagnostics then expose the candidate behaviors, supporting final joint selection against the behavioral plans. Across 12 synthetic mathematical and 10 scientific and engineering application tasks, STaSBO achieves the best average rank against 11 baselines: 2.17 and 2.60, compared with 4.08 and 4.10 for the strongest baseline in each suite, respectively.

[AI-188] On-policy distillation as a Backdoor Propagation Channel

链接: https://arxiv.org/abs/2611.19207
作者: Zane Fu,Yanzhao Zheng,Ziting Qin,Qianshan Wei,Yi Yu,Gongli Xi,Jihuai Zhu,Baohua Dong,Hangcheng Zhu,Ruohui Huang,Gang Yu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:On-policy distillation (OPD) is emerging as an important paradigm for the post training of large language models, improving student models by leveraging supervision generated online by teacher models. However, whether this process also transfers latent security risks from teacher models to student models remains insufficiently studied. In this work, we reveal a new security risk in which an attacker first implants a conditional backdoor into the teacher model through supervised fine tuning (SFT), after which OPD can transfer the backdoor to the student model with only an extremely small proportion of trigger prompt exposure. More concerningly, we find that such backdoor transfer is substantially stronger under OPD than under conventional distillation methods, suggesting that OPD may provide an efficient channel for propagating backdoors across models. Our mechanistic analysis shows that the backdoor signal is primarily transmitted through anomalous differences in the conditional probabilities produced by the teacher model for triggered and clean inputs, and is gradually internalized into the internal representations of the student model during training. Furthermore, we investigate how to suppress such anomalous conditional probability signals at the source of distillation supervision and, based on this analysis, propose an OPD training method tailored to this conditional backdoor propagation setting to mitigate the risk of backdoor transfer induced by untrusted teacher supervision.

[AI-189] What Moves a LoCoMo Score? Memory Context Answer Prompts and Judges

链接: https://arxiv.org/abs/2611.19199
作者: Nikhil Kumar Jha,Amitoj Cheema
类目: Artificial Intelligence (cs.AI)
备注: 13 pages, 2 figures. Answers, judge verdicts, human-check grades and scoring script: this https URL

点击查看摘要

Abstract:LoCoMo is a widely reported benchmark for long-term conversational memory, but published scores mix memory quality with choices of answer prompt, judge and baseline layout that are rarely reported together. We measure how much these choices move a score, and evaluate Becos, a team memory system for coding agents, as a black box under all of them: ten conversations (1,540 questions), two answer models (gpt-4o-mini, gpt-4.1-mini), two answer prompts and three LLM judges, against a whole-conversation baseline rerun under identical conditions. Changing only the judge moves a score by up to 7.2 points and changing only the answer prompt by up to 9.3. A blind human check of 150 answers finds the judges reliable when they agree, and the primary judge’s errors on disputed answers working against Becos rather than for it. Becos answers from a context with 90% fewer characters than the whole conversation. Its configuration was developed on LoCoMo itself, so there is no held-out set. With gpt-4o-mini answers it scores clearly higher than the whole conversation (79.7 vs 73.0 with the short prompt, paired 95% interval for the difference +3.8 to +9.5; 88.8 vs 77.5 with a step-by-step prompt). With gpt-4.1-mini answers the two score about the same (83.2 vs 83.1; 92.5 vs 92.1). We release every answer and verdict and the scoring script.

[AI-190] When Does Confidence Reveal Problem Difficulty?

链接: https://arxiv.org/abs/2611.19173
作者: Xiyu Bao,Langyuan Cui,Hwee Tou Ng,Chun Kai Ling
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:While confidence signals are commonly evaluated for their ability to predict a model’s own correctness, their relationship with problem difficulty remains less understood. In this work, we investigate a wide range of confidence-related signals and whether they can serve as indicators of problem difficulty across various task domains. We find that perturbation inconsistency, semantic entropy, and solution length exhibit strong associations with population difficulty than token-level uncertainty. Notably, these relationships persist even when restricting the analysis to problems correctly answered by the focal model, suggesting that such signals capture information about difficulty beyond their ability to predict the model’s own correctness. We further explore applications of these signals, illustrating the potential and limitations of using confidence-related signals to guide inference-time decisions. Our findings highlight the utility of model-specific signals as imperfect but informative proxies for population-level problem difficulty.

[AI-191] Executable Archaeology: Reanimating the Logic Theorist

链接: https://arxiv.org/abs/2611.19167
作者: David Moews,Jeff Shrager
类目: Artificial Intelligence (cs.AI); Logic in Computer Science (cs.LO); Programming Languages (cs.PL); Symbolic Computation (cs.SC)
备注: This paper builds upon a previous ArXiv paper of a similar name: https://arxiv.org/pdf/2603.13514 . This paper is NOT a mere rehash of the previous paper; appropriate citations and comparison are provided in the discussion. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:The Logic Theorist (LT), created by Allen Newell, J. C. Shaw, and Herbert Simon in 1955–1956, is regarded by some as the first artificial intelligence program. It was originally simulated by hand and written in a pseudocode called the Logic Language (LL), which was later called Information Processing Language One (IPL-I). LT was revised many times. The first machine-executed version was written in IPL-II and ran on the JOHNNIAC, and a late version was published in 1963 by Einar Stefferud as a pedagogical document and written in IPL-V, the only version of IPL that was formally released to the public. We describe a reanimation of both the IPL-I (LL) version of LT and the 1963 IPL-V version. Comparing the two reanimations illuminates the process of the invention of early AI algorithms and computer languages.

[AI-192] From Sensors to Silicon: Causal Cross-Layer Self-Healing Edge AI for Secure and Energy-Adaptive IoT Systems

链接: https://arxiv.org/abs/2611.19165
作者: Mostafa Darvishi
类目: Artificial Intelligence (cs.AI); Hardware Architecture (cs.AR); Emerging Technologies (cs.ET); Signal Processing (eess.SP)
备注: 55 pages, 27 figures, 15 tables

点击查看摘要

Abstract:The migration of artificial intelligence from centralized cloud infrastructure to resource-constrained Internet-of-Things (IoT) devices reduces communication latency and improves data locality, but it also tightly couples sensing quality, inference behavior, microarchitectural activity, hardware reliability, security, and energy management within the same physical node. Existing edge-intelligence solutions predominantly address these dimensions in isolation, making it difficult to determine whether an observed degradation originates from sensor malfunction, data-distribution shift, malicious execution, hardware faults, or resource and thermal stress. This paper presents a causal cross-layer self-healing Edge-AI framework that continuously observes synchronized telemetry spanning sensor health, inference uncertainty, processor hardware-performance counters, accelerator and memory activity, power and thermal state, and communication conditions. A temporally informed causal reasoning engine estimates the most probable origin of abnormal behavior and quantifies diagnostic uncertainty, including under concurrent perturbations.

[AI-193] FundingClaw: Automated Grant Proposal Writing Agents System

链接: https://arxiv.org/abs/2611.19157
作者: Ruizhi Liu,Siru Ge,Siran Wang,Qianli Ma,Zhipeng Zhang
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Grant proposal writing remains a manual, labor-intensive bottleneck: proposals must satisfy rigid, agency-specific templates, and the same intellectual content must be repeatedly re-encoded into incompatible formats across funding organizations. We introduce FundingClaw, a multi-agent framework built on template-driven skill selection, where a Template Parser ingests any agency’s specification and selects the appropriate writing skills for each stage, so that one research brief can produce compliant proposals for NSFC, NSF, KAKENHI, or ERC without modifying generation logic. A blueprint-skeleton strategy drafts a structured outline before any prose is written, mitigating the semantic drift that plagues long-form generation. Three sequential stages produce the rationale, research content, and research foundation sections, with multi-source verification of the PI’s record and optional PDF-grounded extraction of representative works. Optional human feedback checkpoints and a closed-loop self-review complete the pipeline. Evaluated on ten research topics with an eight-dimension rubric, FundingClaw outperforms the baseline, narrowing the gap with human level for laborious funding proposal. The code will be released upon acceptance.

[AI-194] When the Mythical Man-Month Meets Agent ic Web Systems: Structured Coordination for Scalable Development

链接: https://arxiv.org/abs/2611.19140
作者: Zhiyao Wu,Zi Liang,Haibo Hu
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:As agentic systems are increasingly deployed to support complex Web tasks, multi-agent architectures offer greater parallelism, specialization, and task coverage. However, increasing the number of agents does not necessarily improve system performance, as communication, synchronization, redundant work, and coordination overhead may offset the benefits of additional agents. This presents a fundamental scaling challenge for agentic Web systems: how can agent populations be organized to balance task performance and coordination costs? To address this challenge, we systematically investigate the relationship between agent population size, per-agent token consumption, and coordination overhead in collaborative agentic systems. Experiments on complex Web-oriented software development tasks reveal a non-linear scaling pattern: per-agent token consumption initially decreases as the agent population grows, but eventually exhibits diminishing returns. Motivated by these findings, we introduce the Surgical Team Model (STM), a structured coordination framework that organizes agents through role specialization, constrained inter-agent communication, and coordinator-mediated task aggregation. Experiments on full-stack Web application development tasks show that STM achieves competitive task completion quality and comparable token consumption relative to representative baselines, while incurring higher execution latency.

[AI-195] How Much Determinism Can Fine-Tuning Buy? Margin Regularisation Against Batch-Induced Token Flips

链接: https://arxiv.org/abs/2611.19135
作者: Charan Teja Nagisettygari,Siva Raja Sindiramutty
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Greedy decoding at zero temperature is widely assumed to be deterministic, although the exact same prompt may yield different tokens depending on the number of other prompts sharing the batch for inference. The currently available solutions operate during inference, trading determinism for throughput. Can we train a model in such a way that fewer decoding steps would be susceptible? We characterize the perturbation empirically across six model families and three numerical precisions and observe two empirical facts about the measurement: the perturbation depends exactly on the representable precision (8.00x for bfloat16 vs float16, 8192.0x for float16 vs float32, in line with predictions based on mantissa bits) and the batch invariance happens accidentally at certain tensor shapes, allowing a practitioner to achieve zero nondeterminism merely by setting the sequence length. Then we apply hinge penalization of the model’s top-1 / top-2 logit gap, scaled to twice the measured noise quantile. On GSM8K dataset it leads to a 35% reduction in token flips and divergence from 53.8% to 38.1%, at the cost of 1.3 - 2.9 accuracy points, increasing monotonically with the margin size. It does not affect the gap threshold sufficient for determinism: the logit gaps of surviving flips are unchanged (median 0.125 for all arms), as the hinge is a soft penalty and the largest hardware perturbation remains the same. We reject seven hypotheses, including a per-model viability diagnosis, holding on six models and falsified on the seventh.

[AI-196] Agent -Guided Process Verification for Mathematical Self-Correction

链接: https://arxiv.org/abs/2611.19122
作者: Rui Zou,Mengqi Wei,Xin Zhao,Ji-Rong Wen
类目: Artificial Intelligence (cs.AI)
备注: 28 pages, 7 figures

点击查看摘要

Abstract:Large language models have made substantial progress in mathematical reasoning, yet in complex multi-step reasoning, an error at any step may propagate along an incorrect path, and the resulting solution may omit necessary branches. Existing natural-language reflection methods often struggle to locate critical errors precisely, while outcome verification typically checks only whether the available candidate answers satisfy the problem conditions and can therefore overlook necessary branches. To address these limitations, we propose Agent-Guided Process Verification (AGPV) for mathematical self-correction. AGPV decomposes the current solution into an ordered sequence of key claims, verifies them individually while invoking a shared mathematical toolbox when needed, and stops upon detecting the first error. A summarization agent converts the failure evidence into a diagnosis that guides a correction agent to re-derive the subsequent solution from that point; AGPV then adapts the next verification iteration to the revised solution. AGPV also verifies solution completeness. We further prove that AGPV guarantees solution correctness and completeness when key claims provide sufficient coverage and local verification is reliable, provided that all key claims pass verification. The implementation compresses long working contexts and reuses successful tool results within each problem. We evaluate AGPV on five challenging mathematical reasoning datasets with four models from DeepSeek, Qwen, and GPT. AGPV achieves an average accuracy of 63.8%, outperforming the strongest among the representative baselines evaluated (50.3%) by 13.5 percentage points, with higher repair coverage than the programmatic-verification baselines.

[AI-197] A Hybrid State Space-Attention Architecture for Forecasting the Trajectories of Battery Degradation

链接: https://arxiv.org/abs/2611.19116
作者: Shrey Patel,Bhargav Srinivas Rao,Samet Ayhan
类目: Artificial Intelligence (cs.AI); Chemical Physics (physics.chem-ph)
备注:

点击查看摘要

Abstract:Early-life battery degradation trajectory forecasting aims to predict a battery cell’s full state-of-health (SOH) trajectory from its initial charge-discharge cycles, potentially reducing cell qualification and evaluation time from months to days. Despite recent advances, accurate forecasting remains challenging for emerging battery chemistries, data-scarce domains, and cells exhibiting degradation behaviors beyond those observed during training. Existing state-of-the-art approaches increasingly rely on model scaling, large language model-derived aging-condition embeddings, and extensive fold-specific hyperparameter tuning, raising concerns about computational efficiency, reproducibility, and generalization. This work introduces DecayFormer, a compact hybrid state-space and attention architecture designed for early-life degradation trajectory forecasting. DecayFormer combines cycle-level selective recurrence, using a single scalar state per channel, with an attention mechanism queried by the recurrent state. It further incorporates state-of-charge (SOC) interval statistics and a master curve head that jointly estimates a scalar lifetime and a monotonically non-increasing SOH profile parameterized by normalized lifetime. We evaluate DecayFormer on the BatteryLife benchmark using all 22 aging-condition-exclusive folds, with a single run per fold and one fixed configuration across all folds, without fold-specific tuning. DecayFormer achieves mean test MAPE values of 2.25% on Li-ion, 2.76% on CALB, 0.95% on Na-ion, and 7.19% on Zn-ion. These results demonstrate the potential of compact sequence models to deliver accurate early-life degradation forecasts across diverse battery chemistries and aging conditions, without relying on extensive model scaling or per-fold optimization.

[AI-198] Learning inadequacy of Bayesian neural operators for PDE-constrained inference and design

链接: https://arxiv.org/abs/2611.19115
作者: Pratyush Kumar Singh,Lianghao Cao,E. Bruce Pitman,Danial Faghihi
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Partial differential equations (PDEs) underpin predictive modeling across science, engineering and medicine. Their use in digital twins, parameter inference and optimal design requires repeated solutions across parameters and scenarios. For nonlinear, transient and multiphysics systems, these calculations can make numerical simulation prohibitively expensive. Here we introduce the residual-informed Bayesian neural operator (RI-BNO), which learns operator inadequacy using a posteriori error estimates from the governing PDEs and incorporates its correction and uncertainty into Bayesian training and prediction. We establish local error bounds for the residual-informed estimates under stability and smoothness assumptions. We demonstrate RI-BNO on a non-linear reaction-diffusion benchmark with DeepONet and latent-subspace architectures, Markov chain Monte Carlo inference of glioma parameter fields from magnetic resonance imaging, and Monte Carlo risk-averse optimization of three-dimensional thermal breaks in building envelopes under manufacturing uncertainty. Across the evaluated examples, RI-BNO improves agreement with finite element (FE) reference predictions, parameter posteriors and optimized designs while using half as many full nonlinear or transient training simulations as the vanilla Bayesian neural operator (vBNO). For thermal break design, online optimization time decreases from 178 hours with the FE model to 18 minutes with RI-BNO. These results demonstrate the value of explicit inadequacy modeling in neural operators for reliable PDE-based inference and design.

[AI-199] ST-GAT: A Leakage-Free Spatio-Temporal Graph Attention Framework for Robust Network Intrusion Detection

链接: https://arxiv.org/abs/2611.19109
作者: Nikhil Parkar,Sanskar Pandey,Siddhi Prajapati,Suganeshwari G,Syed Ibrahim
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Intrusion detection systems are a central defence in modern networks. Deep learning models report strong anomaly-detection results, but evaluations on standard benchmarks such as CSE-CIC-IDS-2018 are often inflated by two methodological problems: temporal leakage from overlapping sliding windows, and contamination from preprocessing fitted before the train/test split. Per-flow classifiers also ignore the spatio-temporal dependencies across sequences of network flows. We propose a Spatio-Temporal Graph Attention Network (ST-GAT) that represents traffic as a sequence of graph snapshots. It builds k-Nearest Neighbour (k-NN) edges within each window, applies multi-head Graph Attention Network (GAT) layers for spatial aggregation, and uses a Bidirectional Long Short-Term Memory (BiLSTM) network with temporal attention for sequence modelling. To prevent leakage, we use non-overlapping temporal windows (stride equal to window length) and fit all preprocessing on the training split only. We compare three class-imbalance strategies: inverse class-frequency weighting, dynamic balanced sampling, and few-shot metric prototyping. Across three random seeds, ST-GAT with inverse class weighting reached a Macro-F1 of 96.65% (+/- 0.87) and an accuracy of 99.14%. This is 14.3 percentage points higher in Macro-F1 than a per-flow Random Forest baseline (82.34%). Balanced sampling reduced Macro-F1 by 2.3 points (to 94.37%) relative to loss weighting. This suggests that aggressive resampling of small, noisy minority classes can hurt detection in heavily imbalanced traffic.

[AI-200] Guiding AI Agents to Generate Conjectures by Encoding How Theorems Build on One Another

链接: https://arxiv.org/abs/2611.19108
作者: Hita Kambhamettu,Daniel Krashen,Lyle Ungar
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Recently, LLMs have been making headlines, and headway, in mathematical research. Much of this work focuses on proving existing open conjectures. Deciding what to prove is harder and less studied, because of the lack of a clear criterion for what makes a conjecture worth pursuing. We present a method for conjecture generation that builds on the observation that mathematical theorems build on one another: new theorems can modify the setting or hypotheses of earlier theorems, strengthen their conclusions, or extend their scope. We encode these relationships in a new kind of graph, the Theorem Evolution Graph (TEG), which we instantiate as an online search method for LLM agents. An agent builds a TEG in two steps: (1) identifying relevant theorems from papers found by following citations backward and forward from a seed paper; and (2) comparing those theorems to record how one builds on another. LLM judges and an expert mathematician preferred TEG-generated conjectures over those from an unconstrained baseline. LLM judges favored TEG conjectures for significance, ambition, and insight, and favored baseline conjectures for plausibility. Ablations suggest a benefit from constructing and using a TEG even when the available papers are fixed.

[AI-201] Can Agent ic LLM s Write Task-Based Parallel Code? A Study of OpenMP Tasking HPX StarPU and TTG

链接: https://arxiv.org/abs/2611.19105
作者: Linus Bantel,Anna-Lena Roth,Tim Thüring,Jonas Posner,Dirk Pflüger
类目: Distributed, Parallel, and Cluster Computing (cs.DC); Artificial Intelligence (cs.AI)
备注: Will be published at WAMTA2027

点击查看摘要

Abstract:Task-based programming offers more expressive power than loop-level parallelism, but the runtimes providing it differ widely in API style, maturity, and the amount of public code to learn from. This paper asks how well LLM agents write C++ for four of them: OpenMP Tasking, HPX, StarPU, and TTG. Instead of single one-shot prompts, we use an agentic workflow in which the model edits files, compiles, and runs its kernels until it reproduces a reference output. Five locally hosted open-weight LLM implement seven HPC kernels, including matrix multiplication, Cholesky and LU factorization, in all four frameworks and a serial baseline to verify their understanding of the problem in general. Every kernel is evaluated independently for correctness at each concurrency level, strong scaling, and peak memory. Model capability dominates the outcome: three models produce correct, genuinely task-parallel kernels in 57-82% of the cases, while two models rarely compile their own code and almost never succeed. For the capable models, OpenMP Tasking and HPX are handled most reliably, StarPU is close behind, and TTG trails. The best generated GEMM and factorization kernels match or exceed the multithreaded OpenBLAS/LAPACK reference. However, correctness overstates parallel quality: several “parallel” kernels contain no parallelism at all; and serialized task graphs and unsuitable task granularity remain common.

[AI-202] CONTEXT-PCa: Evaluating the role of clinical context in MRI-based AI for detecting clinically significant prostate cancer (study protocol)

链接: https://arxiv.org/abs/2611.19082
作者: Shiqi Huang,Weixi Yi,Shaheer Ullah Saeed,Yipei Wang,Wen Yan,Natasha Thorley,Shonit Punwani,Aqua Asif,Alexander Ng,Veeru Kasivisvanathan,Mark Emberton,Yipeng Hu
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:The objective of this study is to investigate the role of clinical variables - their breadth, source, and quality - in the development, utility, and generalisability of an AI system for bi-parametric MRI (bpMRI)-based diagnosis of clinically significant prostate cancer (csPCa). Clinical context is added to the published PI-CAI-2B image-only model, whose weights are fixed. The study cohort comprises 6,022 biparametric MRI cases from 4,540 patients at University College London Hospitals NHS Foundation Trust (2019-2024), of which 4,834 cases (3,632 patients) are held out for internal testing. Forty-three clinical variables are drawn from structured health records, clinical letters, MRI reports, and automated image measurements. The primary endpoint is the difference in the area under the receiver operating characteristic curve (AUROC) between the image-context and image-only models on the internal test set. Secondary analyses compare combining strategies, the contribution of individual variables, and performance across subgroups. All models are further tested on five external cohorts from primary diagnosis and population-based screening.

[AI-203] Assume/Guarantee Contracts for the Compositional Multi-Agent Reinforcement Learning of Heterogeneous Teams

链接: https://arxiv.org/abs/2611.19078
作者: Dominic Klukas,Cyrus Neary
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Multi-agent reinforcement learning (MARL) problems for heterogeneous teams require agents to collaborate to complete team tasks while making the best use of diverse agent capabilities. We introduce assume/guarantee contracts for heterogeneous multi-agent systems (AGC-MAS) and Iterative Heterogeneous Assignment for MARL (IHA-MARL). AGC-MAS contracts are compositional objects that encode the assumptions provided to, and the guarantees required of, each subtask contributing to the overall team task. We formulate contracts as tuples that specify a subtask, an automaton, timeout and success-probability requirements, and dispatch behavior. The automaton tracks an agent’s high-level progress through the subtask associated with its contract and provides a natural mechanism for synchronization across contracts. Our abstraction enables robust interaction across heterogeneous low-level subtask implementations, allows heterogeneous agents to substitute for one another in satisfying contracts, and supports reuse of previously satisfied contracts in new team contexts. Having defined our theoretical framework, we develop a decomposition algorithm that breaks down a team task into a set of contracts. We then propose an iterative algorithm for task assignment and MARL that alternates between assigning contracts representing subtasks to individual agents, training the assigned agents to satisfy their contracts, and assessing their resulting capabilities to inform the next round of task assignments until the team’s success probability meets the desired threshold. Our experimental results demonstrate the versatility of contracts in enabling agent substitution and contract reuse. Across three benchmarks with five seeds each, IHA-MARL achieves observed 100% final evaluation success. In CPU timing pilots, comparable baselines required approximately 7 – 11\times more training-loop time.

[AI-204] Policy Optimization With Multiple Verifiers via Pareto Rankings

链接: https://arxiv.org/abs/2611.19064
作者: Kai Yamashita,Shohei Taniguchi,Hiroki Furuta,Masahiro Suzuki,Yutaka Matsuo
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Reinforcement learning with verifiable rewards (RLVR) has become an important approach to improving the reasoning capabilities of large language models. In tasks such as code generation and instruction following, multiple verifiers evaluate whether a response satisfies different requirements. Their outcomes provide detailed feedback about partial success, yet this feedback is commonly aggregated into a scalar reward for policy optimization. Rewarding only complete success discards distinctions among partial solutions, while counts and weighted sums impose comparisons between responses that satisfy different subsets of requirements. We introduce Partial-Order Likelihood Policy Optimization (POLPO), which learns from the comparisons established by verification without first constructing an aggregate reward. Pareto dominance identifies responses that satisfy additional requirements without losing any already satisfied, while leaving responses with conflicting outcomes incomparable. POLPO extends the listwise likelihood underlying Direct Preference Optimization by maximizing the total probability of all complete rankings consistent with these partial orders. This objective provides supervision from verified improvements even before complete solutions are available, without requiring a prescribed ordering of incomparable responses. Experiments on code generation and instruction following demonstrate improved learning efficiency and higher rates of complete requirement satisfaction compared with scalar-reward training. These results support partial-order likelihood as an effective approach to learning from multiple verifiers.

[AI-205] More Context Mixed Results: Evaluating LLM s for Identifier Recovery and Code Deobfuscation

链接: https://arxiv.org/abs/2611.19043
作者: Simon Queyrut,Valerio Schiavoni,Giuseppe Antonio Di Luna,Daniele Cono D’Elia,Leonardo Querzoni,Pascal Felber
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 57 pages, 16 figures. Under review

点击查看摘要

Abstract:Source code is often obfuscated through identifier renaming, incentivizing reverse engineers to develop techniques for recovering meaningful names. To assess how LLMs compare with these techniques, we evaluate 4 traditional baselines and 3 LLMs (Qwen, OLMo and DeepSeek) on identifier detection and semantic recovery across 51 recent Java projects, 3 obfuscators, and 6 context/retrieval settings. LLMs generally outperform traditional techniques on both tasks, but benefit from different forms of context and settings: agentic retrieval yields the strongest detection results, reaching 91.56% F_1 , while enclosing-class context generally improves semantic recovery at a higher token cost. Despite these gains, exact recovery peaks at 16.42% overall (under one-hop context) and 32.93% for field names (under repository retrieval), which are the easiest to recover. Class identifiers are the hardest to detect, with a mean detection F_1 of 48.48% across models, settings and obfuscators. More context is not always beneficial: larger methods hinder semantic recovery. Beyond identifier renaming, our experiments show that control-flow obfuscation degrades both tasks more than string encryption. Our findings provide useful ad hoc insights for reverse engineers choosing models and context strategies to deobfuscate code snippets.

[AI-206] Understanding the Role of Reasoning Data in Pretraining

链接: https://arxiv.org/abs/2611.19042
作者: Jianhao Huang,Allan Zhang,Ali Payani,Baharan Mirzasoleiman
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Reasoning data expose models to multi-step reasoning, but reasoning traces may cover only a subset of the available facts. How should pretraining balance atomic facts with reasoning traces? We study this question by constructing a setting in which every fact provides an atomic training signal, while only a subset appears in reasoning supervision, enabling us to measure generalization to transitions observed atomically but never used in training traces. For a linear transformer, we derive the weight assigned to atomic training signals that minimizes held-transition squared error and show that its leading-order optimum increases with reasoning-chain length and decreases with the fraction of facts held out from reasoning traces. These trends persist in a full transformer layer with an MLP, softmax attention, and cross-entropy loss. We further validate the role of reasoning complexity and data coverage in two natural-language settings: controlled multi-hop reasoning and mathematical pretraining. Together, our results provide a principled view of how reasoning-chain structure and data coverage should shape the allocation of training data during pretraining.

[AI-207] he Critic Can Hurt: On the Safety of Critics in LLM -Guided AutoML Search KDD2027

链接: https://arxiv.org/abs/2611.19037
作者: Gaoyuan Du,Xiaoyang Liu,Rex Zhou,Dengdu Jiang,Rong Fu,Meng Ding,Defu Cao
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Accepted at KDD 2027 Research track

点击查看摘要

Abstract:LLM critics increasingly steer iterative search in agentic AutoML, prompt optimization, and self-refinement–evaluated for whether they help, but rarely for whether they can hurt. We isolate the critic’s marginal effect through a four-arm audit and characterize the loop’s tolerance through controlled fault injection. We show a naive 8B critic drives search below the no-critic random-search floor on 7 of 8 tabular datasets, losing up to 0.245 AUC in the worst case; the magnitude varies strongly with dataset, frontier, and budget. When: stronger critics attenuate the loss on the evaluated suite, without establishing a universal capability threshold. Why: controlled fault injection shows that non-directional errors are nearly costless under best-so-far search, whereas directional errors consume finite search opportunities; an idealized effective-sample calculation yields a necessary condition, q(m-1) \beta , not a deployment certificate. How: a simple observable validation gate and a learned detector can reduce realized loss, but neither is guaranteed to preserve the full-budget no-critic baseline. We will release CRITFAULT, to our knowledge the first benchmark for detecting harmful critic feedback in AutoML. A critic is not a free addition: audit before trusting, and gate when capability or verification is insufficient.

[AI-208] Refinement vs. Replanning : Inference Compute Allocation in Robot Policies NEURIPS2026

链接: https://arxiv.org/abs/2611.19034
作者: Alba Maria Tellez Fernandez
类目: Robotics (cs.RO); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: NeurIPS 2026 Workshop: Post-Training Adaptation of Robot Foundation Models

点击查看摘要

Abstract:A frozen chunked robot policy can spend deployment-time inference compute in two coupled ways: more work within each policy call to refine a predicted action chunk, or more frequent policy calls to replan from fresh observations. We study this allocation without updating model weights. Our evaluation procedure reports success with both nominal compute rate and complete attempt cost, selects comparisons on discovery tasks, and tests prespecified comparisons on held-out tasks. Comparisons added after inspecting the held-out results only summarize observed patterns. We demonstrate the procedure with one frozen SmolVLA checkpoint in synchronous LIBERO simulation, using denoising count as the refinement control and fixing chunk length at K=50. Arithmetic and latency characterize cost differently: at d=10, 82.1% of counted arithmetic is independent of denoising, whereas a descriptive affine fit to fixed-input host time assigns 17.7% to its intercept. At h=2, d=2 succeeds on 139/160 attempts versus 122/160 for d=5 (+10.6 percentage points; task-conditional 95% interval: +4.4 to +16.9), while using 5.9% less arithmetic per control step. Executing h=5 rather than h=2 at d=2 yields 138/160 versus 139/160 successes and 60.7% lower arithmetic per attempt; equivalence is not established. The case study shows why deployment adaptation must evaluate generation effort and feedback frequency jointly; it does not define a universal setting.

[AI-209] HeteroWear: Sensor-Aware Self-Supervised Learning for Heterogeneous Wearable Activity Recognition

链接: https://arxiv.org/abs/2611.19027
作者: Yasaman Mohammadpour,Pouya Shaeri
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Signal Processing (eess.SP); Machine Learning (stat.ML)
备注: Extended version of a paper accepted at IEEE BigData 2026. Code: this https URL

点击查看摘要

Abstract:Self-supervised learning for wearable sensing is usually confined to a single dataset or device family, leaving open whether representations transfer across heterogeneous sensor configurations and whether gains come from pretraining rather than architecture or richer sensor input. We present HeteroWear, a sensor-aware self-supervised framework pretrained on an eight-dataset wearable corpus (173 subjects, 358,261 windows) spanning inertial and physiological signals across body locations, devices, and sampling rates. HeteroWear separates within-sensor temporal modeling from cross-sensor relational modeling using sensor-context embeddings, temporal and cross-sensor Transformers, and latent cross-attention fusion. Training combines structured sensor dropout, cross-configuration consistency, and per-axis masked sensor modeling; missing sensors are omitted rather than reconstructed. Under leave-one-dataset-out pretraining and subject-disjoint evaluation on PAMAP2, WISDM, HHAR, HARTH, and UCI-HAR, HeteroWear improves mean macro-F1 from 0.688 to 0.724 over the identical architecture trained from scratch, with the largest gain on HHAR (0.823 versus 0.741), and achieves the highest macro-F1 on four of five benchmarks relative to single-accelerometer supervised baselines. Held-out-device evaluation yields a mean +0.066 macro-F1 improvement. Under 50% test-time channel removal, HeteroWear retains 73.8% of full-channel performance versus 64.1% for the same architecture trained from scratch without sensor dropout. Ablations show that structured sensor dropout primarily drives robustness, while cross-configuration consistency and sensor-context encoding contribute to transfer.

[AI-210] Match and Correct: A Repeated Computational Motif in Evoformer Layers ATC NEURIPS2026

链接: https://arxiv.org/abs/2611.19016
作者: Ishan Khire,Naren Manikandan,Tyler Hayes,Giri Krishnan
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Quantitative Methods (q-bio.QM)
备注: 27 pages, 4 figures (main text), 16 supplementary figures. NeurIPS 2026 Workshop: Interpretability for Discovery. Code: this https URL

点击查看摘要

Abstract:Triangle attention and triangle multiplicative updates form key components of the Evoformer, and were introduced specifically to encourage consistency with the triangle inequality. Yet whether representational changes over layers reduce triangle inequality violations has not been tested. We probe the pair representation at each of OpenFold’s 48 Evoformer layers, decoding final C _\alpha -C _\alpha distances with four probe families (a shared linear probe, per-protein linear probes, a sequence-separation-gated linear probe, and XGBoost) and reconstructing intermediate 3D structures from the decoded distances via multidimensional scaling. Across a cohort of 552 CAMEO and ECOD proteins, distance decodability ( R^2 ) and reconstruction quality (TM-score) rise in early layers, plateau in the middle of the network, and rise again in late layers. Directly measuring triangle-inequality violations in the decoded distances reveals a similar, more exaggerated pattern: violations rise to a peak, fall to a trough, rise to a second peak, and only then decline, revealing two distinct episodes of decreasing normalized triangle violations rather than one continuous improvement. Cumulative forward and reverse layer-wise ablations of the Pair Transition, MSA Transition, and triangle mechanisms localize the regions in which specific Evoformer components become essential. The two stages of geometric refinement coincide with the use of the Pair Transition MLP and the triangle mechanisms. These results characterize how distance information and decoded geometric consistency evolve across layers, and motivate hypotheses about the computations behind their trajectories.

[AI-211] Agent -ZTA: A chain-Aware Zero Trust Model for Bidirectional Security-State Propagation and Runtime Containment in Multi-Agent Delegation

链接: https://arxiv.org/abs/2611.18999
作者: Asadullah Abdullah Khan,Toqeer Ali Syed
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: Conference version

点击查看摘要

Abstract:Multi-agent systems increasingly delegate tasks and permissions across chains of autonomous agents that invoke protected tools. Although agent identities and delegated credentials may remain valid throughout a workflow, the security condition of any agent along the chain can deteriorate during runtime execution-either through upstream compromise or downstream ingestion of untrusted inputs. Authorization models based solely on the immediate caller and its static delegated scope fail to account for these mid-execution state changes. This paper develops Agent-ZTA, a chain-aware Zero Trust authorization and containment model that links runtime security-state changes across multi-agent delegation graphs. The model distinguishes an agent’s path-specific delegated authority bound from the effective permissions currently exercisable under trusted policy. It defines an architecture featuring independently enforced policy decisions, authenticated delegation provenance, a dynamic chain-state view, and explicit fail-safe handling of missing or stale authorization signals. Furthermore, Agent-ZTA defines bidirectional security state propagation, allowing downstream security events to affect relevant upstream workflow states, and specifies adaptive branch-scoped containment through restricted permissions and a sandbox (“jail”) execution mode where an isolation boundary is available. The deterministic decision procedure, five stated security properties, and worked authorization traces show how accepted security-state changes constrain affected operations while preserving original delegation bounds and separating unrelated workflow branches under the defined policies. Finally, the paper states its trust assumptions and the exact scope of its security properties.

[AI-212] Model Maps: A Documentation Framework for Tracing Foundation Model Ancestry and Understanding AI Entanglement

链接: https://arxiv.org/abs/2611.18998
作者: Rachel Hong,Madeleine Grunde-McLaughlin,Sarah Cen,Jevan Hutson,Jamie Morgenstern,Tadayoshi Kohno
类目: Computers and Society (cs.CY); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:The development of foundation models incorporates pre-existing AI artifacts like models, datasets, and benchmarks. These relationships, as part of the broader AI supply chain, produce “entangled” webs in which properties of artifacts upstream can influence the outputs of a foundation model downstream - a phenomenon known as AI entanglement. Without visibility into these connections, a black-box model can inherit upstream properties that result in security vulnerabilities, hallucinations, or biased predictions, yet developers are left with no recourse to identify or mitigate the root issue. To address these concerns, we introduce model maps as a conceptual documentation framework for tracing model ancestry that builds on previous documentation frameworks like ecosystem graphs, datasheets, and model cards. A model map for a given foundation model is a detailed knowledge graph that represents the relationships between artifacts that influence the resulting foundation model. Our framework incorporates a taxonomy of 13 dependency mechanisms including model fine-tuning and synthetic data generation. We apply this framework to construct model maps of 15 closed and open-weight foundation models to illustrate the complexity of AI entanglement in current model development practices. From our model maps, we record 2,778 unique dependencies and trace artifacts that are located up to 9 hops away from a resulting foundation model. Our findings suggest that AI transparency requirements should extend beyond individual models and training-data disclosures to encompass upstream dependencies, which would enable more effective risk assessment, auditing, and accountability across the AI supply chain. To help developers, regulators, and researchers understand and analyze the interactions that can influence the behavior of a foundation model, we release these these model maps at this https URL.

[AI-213] Context Matters: Improving LLM -Based Code Summarization with Hierarchical Information

链接: https://arxiv.org/abs/2611.18964
作者: Alexandre Petrachini,Breno Bernard Nicolau de França,Guido Araujo
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 11 pages, 4 figures, 10 tables. Accepted and presented at the Brazilian Symposium on Software Quality (SBQS 2026), Technical Track. Replication package: this https URL

点击查看摘要

Abstract:Source code documentation plays an important role in software quality activities, such as code review, program comprehension, maintenance, and bug fixing. However, it is often fragmented across individual code elements and focuses on local implementation details, providing little information about their surrounding context. This limits both developers’ understanding of software systems and the effectiveness of Large Language Models (LLMs), whose reasoning is constrained by finite context windows. Enriching documentation with hierarchical context has the potential to improve software comprehension and support downstream, quality-related tasks for both developers and LLM-based tools. In this paper, we present an in silico study that showcases the impact of augmenting context on the generation of code-level documentation, using a variety of LLMs. By varying the function, class, and module hierarchy visible to the LLMs, we assess whether this variation produces a significant difference in the resulting code documentation compared with the existing documentation, which serves as a proxy for developer intent. Our results show that Documentation-Aware prompts with class or module-level context consistently yield summaries that are closest to the original documentation. Adding surrounding documentation yields statistically significant improvements over the baseline across models, with effect sizes ranging from negligible (DeepSeek) through weak (GPT-4o-mini) to medium (Gemini). We restrict the study to this immediate hierarchical context; call-graph traversal and cross-module dependency context fall outside our scope and are left to future work. These findings provide a clear path to improve automated documentation tools and enhance LLMs’ ability to reason about codebases for downstream tasks, such as code retrieval.

[AI-214] SpecRelay: Amortize or Avoid the Round Trip in Edge-Cloud Speculative Decoding

链接: https://arxiv.org/abs/2611.18963
作者: Jai Jain,Chihyo Ahn,Yong Liu,Charles Chen,Duc N.M Hoang,Ajay Kumar Jaiswal,Fatih Ilhan,Han-Byul Kim,Lauren Hannah,Richa Dixit,Minsik Cho,Yingyan Celine Lin,Hyesoon Kim
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Edge-cloud speculative decoding moves autoregressive drafting to an edge device while a larger cloud model verifies draft tokens in parallel, so that edge drafting and cloud verification overlap under a synchronous draft-verify schedule. Beyond low-latency local networks, however, the next verification request of that schedule can be sent only after the previous reply arrives, so every round still waits out a full round trip, and throughput falls as the network RTT grows. To recover this loss, we present SpecRelay, an edge-cloud LLM inference system that addresses this cost at two levels: it amortizes round trips when cloud verification is necessary and avoids them when local reasoning suffices. For cloud-bound queries, an asynchronous verification pipeline keeps fixed-budget draft trees in flight while earlier replies are still outstanding, and the cloud rejects stale requests before the target-model forward pass (the asynchronous schedule). At the query level, an RTT-conditioned router (the Latency-Aware Router) estimates quality and completion latency for three on-device modes (direct generation, Chain-of-Draft, and full reasoning) and cloud speculative decoding, selecting among them under a configurable accuracy-latency objective without adding latency during inference, since the router runs concurrently with the draft-model prefill. The asynchronous schedule alone improves decoding throughput by up to 50% over a reproduced synchronous baseline across multiple benchmarks. At matched target-model accuracy and 250 ms of added round-trip latency, the full system with the router delivers 1.95x and 1.55x the throughput of the same baseline on SPROUT and OlympiadBench, respectively, while sending only 47% and 39% of the queries to the cloud.

[AI-215] Dual-Stage Channel-Posterior Integration for MLM-Aided Token Communications

链接: https://arxiv.org/abs/2611.18958
作者: Hyelin Nam,Jihoon Lee,Seung-Woo Ko,Jinho Choi,Seong-Lyun Kim
类目: Artificial Intelligence (cs.AI); Information Theory (cs.IT); Signal Processing (eess.SP)
备注: 6 pages, 5 figures, This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:Token transmission over conventional digital modulation is vulnerable to channel impairments, leading to bit errors that cause received tokens to become contextually inconsistent. This paper proposes a receiver-side dual-stage channel-posterior integration framework for masked language model (MLM)-aided token reconstruction. In the proposed framework, the receiver detects unreliable token positions from channel reliability and constructs a vocabulary-level channel-posterior. In particular, within the receiver, Posterior-Weighted Input Embedding (PWE) injects this posterior into the MLM input, while Adaptive Output Logit Fusion (AOF) reuses it at the output with a channel-adaptive weight. Consequently, multiple unreliable tokens can be reconstructed in a single bidirectional MLM inference without retransmission, transmitter-side encoding, or additional MLM training. Experiments over noisy channels show up to a 67.2 % improvement in cosine similarity over context-only MLM reconstruction, with consistent gains across different MLM architectures and existing language-model-aided methods.

[AI-216] From Language-Specific to Cross-DSL: Semantic-Role-Guided Model Completion with LLM s

链接: https://arxiv.org/abs/2611.18952
作者: Meriem Ben Chaaben,Théo Giraudet,Benoit Combemale,Houari Sahraoui
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:The rapid emergence of large language models (LLMs), trained on vast and diverse corpora, has positioned them as a promising source of domain knowledge to assist modelers. However, leveraging LLMs for model completion remains challenging: suggestions must respect both the syntax of a domain-specific language (DSL) and the intended domain semantics, while remaining consistent with an evolving model. This paper presents RAFiC, an LLM-based approach for iterative domain model completion that generalizes across DSLs. It drives completion using metamodel information to ensure syntactic validity and enriches prompts with semantic-role annotations to better capture the intended modeling role of DSL elements. The approach provides context-aware, incremental guidance during model construction. We evaluate RAFiC through an ablation-based comparative study over three DSLs and two scenarios per DSL to isolate the contribution of syntactic and semantic guidance. Across these six scenarios, semantic role guidance improves contribution rate in every case, and RAFiC improves completion effectiveness over weaker prompting configurations in most of them. The largest gains occur for DSLs that are less represented in LLM training data. In these cases, iterative guidance leads to earlier and more stable accepted suggestions, while semantic role guidance improves their relevance.

[AI-217] he Human Side of ML Engineering: Analyzing the Collaboration and Communication Skills in ML Job Postings

链接: https://arxiv.org/abs/2611.18915
作者: Aidin Azamnouri,Phuong Thao Le,Justus Bogner,Stefan Wagner
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
备注:

点击查看摘要

Abstract:Studies of Machine Learning (ML) practice highlight the importance of combining technical expertise with interpersonal capabilities, particularly collaboration and communication. However, evidence on how ML job postings express these skills is scarce, and prior work largely relies on keyword-based analyses that overlook implicit expression. We analyze demand for these skills in ML job postings across regions, industries, and seniority levels, and compare it with software engineering (SE) roles. We collected 1,000 postings (500 ML, 500 SE) from five continents via a semi-automated pipeline. For annotation, we synthesized a taxonomy of collaboration and communication sub-categories from prior work and used an LLM as a third annotator alongside two humans, enabling detection of context-dependent and implicit skill references. Collaboration appeared in 88.6% of ML postings and communication in 72.2%, far above prior keyword-based estimates for AI roles. Roughly one in five collaboration and one in four communication requirements were expressed only implicitly. Decomposing these requirements into the taxonomy’s sub-categories reveals what aggregate figures conceal. Collaboration is demanded predominantly as integration across role boundaries rather than team membership. Communication is mostly required in generic terms, with fewer than half of the postings requiring it naming whom the candidate must communicate with. ML and SE postings demand interpersonal competence at statistically indistinguishable levels but orient it differently, SE toward team practices and ML toward presentation and audience-facing interaction. Within ML, the advertised role separates interpersonal profiles more sharply than any other dimension, including the contrast with SE.

[AI-218] Ask Precisely Assume Wisely: Interactive Evaluation of Ambiguity Resolution in Large Language Models NEURIPS2026

链接: https://arxiv.org/abs/2611.18905
作者: Avik Dutta,Aayush Kumar,Zahra Ashktorab,Hosein Hasanbeig,Austin Z. Henley,Gustavo Soares
类目: Artificial Intelligence (cs.AI)
备注: 40th Conference on Neural Information Processing Systems (NeurIPS 2026). Workshop: Evaluation of Interactive Agents

点击查看摘要

Abstract:Ambiguity is a core challenge in human–AI interaction, where models must decide when to make assumptions and when to seek clarification. Existing approaches to ambiguity evaluation largely focus on coverage over multiple interpretations using static metrics, offering limited insight into how efficiently a model converges toward a single user-intended interpretation through interaction. We introduce a preference-grounded simulated user framework for controlled, multi-turn evaluation of ambiguity resolution, grounding simulator responses strictly in user disambiguation preferences to avoid hallucination and over-disclosure. Alongside this, we define three observable dissatisfaction signals — incorrect assumptions, repeating clarifications, and irrelevant clarifications — that characterize interaction quality beyond task accuracy. Our analysis reveals a clear trade-off: models that frequently seek clarification tend to over-clarify and prolong interactions, while assumption-driven models are more prone to misaligning with user intent. We demonstrate that system prompts optimized on these dissatisfaction signals achieve a Pareto-optimal balance across all three metrics, reducing overall error rates by up to 7.65%, underscoring the value of our framework in designing agents that ask precisely and assume wisely.

[AI-219] Benchmark Factory: Agent ic Benchmark Creation via Dataset Curation and Label-Free Quality Anchors

链接: https://arxiv.org/abs/2611.18890
作者: Shervin Shahidi,James Poon,Hang Xie,Kenneth O. Xu,Robin Håkanson,Vikram Nair,Jens Pillgram-Larsen,Nachiappan Nagappan
类目: oftware Engineering (cs.SE); Artificial Intelligence (cs.AI)
备注: 5 pages, prepared for submission to AGENT 2027

点击查看摘要

Abstract:Teams deploying coding agents need benchmarks built from their own code, but curating and validating them by hand is resource intensive, and an unvalidated benchmark can mislead silently. We present Benchmark Factory, an autonomous meta-agent loop with human checkpoints: an LLM agent writes sourcing queries over landed production code changes, runs models of known relative strength on each candidate, diagnoses why a candidate fails, and repairs it by editing its queries, prompts, and configuration until it clears quality anchors that need no per-task human labels. We used Benchmark Factory to build benchmarks for five developer cohorts at our organization. Engineering raters scored a 5% sample 2.3 for task quality and 2.6 for grader verdicts on a 1-to-3 scale. On configurations never used to build them, the benchmarks ranked Claude Opus~4.7 at high reasoning effort above low effort (60.6% vs. 54.4% of tasks passed, p=0.002 ) and GPT-5.5 above GPT-5.4 (65.8% vs. 60.5%, p=0.004 ); 9 of the 10 per-cohort comparisons pointed the same way. A deliberately trivial benchmark failed the anchors; five broken graders failed them, even under a relaxed rule fixed in advance that the real grader passes; and with model labels shuffled so that no true ordering existed, random redraws of a candidate passed in at most 1.4% of simulations. The anchors are necessary, not sufficient, conditions for a valid benchmark, designed to focus human review on what the loop cannot repair.

[AI-220] We Have All the Cards: Securing Agent Cards at the Operational Trust Boundary of the Agent Agent 2Agent Protocol ACSA

链接: https://arxiv.org/abs/2611.18887
作者: Miguel A. Guirao Aguilera,Satyajayant Misra,Efren Lopez-Morales
类目: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI)
备注: To be published at the Workshop on Agentic AI in Offensive and Defensive Cyber Operations (AIDC’26) proceedings co-located with the IEEE Annual Computer Security Applications Conference (ACSAC)

点击查看摘要

Abstract:AI agent interoperability protocols enable orchestrators to discover and invoke external agents using machine-readable metadata. In the Agent2Agent (A2A) protocol, Agent Cards contain attacker-controllable natural-language fields that may influence agent selection before a provider is invoked, creating a pre-invocation trust problem. We characterize this discovery-time attack surface through a taxonomy of ten Agent Card manipulation attacks and present SECCARD, a layered validation framework combining protocol checks, deterministic analysis, and LLM-based semantic reasoning. We measure the emerging public A2A ecosystem by probing 36.1 million Internet domains and collecting 310 deployed Agent Cards. Two annotators independently label the corpus with 96.8% agreement. We then construct a 660-card real-card perturbation benchmark with matched benign controls and a 260-card synthetic adversarial benchmark spanning all ten attack categories. On controlled attacks injected into real Agent Cards, the strongest semantic judge achieves 97.0% security recall at a 1.7% false-positive rate. Across nine LLM judges on the synthetic benchmark, security recall ranges from 25.8% to 92.3%, demonstrating substantial model-dependent variation. These results show that SECCARD effectively identifies Agent Card manipulations spanning all ten discovery-time attack classes while preserving benign Agent Cards, but also reveal that its effectiveness depends critically on the choice of the underlying LLM judge.

[AI-221] SPLICE: Structured Prompt Local Iterative Combinatorial Evolution

链接: https://arxiv.org/abs/2611.18883
作者: Anish Acharya,Phillip Studans,Amit Dhanda,Sina Amini Niaki,Ninad Vivek Rao,Vishal Amarlal Khatri,Brian Verkhovsky
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Natural language optimization (NLO) improves an agent by having a model critique and rewrite its natural language policy parameters: system prompts, tool instructions, reasoning directives, few-shot demonstrations. These artifacts are multi-section documents in which some sections encode contracts (safety clauses, tool schemas, output formats) that must not drift, while others are legitimate optimization targets. Most iterative NLO methods treat the artifact as a single editable string, producing three recurrent failure modes: unsafe global rewrites, beam collapse onto a single lineage, and unchecked bloat, none addressed by formal guarantees. We introduce SPLICE, a unified NLO framework that poses optimization as constrained combinatorial search over an edit graph of block-structured, selectively-mutable states. We map each failure mode to a matched guarantee: (i) per-block mutability preserves contract text verbatim, with a price-of-freezing bound; (ii) elitist beam search with diversity slots never lets the leader regress on validation, bounds regression under noisy re-evaluation by an exponential tail, and keeps a competitive candidate from a different parent whenever one exists; (iii) a ratio length regularizer comes with an explicit accuracy-length bias bound. Candidates are proposed from dual section-local textual gradients; under a proposer-competence condition whose proxy we measure, edit-graph reachability and a convergence rate bound how quickly improving children arrive. Across seven public benchmarks under five reproduced protocols, SPLICE wins on four and matches the strongest prior method on three. On a production-style contract stress test it keeps the contract clauses that unconstrained optimizers trade away to raise the proxy metric, and ablations attribute gains to critique and elitist search.

[AI-222] Découpage: Directorial Audio-Visual Control for Long Multi-Shot Video Generation

链接: https://arxiv.org/abs/2611.18866
作者: Yifan Wang,Hao Ouyang,Qiuyu Wang,Yihao Meng,Wen Wang,Zeyue Tian,Yichong Lu,Hanlin Wang,Qingyan Bai,Zichen Liu,Yujun Shen,Qifeng Chen
类目: Artificial Intelligence (cs.AI); Sound (cs.SD)
备注: Project page: this https URL

点击查看摘要

Abstract:In cinematic storytelling, sound and visual transitions do not always coincide in lockstep: dialogue frequently bridges across cuts to reveal reactions or introduces new scenes via J- and L-cuts, while deliberate silences pace the scene. Existing audio-visual generative frameworks, however, lack explicit mechanisms to coordinate independent visual and acoustic timing over extended horizons. As narrative duration and shot complexity scale, these systems struggle to maintain long-range dialogue pacing, often hallucinating speech during intended silences or erroneously driving mouth movements on silent listeners. To enable fine-grained cinematic control, we present DECOUPAGE, a screenplay-driven framework for long-form, multi-shot, and multi-character joint audio-video generation. Rather than coupling modality boundaries by default, DECOUPAGE uncouples generation into independently bounded visual and acoustic tracks governed by persistent character identities. DECOUPAGE builds on a joint audio-video diffusion model with dual-schedule temporal cross-attention, incorporating a directorial Lip-sync mechanism that selectively decouples speech conditioning from visual facial dynamics during off-screen dialogue. On multi-shot conversational benchmarks, DECOUPAGE achieves precise control over shot transitions, audio timing, and visibility-aware lip synchronization, directing 15- to 60-second screenplay-driven audio-video generation with multimodal coherence.

[AI-223] Self-Improvement Memory Enables Continual Recursive Self-Improvement

链接: https://arxiv.org/abs/2611.18856
作者: Xin Yu,Fei Xia,Jiaqi Pan,Youbao Tang,Chengzhi Liu,Adams Yu,Sicong Jiang,Yanfei Chen,Lingzhou Xue,Ting Yu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Harness optimization is often studied with a fixed policy model and benchmark. When either changes, the optimizer must decide which parts of an evolved harness to reuse or revise, but the harness alone may not explain why earlier edits were made or how they performed. In this work, we study Continual Recursive Self-Improvement (CRSI), a setting for continuing existing harness-optimization methods across compatible benchmarks or policy models. We propose Self-Improvement Memory (SIM) for the harness optimizer, rather than the task-solving policy. SIM groups edits by recurring harness problem and records their rationale, predicted and observed outcomes, and evaluation context. New attempts extend this history as optimization continues. In our experiments across four domains, SIM achieves higher final scores than harness-only and root-cause-only continuation in cross-dataset and pairwise cross-model comparisons. Across ten independent source runs in coding, paired continuations with SIM improve final scores by an average of 2.36 points over harness-only continuation ( p=0.003 ).

[AI-224] What to Practice Next? Latent Profile-Guided Task Selection for Skill-Evolving LLM Agents

链接: https://arxiv.org/abs/2611.18818
作者: Qijin Zhang,Hanzhen Wang,Yan Huang,Zhongxiang Dai,Shuang Qiu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Task selection shapes the experiences available for skill evolution in LLM agents, yet fixed practice schedules overlook differences in skill requirements and evolving task value. Most existing methods, however, rely on fixed schedules or scalar measures of progress, leaving nonstationary skill demands unaddressed. We propose SPG-Select, a framework for allocating limited practice budget using a multidimensional profile inferred from binary outcomes via MIRT. The framework predicts surrogate progress, defined as the difference between consecutive profile estimates, and prioritizes progress in dimensions with larger estimated proficiency gaps. Our SPG-Select-R policy combines these predictions with uncertainty-based exploration and periodic restarting to accommodate nonstationarity. We evaluate SPG-Select on ALFWorld and WebShop with three skill-evolution backends. In the reported runs, SPG-Select-R achieves higher mean held-out terminal success than uniform sampling in all six backend-benchmark combinations, with gains of 1.49-13.18 percentage points on ALFWorld and 0.93-3.13 points on WebShop. Under a predictable linear surrogate model and boundedness assumptions, we establish a high-probability dynamic pseudo-regret bound for surrogate task allocation in terms of parameter path variation.

[AI-225] LEAP-LLM : Preserving lead-time structure for question-conditioned ECG interpretation

链接: https://arxiv.org/abs/2611.18809
作者: Aakash Agarwal,Yidan Shen,Xin Fu,Renjie Hu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: 23 Pages including Appendix

点击查看摘要

Abstract:The electrocardiogram (ECG) records cardiac electrical activity across multiple leads and is used for detecting cardiac abnormalities. Recent ECG-language models enable large language models (LLMs) to interpret cardiac signals and answer clinically relevant questions. However, existing ECG-language models typically encode the ECG independently of the clinical question, forcing the ECG representation to retain broad and potentially irrelevant signal information. Moreover, these models often aggregate information across leads which can obscure lead-specific details needed for a particular question. To tackle these limitations, we introduce LEAP-LLM, an ECG-LLM architecture for Lead-aware ECG Alignment with Prompt conditioning. LEAP-LLM uses the clinical question to guide ECG feature extraction while preserving explicit lead and temporal structure. Task-trained instruction representations distill information from the question and guide two complementary branches. The Lead branch captures question-relevant temporal patterns within each ECG lead, while the Temporal branch summarizes relevant information across leads at each temporal position. Together, they provide the LLM with compact, question-relevant ECG information while preserving lead and temporal structure. LEAP-LLM reduces modality-token count by 89.0% and prefill FLOPs by 7.2x on average relative to other ECG-LLMs baselines. Across four ECG benchmarks, LEAP-LLM achieves better performance, including a 3.4-percentage-point improvement in G12EC over the strongest prior ECG-specific method. We further show that the model adaptively emphasizes the clinically relevant anatomical lead groups for different cardiac abnormalities, with lead importance patterns consistent with established diagnostic knowledge, providing interpretable evidence that the predictions are supported by clinically meaningful ECG structure.

[AI-226] Multimodal Fusion: A Review of Breast Cancer Advanced and Traditional Modalities

链接: https://arxiv.org/abs/2611.18793
作者: Jawaher Al-Tuweity,João Paulo Papa
类目: Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Purpose: Multimodal medical imaging fusion (MMIF) is an increasingly important approach to breast cancer detection (BCD), as different imaging modalities rely on different physical principles and provide complementary information. No single modality can capture all the structural, functional, and biological information needed for accurate diagnosis and characterization. Methods: This review examines more than 15 imaging modalities used in breast cancer detection, including conventional techniques such as mammography, ultrasound, and magnetic resonance imaging (MRI), as well as emerging approaches such as thermal imaging, photoacoustic imaging, microwave imaging, and multimodal microscopy. Results: Mammography, ultrasound, and MRI remain among the most widely used modalities, providing complementary structural and functional information. Recent AI advances have enabled fusion at the data, feature, and decision levels, with applications in detection, classification, segmentation, prognosis, and personalized cancer management. Multimodal approaches increasingly combine imaging data with clinical, histopathological, and genomic information. However, important challenges remain, including limited availability of large and diverse multimodal datasets, variations in imaging protocols, missing modalities, limited external validation, and the need for interpretable and clinically reliable AI models. Conclusion: MMIF offers promising opportunities to improve breast cancer detection and characterization by combining complementary imaging and non-imaging information. Future research should prioritize robust fusion strategies, standardized and diverse datasets, multicenter validation, explainable and privacy-preserving AI, and the integration of imaging with clinical and molecular data to facilitate reliable translation into clinical practice.

[AI-227] Gradient-Guided Hardware-Native MIXFP4

链接: https://arxiv.org/abs/2611.18769
作者: Jian-Jia Chen,Brian J. Chan,Wen-Hung Lee,Yi-Ting Li,Po-Shen Chen,Kuan-Te Kuo,Hsien-Cheng Huang,Cody Hao Yu,Chi-Chih Chang,Mohamed S. Abdelfattah,Kai-Chiang Wu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Blackwell tensor cores support native inference with 4-bit weights and activations (W4A4), but NVFP4 leaves a perplexity (PPL) gap to BF16. Choosing a format for each 16-element block can improve accuracy, but current matrix multiply instructions cannot execute choices this fine. Applying these rules at hardware-compatible tiles loses most of their benefit. We introduce FlipQuant, which selects E2M1 or E0M3 for each MMA weight tile. It keeps both quantized weight candidates fixed and learns a binary format map from end-to-end gradients of a bucketed KL loss to a BF16 teacher, re-evaluating them after each update. Native kernels execute 16 x 64 tiles on SM120 and 256 x 64 tiles on SM100 with median GEMM overheads of 1.9% and 1.7%, respectively. Across six language models and two corpora, FlipQuant at 16 x 64 closes 42.5% of NVFP4’s PPL gap to BF16 on average, compared with 13.3% for FourOverSix. Code is available at this https URL.

[AI-228] SART: Sign-Absolute Reformulation Theory for Binary Variable Reduction in Neural Network Verification

链接: https://arxiv.org/abs/2611.18657
作者: Jin Xu,Miaomiao Zhang,Bowen Du
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Neural and Evolutionary Computing (cs.NE)
备注: OOPSLA 2026

点击查看摘要

Abstract:Complete formal verification of neural networks is crucial for their deployment in safety-critical domains. A key bottleneck stems from encoding complexity: traditional methods assign one binary variable per unstable ReLU neuron. We propose the Sign-Absolute Reformulation Theory (SART), which fundamentally breaks the conventional one-to-one mapping between unstable neurons and binary variables by establishing formal reducibility criteria. This allows for finer-grained modeling, where each unstable neuron corresponds on average to fewer than one binary variable, thereby reducing verification complexity at its source. Based on SART, we derive a theoretical lower bound on the number of binary variables required for complete verification and, under the assumption that P \neq NP , prove that variables in the final layer can be compressed by 50%, while the number of variables in intermediate layers cannot be further reduced. To overcome the apparent ``last-layer-only’’ limitation, we recast verification as a sequential process and, crucially, show that the gain lifts to the entire network: LayerABS, a SART-based progressive tightening verifier, iteratively treats intermediate layers as temporary final layers and propagates tight bounds that shrink the global search space and binary-variable counts. Furthermore, we reveal a structural law influencing verification complexity: when the signs of weights of unstable neurons satisfy numerical symmetry, with positive and negative weights equal or differing by at most one, the worst-case verification complexity achieves the theoretical optimum, offering theoretical guidance for the design of verification-friendly architectures. As a general-purpose underlying encoding, the value of SART is independent of specific algorithms.

[AI-229] Algebraic Interaction: A Recipe for Neural Architecture Generation

链接: https://arxiv.org/abs/2611.18649
作者: Haonan Dong,Angelica I Aviles-Rivero
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:We introduce an algebraic foundation for neural architecture design in which convolution, attention, state-space models, and equivariant layers are expressed through compositions of bilinear product operators over two structures: a position algebra governing positional routing and a feature algebra governing feature interactions. This decomposition makes routing, feature algebra, and operator composition explicit design this http URL focus on the feature-algebra axis through an Algebraic Recipe: fix the routing and operator composition of an existing architecture while replacing its feature algebra. This preserves the information-flow structure while allowing computational and symmetry properties to be modified through algebraic this http URL instantiate the recipe in linear attention, Mamba, and equivariant networks. Structure-constant and Clifford algebras improve accuracy at comparable parameter count in linear attention; structure-constant algebras yield up to 3\times faster training and inference in Vision Mamba at competitive accuracy; and an \mathrmSO(3) -representation algebra provides rotational equivariance by construction. These results demonstrate that feature algebra alone can control accuracy, efficiency, and symmetry without redesigning positional this http URL turns algebraic structure into an explicit design variable and reduces architecture design, within our framework, to algebraic specification.

[AI-230] RSIMed: Scaling Medical Foundation Models beyond Limited Clinical Data via Retrieval-Grounded Self-Evolution

链接: https://arxiv.org/abs/2611.18646
作者: Wei Dai,Qianchu Liu,Guanghui Qin,Xing Han,Tristan Naumann,Sheng Zhang,Paul Pu Liang
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Rare diseases are hard to diagnose, and the labeled clinical records needed to train models for them are few, private, and slow to extend, since a specialist must read every case. A model that writes its own training tasks could learn beyond these records, but no program can check a model-written diagnosis, and a wrong answer key corrupts training. To tackle this dilemma between data scarcity and difficulty of data annotations, we introduce RSIMed, Recursive Self-Improvement in Medicine, one of the first self-evolving pipelines in which a task proposer co-evolves in context with a diagnostic solver on real clinical records. Seeded by training admissions, a frozen copy of the solver writes each clinical question from retrieved medical evidence, so its answer rests on a source, and steers topics and difficulty toward what the solver still gets wrong. On rare-disease diagnosis from MIMIC-IV admissions, the accuracy of RSIMed-27B grows linearly with training compute, while a static pool of generated tasks collapses. Without any external model, RSIMed-27B and RSIMed-9B reach 0.516 and 0.488 diagnostic accuracy, above every off-the-shelf GPT-5.6 setting, whose best reaches 0.399 at 38 times the inference cost of RSIMed-27B. Tasks written by four external teachers, including GPT-5.5, train a less accurate diagnostician than the solver’s own. Trained only on diagnosis, RSIMed-RL-27B also improves on the physician-written tasks of HealthBench Professional, from 0.291 to 0.303 in adjusted score.

[AI-231] Minari: Infrastructure for Collaborative Reinforcement Learning Datasets

链接: https://arxiv.org/abs/2611.18644
作者: Omar G. Younis,Rodrigo Perez-Vicente,John U. Balis,Alex Davey,Will Dudley,Jordan K. Terry
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Offline reinforcement learning (RL) depends on shared datasets, but sharing them is still hard. Each benchmark ships its own format, documentation is thin, and most datasets cannot be regenerated once the environment code moves on. We introduce Minari, an open-source Python library for creating, sharing, and loading RL datasets. Its API follows Gymnasium, so recording a dataset means wrapping an environment and stepping through it as usual. From there, users can attach metadata, filter, combine, and shuffle datasets, and publish or download them with one command through the HuggingFace Hub. Every dataset stores the environment specification, seeds, applied wrappers, and pinned package requirements needed to rebuild the environment and regenerate the data. TorchRL, AgileRL, and d3rlpy already load Minari datasets natively, so switching datasets in those libraries takes one line of code. Minari also hosts 338 reproducible datasets (350K trajectories, 55M steps) covering robotic locomotion and manipulation, maze and grid navigation, language-conditioned instruction following, Atari, and web navigation, including a full regeneration of D4RL with public generation scripts. We report baseline scores for six offline RL algorithms on the regenerated D4RL and MuJoCo datasets.

[AI-232] FedFit: Federated Dynamic Sparse Training via Fisher Information scoring ICML2026

链接: https://arxiv.org/abs/2611.18643
作者: Meng Bi,Hong Huang,Jinlong Song,Charles Wang,Chengming Hu,Xi Chen,Ting Yu,Xue Liu
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: The paper was accepted by ICML 2026

点击查看摘要

Abstract:Cross-device Federated Learning (FL) is frequently bottlenecked by the prohibitive memory and communication costs of training deep neural networks on resource-constrained edge hardware. While federated dynamic sparse training aims to alleviate these costs by adjusting sparse structures during training, existing methods rely on magnitude-based heuristics that are fundamentally ill-suited for the non-convergent, heterogeneous environments inherent to FL. To address this challenge, we propose FedFit, a federated dynamic sparse training framework that replaces simple heuristics with optimization-centric criteria for structure adjustment. By leveraging a second-order approximation of the loss landscape via the Fisher Information Matrix, FedFit enables precise and efficient structure adjustment without the overhead of explicit Hessian computation. Empirical evaluations across computer vision and natural language processing benchmarks demonstrate that FedFit significantly narrows the sparse-to-dense accuracy gap, outperforming state-of-the-art methods while maintaining high communication efficiency. Our code is available at this https URL .

[AI-233] Agents Cannot Hand Over What They Have Not Yet Understood

链接: https://arxiv.org/abs/2611.18642
作者: Yibo Liu
类目: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 16 pages, 6 figures

点击查看摘要

Abstract:Agents that work beyond a single context window, or in teams, hand their work to a successor as a short written summary. Evaluation has focused on whether an agent can solve a task, and relatively little is known about how much of what an agent has learned survives being written down. We study this in 56 interactive worlds with hidden rules, where an agent that is told nothing fails and an agent given a 300-character statement of the rules succeeds. We compare one agent that plays eight lives and keeps its whole history with a chain of eight agents that each play one life and inherit a 300-character note. All six models we test learn less in the chain: after eight generations it stands where the single agent stood after about four lives. Controlled variants of the chain decompose this handoff loss into three components: none from changing the agent, a smaller loss from compressing to 300 characters, and a larger loss from compressing repeatedly. When each note is written from the full record of earlier lives instead of from the previous note, the same 300 characters recover 70% to 82% of the loss. The loss is robust across model sizes and is not alleviated by instructing the writer to record evidence, to revise the inherited note, or to keep a structured ledger. In simpler terms, an agent cannot hand over what it has not yet understood: a note holds what its writer has worked out, and a rule that takes several lives to see is lost with the lives that were not passed on.

[AI-234] Multimodal Fusion with Pre-trained Models for Psychotherapy Outcome Prediction from Emotional Dynamics

链接: https://arxiv.org/abs/2611.18635
作者: Tobias Hallmen,Marius Funk,Antonia Vehlen,Jeanne Hierse,Gerhard Prinz,Steffen Eberhardt,Wolfgang Lutz,Elisabeth André
类目: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
备注: Preprint. 18 pages, 7 figures; 6-page supplementary material as ancillary file. This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

点击查看摘要

Abstract:In-session emotional dynamics are associated with psychotherapy outcomes but difficult to measure at scale. Pretrained multimodal models provide automated access to facial, vocal, and textual signals, yet clinical datasets remain small. Our research question is how pretrained multimodal representations can best support psychotherapy outcome prediction under data scarcity. Three studies use 106 sessions from 88 patients and nine session-level outcome scales (HSCL-11, EPO-1, POMS). Study 1 asks whether deep learning outperforms classical machine learning: a two-stage deep pipeline, first predicting window-level valence/arousal and then session outcomes, transferred poorly (best single configuration, r=.078 ). Study 2 asks which representation and preprocessing choices work best: statistical pooling of frozen pretrained embeddings with regularized classical models achieved average r=.406 across nine scales under per-scale configuration selection, up to r=.490 for symptom distress, with complementary modalities and no significant benefit from manual affect annotations or therapist features. Study 3 asks whether deep learning remains inferior under these choices: classical models outperformed an optimized single-stage deep model at the single-configuration level ( r=.281 vs. .222 ) and on all nine scales under matched per-scale selection ( .406 vs. .252 ). The results suggest that under clinical data scarcity, pretrained representations with fixed statistical pooling and regularized classical models are more reliable than learned aggregation.

[AI-235] Minimum cores in lattice graphs: zonotope structure and quartic growth

链接: https://arxiv.org/abs/2611.18633
作者: Navid Mehrdad
类目: ocial and Information Networks (cs.SI); Artificial Intelligence (cs.AI); Combinatorics (math.CO)
备注:

点击查看摘要

Abstract:A core of a graph is a nonempty vertex set in which every member has at least as many neighbours inside the set as outside. We study the minimum size c_m,n of a finite core in the graph on \Z^m joining points at sup-norm distance at most n . The local majority condition imposes a global geometric constraint: the convex hull of every core has an explicit lattice zonotope H_n^(m) as a Minkowski summand. In the plane, this zonotope has width asymptotic to 6\zeta(3)n^3/\pi^2 . We prove that all its lattice points form a core and construct a shell with O(n^4) points. Together with a packing argument, this gives c_2,n=\Theta(n^4) and coordinate widths \Theta(n^3) for every minimum core. We determine c_2,1=12 and c_2,2=96 by analytic arguments and prove uniqueness up to translation in both cases. Explicit forced-set and deficit certificates determine further minima within the prescribed zonotope hull. We distinguish these restricted results from global optimality and state the remaining geometric conjectures. We also give a direct reduction from \textscClique showing that the core problem is NP-complete on general finite graphs, which settles a conjecture of Hassanpour in the terms in which it was posed.

[AI-236] Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model

链接: https://arxiv.org/abs/2611.20986
作者: Yong Liu,Zhanpeng Shi,Peicong Lin,Zhenyi Zhong,Yizhou Dang,Wendi Wu,Haoling Li,Shuangjia Zheng
类目: Quantitative Methods (q-bio.QM); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Biomolecular foundation models predict the structures of proteins and their complexes accurately, and large language models (LLMs) reason in natural language, but the two remain separate: the abilities of the former are not part of a general-purpose LLM, and the reasoning of the latter does not enter a structure prediction. Calpha aims at one general model that folds proteins and complexes, understands protein sequence and structure, and lets its language reasoning influence the structures it predicts. This report covers the first step, in which an LLM outputs single-chain protein structures. Qwen3.5-9B reads a sequence enriched with ESM2 residue features. It can respond in text or emit a token that invokes a pretrained SimpleFold structure head, which receives information about the protein only through Qwen’s residue states. Under a fixed folding request, the head achieves mean TM-scores of 0.808 on the FoldBench-334 development set, 0.635 on CASP16 domains and 0.900 on a CAMEO-style PDB set. We compare these predictions with a baseline in which ESM2-3B encodes the sequence and a fine-tuned SimpleFold-3B head predicts the coordinates. The two systems also differ in trainable weights, training duration and crop length. For chains of 100–299 residues, all three paired confidence intervals include zero. At 300–600 residues, the mean differences are -0.044, -0.160 and -0.025, respectively. In a separate test, four English request types triggered folding on every evaluated example. Protein training reduced accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. An LLM can therefore output single-chain structures from its own states, with lower long-chain scores in the current comparisons and reduced accuracy on the evaluated MMLU-Pro subset. Complex prediction, structure understanding and the effect of language reasoning on the predicted coordinates remain future work.

[AI-237] GranularGPT : Natural-Language-Driven End-to-End Discrete Element Method Simulations

链接: https://arxiv.org/abs/2611.20551
作者: Erlei Li,Sirui Tang,Ai Wang,Mingwei Hu,Yuchen Dai,Geoff Wang
类目: Computational Physics (physics.comp-ph); Artificial Intelligence (cs.AI)
备注: 21 pages, 9 figures

点击查看摘要

Abstract:Discrete element method (DEM) simulations are widely used to investigate granular and particulate systems, but their application often requires specialized expertise in particle physics, numerical modelling, and simulation post-processing. Large language models (LLMs) offer new opportunities to simplify these workflows by enabling users to interact with simulation tools through natural language. This study offers GranularGPT, an LLM-assisted agentic framework for end-to-end DEM simulation workflows, from natural-language problem descriptions to simulation execution and automated post-processing. The framework enables users to generate DEM simulation cases, execute and monitor simulations, and automatically extract and report physical and process-level metrics. Evaluation across 60 normal-condition cases covering 12 DEM scenarios and parameter variations achieve a simulation success rate of 91.7%. In the error-handling evaluation, all 10 repairable errors are detected and 9 are successfully resolved, while all 5 non-repairable cases are correctly identified and reported. The framework automatically extracts a range of DEM-relevant metrics, including particle size distributions, packing and density measures, velocity profiles, discharge rates, mixing indices, contact force characteristics, and thermal quantities. GranularGPT provides a foundation for improving the accessibility and efficiency of DEM simulation workflows while reducing the specialized expertise required for their operation.

[AI-238] Multi-output neural network surrogate for double-pulse femtosecond optothermal control of thin silicon films on fused silica

链接: https://arxiv.org/abs/2611.20301
作者: George D. Tsibidis
类目: Optics (physics.optics); Disordered Systems and Neural Networks (cond-mat.dis-nn); Materials Science (cond-mat.mtrl-sci); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:We introduce a multi-output feedforward neural network surrogate trained on a coupled multiple-reflection and two-temperature model to map double-pulse femtosecond laser excitation (1030 nm) of crystalline silicon films on fused silica across film thicknesses (50-500 nm) and inter-pulse delays (0-10ps). Trained via Levenberg-Marquardt optimization on 2000 physical simulations, the dual-hidden-layer network ([40,30] neurons) simultaneously predicts six strongly coupled optothermal observables, peak electron temperature, peak lattice temperature, peak carrier density, minimum and maximum absorptivity and minimum refractive index achieving coefficients of determination R2 larger than 0.9998 across all targets. Evaluating a dense 200 by 200 design grid in milliseconds resolves pronounced thin-film interference fringes and carrier-induced index modulation, identifying an optimal optothermal coupling regime at d=156 nm and Dt=4.572 ps. This unified machine-learning workflow eliminates computationally expensive multi-physics runtime bottlenecks for real-time laser processing control and inverse photonic thin-film design.

[AI-239] Integrating Latent Block Models into Variational Autoencoders for Disentangled Representation Learning

链接: https://arxiv.org/abs/2611.20253
作者: Anasse Essalih,Reda Khoufache,Laurent Durville,Mohamed-Cherif Rahal,Mustapha Lebbah,Hanene Azzag,Walid Kheriji,Fouad Hadj Selem
类目: Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:We introduce LBM-VAE, a novel framework that integrates Latent Block Models (LBMs) into Variational Autoencoders (VAEs) to jointly cluster observations and latent dimensions. Beyond its clustering capabilities, LBM-VAE provides a block-structured latent representation that naturally supports disentangled representation learning. In contrast to conventional approaches that associate each factor of variation with an individual latent dimension, our framework allows factors of variation to be encoded across groups of latent dimensions. Experimental results demonstrate the promising clustering performance of LBM-VAE, while achieving competitive results in disentangled representation learning. LBM-VAE highlights the potential of combining co-clustering and variational inference to learn structured and interpretable latent representations.

[AI-240] Markets for Compute and Inference: Structure Financing and Risk Transfer

链接: https://arxiv.org/abs/2611.20242
作者: Dr Gabriel Borrageiro
类目: Computational Finance (q-fin.CP); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Capacity to run large language models is rented by the accelerator-hour, and their output is sold by the million tokens. Both goods now have published price indices, an over-the-counter forward has traded on one, a perpetual future on one is listed offshore, and CME and ICE have specified futures on rental indices that await the United States regulator. We ask what in these two markets can be hedged, by whom, and at what price. Three findings stand out. First, the price that exchange contracts settle on is not the price that drives operators’ revenue. Every contract references the published short-dated rental rate, while most revenue, and with it credit quality, comes from private multi-year contracts whose prices are largely undisclosed. Second, the token price and the rental rate are linked by an identity through throughput and margin, but they are different exposures, so a hedge on the rental rate does not hedge the cost of inference; at the measured capability of GPT-5 the cheapest list price fell by a factor of 17 in twelve months, and by far less at lower capability. Third, the demand for hedging already sits in credit agreements, as the years by which operators’ debt outlives their contracted revenue. Filings cannot measure that gap, because debt is scheduled year by year while contracted revenue rarely is, and the one loan that prices it settles the risk by cash prepayment against a model only the parties see. We design a reverse convertible note through which leveraged operators could buy protection against falling rental rates. Its coupon depends steeply on the annualised volatility of the rental rate, which the two public indices put between about 5% and 89% according to how each is built and the horizon measured; across that range the fair one-year coupon runs from 4.1% to 31.5%.

[AI-241] Clinician Assessment of Laryngeal Disease with Auditory AI before Laryngoscopy: A Paired Reader Study

链接: https://arxiv.org/abs/2611.20050
作者: Siyuan Hou,Shuting Yang,Zengrui Jin,Liyu Cheng,Siyin Wang,Yunhan Shi,Xueyan Li,Rong Hu,Yuhong Lin,Qiyu Chen,Yanru Li,Ji Wu,Bowen Zhou,Mangsuo Zhao,Wen Xu,Chao Zhang
类目: Audio and Speech Processing (eess.AS); Artificial Intelligence (cs.AI)
备注: 45 pages, including supplementary material

点击查看摘要

Abstract:Background: Before laryngoscopy, clinicians assess the likely cause of hoarseness and whether to suspect a precancerous or malignant lesion. Whether voice-based artificial intelligence (AI) supports pre-laryngoscopic categorization is unknown. Methods: We developed an auditory large language model that generates a patient-level assessment card (provisional category, voice-quality ratings, and rationale) and a prespecified malignancy alert from voice recordings, patient information, and acoustic measures obtained before laryngoscopy at one center. We evaluated both in a held-out test set and a later cohort. In a sequential reader study, 32 otolaryngology clinicians categorized held-out cases without AI and, 1 to 2 weeks later, reviewed their initial categories with the card (without the alert). Results: Across 800 paired assessments, reader accuracy was 49.8% before and 56.8% after review (difference, 7.0 percentage points; 95% CI, 1.7 to 13.5; P 0.001); 99 categories changed from incorrect to correct and 43 from correct to incorrect (ratio, 2.30; 95% CI, 1.22 to 5.31). Accuracy rose 14.7 points when the model was correct and fell 8.7 when incorrect. Provisional-category accuracy was 72% in held-out testing and 67% in temporal validation. The malignancy alert flagged 85.7% and 86.8% of precancerous or malignant lesions. Sensitivity for precancerous or malignant lesions increased from 53.9% to 75.0% after AI-assisted review (difference, 21.1 percentage points; 95% CI, 4.3 to 42.0). Conclusions: In this single-center study, clinicians’ pre-laryngoscopic categorization accuracy was higher after reviewing the AI assessment card, although the sequential design cannot fully separate AI assistance from reassessment; gains depended on model correctness without clearly selective adoption. Randomized and external evaluations, including of the malignancy alert, are needed. Comments: 45 pages, including supplementary material Subjects: Audio and Speech Processing (eess.AS); Artificial Intelligence (cs.AI) Cite as: arXiv:2611.20050 [eess.AS] (or arXiv:2611.20050v1 [eess.AS] for this version) https://doi.org/10.48550/arXiv.2611.20050 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Zengrui Jin [view email] [v1] Sun, 15 Nov 2026 14:53:14 UTC (4,602 KB)

[AI-242] Equal Dose Different Coverag e:Tracing Rate and Duration in Image Editing

链接: https://arxiv.org/abs/2611.19955
作者: Yi Liu
类目: Image and Video Processing (eess.IV); Artificial Intelligence (cs.AI); Multimedia (cs.MM)
备注:

点击查看摘要

Abstract:Benchmarks for scientific image editing test whether a model draws the outcome of a described process, leaving open which representations carry its errors and whether correcting them yields the right quantity. We study this question with an airbrush of fixed deposition profile, under which exchanging rate and duration preserves the dose and should preserve the endpoint, yet changes the edited coverage. Matched-norm activation patches across text positions and denoising stages show that, in distilled FLUX with Chinese prompts, representations at unchanged text positions (context) carry this discrepancy: early context transfers 3.51 of the 3.98 percentage-point (pp) signed route difference, against 0.09 pp for equal-norm updates at the two changed digits. Early image latents transmit native (unpatched) route differences across two FLUX configurations and OmniGen, while text-position effects depend on configuration. The trace predicts an intervention: share early context between routes while each retains its own digits. On independent English layouts, this reduces the distilled-FLUX absolute route gap from 1.13 to 0.60 pp, outperforms early digit sharing, and improves agreement more than in base FLUX under directly paired tests. Calibrated evaluation, which prescribes an area for each dose, separates this gain from accuracy: sharing the full prompt encoding produces identical route images with target errors of 8-16 pp, and stating the target percentage directly retains a compressed area response. Localization predicts a consistency gain on new layouts, and calibration measures the quantity error that agreement leaves unresolved.

[AI-243] Learning Autoregressive Neural Operators on Solver-Defined Physical Constraint Manifolds KDD2027

链接: https://arxiv.org/abs/2611.19753
作者: Nannan Zhou,Guanmeng Xian,Ning Yang
类目: Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Fluid Dynamics (physics.flu-dyn)
备注: Accepted by KDD 2027

点击查看摘要

Abstract:Autoregressive neural operators approximate time-dependent PDE solvers by repeatedly composing a learned transition, although the resulting states need not remain feasible under the discrete constraints of the numerical problem. We study this issue for incompressible flow, where each velocity state must satisfy solver-level incompressibility together with prescribed boundary values. We formulate autoregressive operator learning as a map between solver-defined affine feasible sets and introduce PCM-NO, a projected-residual realization of this feasible-transition principle. The neural backbone predicts a candidate state increment, a fixed tangent projection extracts its admissible component, and a metric retraction maps the advanced state to the prescribed feasible set. With exact affine constraint solves, the resulting transition is closed on the feasible set, while an affine identity separates admissible-update parameterization from state-level closure. With approximate solves, tangent-only residuals admit a cumulative upper bound, whereas per-step retraction resets feasibility to the residual achieved by the current solve. Experiments on periodic Navier–Stokes flow and lid-driven cavity flow demonstrate the superiority of PCM-NO in preserving solver-level constraints over long autoregressive rollouts and under parameter, boundary-value, and resolution shifts, while achieving lower rollout errors and better flow-specific diagnostics across both in-distribution and stress settings.

[AI-244] Joint Learning of Factual and Counterfactual Worlds for Causal Prior-data Fitted Networks

链接: https://arxiv.org/abs/2611.19710
作者: Furkan Danisman,Luke Zhang,Zhijing Jin,Dehan Kong
类目: Machine Learning (stat.ML); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注:

点击查看摘要

Abstract:Estimating treatment effects is central to science across many disciplines and requires counterfactual reasoning about alternative worlds. In the real world, however, each unit reveals its outcome only under the factual world. Thus, the factual and counterfactual outcomes can never be jointly learned by any model trained on real data. The problem is therefore tractable only in synthetic settings, where both worlds can be generated. Recent progress on prior-data fitted networks (PFNs) has shown that models pre-trained purely on synthetic data can transfer to real-world prediction. This paradigm has since carried over to causal effect estimation. Despite this fully synthetic setup, existing causal PFNs carry the real-world constraint into their training objective, confining learning to a single world’s outcome per unit. The dependence between the two worlds is thereby lost, and the density of the treatment effect relies by default on independence, which is implausible in practice. We therefore propose joint learning, a training setup in which a causal PFN learns the joint distribution of the factual and counterfactual outcomes of each unit. From the joint distribution, we obtain both the response and the individual treatment effect densities, and at inference, we apply optimal transport across units to obtain an approximation of the average treatment effect density. Ours is therefore the first causal PFN approach to provide all three densities together while accounting for the cross-world dependence. We apply joint learning and optimal transport to existing causal PFNs and evaluate it across semi-synthetic, synthetic, and real benchmarks spanning diverse causal scenarios. The jointly trained models significantly improve upon their baselines in point estimation and uncertainty quantification in most settings, indicating that joint learning is a preferable training setup for causal PFNs.

[AI-245] Generation of protein conformations using a simple inverted transformer

链接: https://arxiv.org/abs/2611.19321
作者: Luigi Leonardo Palese
类目: Quantitative Methods (q-bio.QM); Artificial Intelligence (cs.AI); Biomolecules (q-bio.BM)
备注:

点击查看摘要

Abstract:Molecular dynamics is an extremely powerful approach for studying protein behavior, but it requires enormous computational resources to obtain an adequate sampling of possible conformations. Recently, thanks to successes in predicting three-dimensional protein structures, there has been an explosion of research exploring the opportunities offered by deep learning to address the problem of protein conformational landscape. In this work, an artificial intelligence system based on a variant of the Transformer architecture is introduced and applied to the problem of determining the ensemble of three-dimensional structures that a protein can access and their statistical weight. After being trained on a dataset of conformations of a particular protein obtained using classical all-atom molecular dynamics, this neural network is able to directly generate physically plausible protein conformations similar to those obtained both through molecular dynamics and experimentally. The ordered sequence of conformations obtained as output of the neural network rapidly converges towards the most populated regions of the protein’s conformational landscape, suggesting that the system may have learned not only the protein’s three-dimensional organization but also the thermodynamic weights of the various feasible conformations. These features, combined with its simplicity and low computational load, make this Transformer variant an ideal candidate for the development of systems capable of performing generative molecular dynamics.

[AI-246] Mean-field Token Steering

链接: https://arxiv.org/abs/2611.19163
作者: Md Zulfiqur Haider,Asmaa Eldesoukey,Abhishek Halder,Yongxin Chen
类目: Optimization and Control (math.OC); Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Systems and Control (eess.SY); Machine Learning (stat.ML)
备注:

点击查看摘要

Abstract:We consider stochastic token steering through a deep stack of transformer layers as a mean-field control problem for an interacting particle system on the unit sphere, where interactions arise from the attention mechanism. In the mean-field limit, the token population evolves via a controlled integro-partial differential equation where the feedback control acts as a corrective vector field guiding the population toward an aggregate configuration of tokens in the representation space. We deduce the minimum effort control to accomplish this task over a fixed time horizon interpreted as normalized layer depth. We show that the resulting controlled dynamics for the token population can be understood as a mean-field Schrödinger bridge adapted to the sphere geometry. Building on this, we propose a novel generalization of the dynamic Sinkhorn recursion to compute the consistent token evolution. Numerical experiments confirm that the proposed approach achieves exact steering of the token population.

[AI-247] SPARC: Evidence-guided adaptive model composition for materials discovery

链接: https://arxiv.org/abs/2611.19050
作者: Sissi Feng,Yuchen Han,Runze Zhang,Jiaru Bai,Brandon Sutherland,Ali Shayesteh Zeraati,Jason Hattrick-Simpers,Varinia Bernales,Alán Aspuru-Guzik,Yang Bai
类目: Materials Science (cond-mat.mtrl-sci); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
备注: 52 pages, 5 figures; main text with Supplementary Information

点击查看摘要

Abstract:Closed-loop experimentation requires more than selecting the next experiment. It must also determine which existing measurements and model-derived relationships are sufficiently supported to influence that decision. Here we introduce SPARC (Sparse Predictive Aggregation of Regression Components), a framework that continually revises this support as experimental evidence accumulates. Competing model families are evaluated through out-of-fold predictions, so that each model is tested on observations not used in its corresponding fit. Sparse posterior composition assigns model influence according to predictive contributions supported by the out-of-fold evidence, retaining complementary contributions from several model families. Candidate-dependent predictive dispersion and residual variability define trust, which modulates exploration alongside the predicted response. Independent diagnostics also prioritize existing measurements for prospective verification, with model-derived trust reported alongside them. Across eight published materials datasets, SPARC achieved higher R^2 than the corresponding literature-reported values in six, with relative gains of up to 77% without the density-functional-theory- or experiment-derived descriptors used by several published pipelines. In an automated alkaline oxygen-evolution-reaction campaign, 8 of 11 SPARC-selected experiments reached overpotentials below 300 mV, with a best value of 242 mV, compared with 18 of 125 initial experiments. Measurements prioritized for verification before repetition subsequently deviated by 76 mV on average, compared with 27 mV for matched controls. SPARC therefore turns experimental evidence into a basis for revising model influence, directing exploration and identifying measurements that warrant verification.

[AI-248] he opposing informational effects of AI trading

链接: https://arxiv.org/abs/2611.18872
作者: Hajime Shimao
类目: General Economics (econ.GN); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:Better artificial intelligence (AI) can improve individual forecasts without making prices better at revealing new information. In a competitive market with resource-constrained forecasting, more capable algorithms can substitute analysis of public records for private observations, reducing the private information aggregated into prices. The same trades can convey sophisticated public-data analysis to investors unable to perform it themselves. We establish conditions under which these effects coexist and the private benefit of further computation falls. Numerical equilibrium calculations and finite-data neural-market experiments exhibit corresponding patterns. Out-of-sample earnings forecasts and controlled language-model tasks show related patterns of information substitution. In the neural experiment, wider adoption increases the predictive value of price for limited observers, reduces it for advanced observers, and lowers the benefit of upgrading by 44%. Retraining by investors who did not upgrade can amplify the loss of private-information precision. The informational consequences of AI trading depend on what prices reveal and who can use them.

[AI-249] An LLM -assisted theoretical model for aeroengine installation noise scattering

链接: https://arxiv.org/abs/2611.18796
作者: Xun Huang
类目: Fluid Dynamics (physics.flu-dyn); Artificial Intelligence (cs.AI)
备注:

点击查看摘要

Abstract:A theoretical model is developed for the canonical problem of aeroengine fan noise impinging on, and scattering from, a nearby semi-infinite rigid plate. The model chains two Wiener–Hopf problems in different coordinate systems. The first stage solves the duct-radiation problem analytically in axisymmetric cylindrical coordinates. The second stage expands the field incident on the plate into a double angular spectrum of plane waves: the axial spectrum is inherited from the first stage, while the spanwise spectrum is generated by a Weyl–Sommerfeld expansion of the cylindrical harmonics. Each spectral component is then scattered by the plate. The mathematical crux is the coordinate transformation between the axisymmetric duct frame and the Cartesian plate frame. The core physical insight–that the Weyl–Sommerfeld identity provides an exact spectral bridge between these systems–originates entirely with the authors. The heavy algebra of translating between the two spectral representations, together with the associated numerical verification, was carried out with the assistance of a large language model (LLM), under strict human oversight and independent numerical cross-checking. This division of labour enables the LLM to handle tedious symbolic manipulation without compromising rigour. Results are presented for spinning modes radiating from a cylindrical duct with a subsonic jet into a quiescent ambient. The case study demonstrates that, under systematic verification, LLMs can accelerate tedious algebra in theoretical fluid mechanics, contributing a concrete empirical example to the broader discussion on AI’s role in theoretical research.

[AI-250] Vibe synthesis: hierarchical agent ic decision-making for intent-driven multistep molecule synthesis

链接: https://arxiv.org/abs/2611.18772
作者: Jie Chen,Guikun Chen,Feng Wan,Ning Xu,Wen Liu,Yixiang Ruan,Yiyi Qiu,Haohua Li,Wei Wang,Yi Wang,Xiaoxue Wang,Peng Wang,Jiazhou Zhu,Jun Xuan,Chengxi Li,Jianhong Xu,Wenguan Wang,Yiming Mo
类目: Chemical Physics (physics.chem-ph); Artificial Intelligence (cs.AI)
备注: 30 pages, 5 figures

点击查看摘要

Abstract:The ability to construct complex molecules underpins advances in medicine and the molecular sciences. Yet long-horizon multistep synthesis requires decisions to be revised as evidence accumulates across chemically dependent operations, a capability not readily supported by artificial-intelligence tools and automated platforms organized around predefined tasks and hard-coded workflows. Here we propose and demonstrate a vibe synthesis system (VibeSyn), an intent-driven architecture in which hierarchical agents translate chemist-defined molecular objectives and accumulating evidence into adaptive decisions executed through computational tools and laboratory automation. We evaluate VibeSyn in three demonstrations of increasing scope and autonomy, progressing from evidence-grounded experimental decisions to adaptive execution of a ten-step total synthesis of catharanthine and structure-to-compound synthesis of cyclic peptides containing unnatural amino acids (uAAs). Across the long-horizon campaigns, active human input and simple confirmation account for less than 7.5% and 5.2%, respectively, of recorded workflow actions and outcomes. VibeSyn points towards a future in which chemists define synthetic intent and agent-directed laboratories autonomously navigate the evolving experimental trajectory required to deliver the target compound.

[AI-251] Selection-Aware Certification of Client-Constrained Fraud Rule Portfolios Under Synthetic Distribution Shift

链接: https://arxiv.org/abs/2611.18687
作者: Rishabh Pande,Yu Zeng
类目: Risk Management (q-fin.RM); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR); Machine Learning (cs.LG)
备注: 7 pages, 2 figures, 4 tables

点击查看摘要

Abstract:Selecting fraud-rule portfolios to satisfy a precision or recall floor can overfit the data used for selection. We study finite-family certification for bounded OR portfolios, comparing point-estimate selection, a fixed safety margin, and one-sided Clopper-Pearson lower bounds with Bonferroni adjustment. In a factorial follow-up with 6,400 datasets from two author-designed synthetic generators, full-validation certification reduced the held-out floor-violation rate by 29.1 percentage points on average versus full-validation point selection, while reducing deployment by 31.7 points. Compared with a fixed 0.05 margin, certification reduced violations by 22.2 points and deployment by 29.4 points; results varied across conditions. Using the full rather than half validation sample increased deployment by 17.1 points on average, without a uniform reduction in shifted-test violations. In a separate 400-dataset no-shift sweep, certification deployment rose from 61.3% to 90.2% as validation size increased from 1,500 to 15,000; no violations occurred among 2,887 deployed portfolios tested on 100,000-row holdouts. A temporal case study on the public synthetic Bank Account Fraud Base dataset used recall at a 5% false-positive-rate cap; the selected candidate pool captured less than 1% of test fraud, showing limited coverage in this setup. Together, these experiments characterize how finite-family certification changes the reliability-deployment tradeoff for bounded OR portfolios under specified synthetic conditions. The evidence covers the evaluated generators and one public synthetic BAF case study; it does not estimate performance on real transaction populations or robustness to arbitrary distribution shift. The authors revised and rewrote this abstract after receiving a draft from OpenAI Codex.

机器学习

[LG-0] Hyperlyve: Parallelized Enumeration and Exact Lyapunov Stability Certification for Deep ReLU Networks

链接: https://arxiv.org/abs/2611.21599
作者: Brian Yarbrough,Jesse Wayment,Jingbo Wang,Philip E. Paré,Shreyas Sundaram
类目: ystems and Control (eess.SY); Computational Geometry (cs.CG); Machine Learning (cs.LG); Logic in Computer Science (cs.LO); Dynamical Systems (math.DS)
*备注:

点击查看摘要

Abstract:We introduce Hyperlyve (Hyperplane Partitioned Lyapunov Verification and Enumeration), a framework for proving the asymptotic stability of nonlinear dynamical systems, including those with neural network controllers. Hyperlyve learns and formally certifies continuous piecewise-affine (CPA) neural Lyapunov functions within a counterexample-guided inductive synthesis (CEGIS) loop. Hyperlyve’s key innovation is a parallel algorithm that enumerates the hyperplane arrangement formed by the ReLU activations of deep networks and certifies that Lyapunov conditions hold within each polyhedral region. We derive sound and complete regional tests to verify the asymptotic stability of the exact smooth dynamics (rather than a piecewise-linear approximation) within a compact domain of interest. Furthermore, we formally characterize the computational complexity of the CPA neural Lyapunov verification problem as co-NP-complete, filling a key gap in the literature. Empirical results demonstrate that Hyperlyve enumerates ReLU networks over 150-times faster than competing tools, which contributes to Hyperlyve proving the asymptotic stability of systems that are currently intractable for state-of-the-art verifiers.

[LG-1] Optimal Autoregressive Chain-of-Thought Learning via Bagging

链接: https://arxiv.org/abs/2611.21597
作者: Awni Altabaa,John Lafferty,Omar Montasser
类目: Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:We study how to learn an autoregressive next-token generator f\colon\Sigma^\star\to\Sigma from a class \mathcalF , given chain-of-thought (CoT) demonstrations of length T generated by f . We present a simple algorithm based on bagging, which outputs a majority vote of generators that are trained to be CoT-consistent on bootstrap samples of the training data. For binary token sets, it achieves the optimal sample complexity of O((d_\mathrmVC+\log(1/\delta))/\epsilon) , independent of the CoT length T , where d_\mathrmVC is the VC dimension of \mathcalF . For finite token sets \Sigma , the VC dimension is replaced by a multiclass analogue, the Graph dimension d_\mathrmG , and the sample complexity is optimal up to a \log|\Sigma| factor. The number of generators in the majority-vote ensemble is O(\log(T/\delta)) , and we show that the ensemble can be distilled into a single generator in \mathcalF at the expense of n\log(nT) unlabeled prompts.

[LG-2] SADL: On the Demand-Supply Geometry of Representation Collapse

链接: https://arxiv.org/abs/2611.21596
作者: Junhan Sun,Yigui Feng,Mingzhe Liu,Haoyan Xu,Lei Ye,Ziqi Han,Hanru Bai,Yanzhao Zhang,Hao Zhao,Guofeng Zhang
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注: 49 pages, including references and appendices

点击查看摘要

Abstract:In Joint-Embedding Predictive Architectures (JEPAs), a well-constrained projector output does not by itself prevent representation collapse. We find that, from initialization, the encoder representation is dominated by a component shared by different inputs. A batch-normalized projector removes this component, so regularization after the projector cannot correct it, and the differences between representations shrink until they collapse. We propose SADL (Supply And Demand Loss), which keeps the original training objective and constrains the encoder directly when needed. SADL compares the representation differences that the projector output demands with the differences the encoder can retain and pass on to the projector, and adds a constraint on the encoder only when supply falls short of demand. It works with different Demand Regularizers and introduces no auxiliary networks, additional trainable parameters, or architectural changes. In single-task and jointly trained four-task LeWorldModel (LeWM) and INTACT experiments, SADL lowers the collapse rate over ten random seeds from 50% to 0%. After one epoch of training, encoder effective rank is 3.4-6.4 times that of the corresponding baseline without SADL, and latent prediction error falls to 1/20-1/303 of it. Without additional tuning, SADL stays robust as its loss weight spans orders of magnitude, and with several Demand Regularizers INTACT exceeds 95% mean search-free control success. On four image datasets, SADL also consistently improves accuracy and stability across random seeds. These results show that matching the representation differences a regularizer demands with the differences the encoder can supply leads to faster and more stable representation learning in control and image classification.

[LG-3] CertiSum: Bitwise-Identical MXFP4 Training Across Platforms and Parallelism Layouts

链接: https://arxiv.org/abs/2611.21584
作者: Steve Rhyner,Yubo Gao,Stefan de Lasa,Sankeerth Durvasula,Nandita Vijaykumar
类目: Machine Learning (cs.LG); Distributed, Parallel, and Cluster Computing (cs.DC)
*备注:

点击查看摘要

Abstract:Training with MXFP4, a 4-bit microscaling format, is non-deterministic: identical inputs can produce different results across parallelism layouts and GPU vendors. These differences complicate debugging and training audits and can cause training and inference engines to disagree. We present CertiSum to make MXFP4 training bitwise invariant across runs, parallelism layouts, and GPU vendors. Its key insight is that native FP32 accumulation of MXFP4 products is often already exact for every summation order. CertiSum collects statistics during quantization to certify native matrix multiplication results, recomputes uncertified outputs exactly, and uses exact kernels directly where certification rarely succeeds. CertiSum keeps quantization blocks unchanged across layouts and merges partial sums without rounding, including in attention and weight-gradient computation. In 200-step Qwen3-8B-Base experiments, CertiSum preserves training states bit for bit across AMD MI350X and NVIDIA GB200 and across changes in data, tensor, and context parallelism on MI350X. Across three 8B–14B models on FineWeb-Edu using MI350X, it adds 11.8–60.5% to training-step latency relative to its non-exact MXFP4 counterpart at the larger evaluated input shapes, with speedups of up to 1.24\times over BF16 and 1.70\times over deterministic BF16.

[LG-4] DynSurvX: Time-Resolved Explanations for Dynamic Survival Predictions

链接: https://arxiv.org/abs/2611.21571
作者: Van Tuan Nguyen,Agathe Guilloux
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Recent survival models integrating longitudinal data achieve strong predictive performance, but their predictions are often difficult to interpret, limiting their clinical use. We propose DynSurvX, a model-agnostic explainability framework for dynamic survival prediction with mixed static and longitudinal covariates. Based on Shapley values, DynSurvX introduces a hierarchical two-phase strategy that first identifies influential longitudinal features and then adaptively refines their temporal importance, enabling fine-grained explanations while reducing computational cost. The framework provides time-resolved explanations that evolve as new patient data become available. Experiments on simulated and two clinical datasets show varying agreement with theoretical importance across prediction models, with closer agreement for better-performing models, alongside clinically interpretable insights in real-world settings. It remains robust across prediction settings and reduces computation time by 41–80% relative to the temporal explanation baseline, making dynamic explanations practical.

[LG-5] nsorized Automata: Algebraically Informed Representation Learning for Sequential Decision Making

链接: https://arxiv.org/abs/2611.21514
作者: Anand Balakrishnan,Dweep Trivedi,Swarat Chaudhuri
类目: Formal Languages and Automata Theory (cs.FL); Machine Learning (cs.LG)
*备注: This manuscript is an extended version of the paper submitted to L4DC 2027

点击查看摘要

Abstract:Task representations for sequential decision making should describe both an objective and progress toward it. We present an algebraically informed pipeline that uses the relational structure and language semantics of alternating finite automata (AFAs) to learn control features and compile compact execution operators. AND-OR message passing learns task and state vectors; selected configurations define an event-driven Boolean matrix operator. We prove that initial and successor language coverage make its carrier represent the exact residual after every prefix, with a sufficient capacity bound in terms of reachable configurations. Conjunction and disjunction use the same pipeline. Fidelity and composition audits evaluate copied formulas and learned disjunct assignments. In five-policy-seed XLand experiments with larger training pools, pooled execution-state vectors outperform raw carrier bits and task-blind conditioning across three task pools under two rejection rewards. They achieve 77.3% success on multi-event tasks and 99.4% on a four-task stratum with 33–128-state minimal DFAs, using 16 operator slots. LetterWorld results show useful residual conditioning alongside a remaining gap to RAD. These results connect algebraically informed representation learning to compact task execution and useful policy conditioning.

[LG-6] Leverag e-Enhanced Sampling Owen Method for Explainable Artificial Intelligence

链接: https://arxiv.org/abs/2611.21504
作者: Jing Liu,Elena Parilina,Chi Zhao
类目: Computer Science and Game Theory (cs.GT); Machine Learning (cs.LG)
*备注: 25 pages, 6 figures

点击查看摘要

Abstract:Recently, the Shapley value has been widely applied in machine learning for feature importance analysis and model explainability. The multilinear extension of games proposed by Owen (1972) expresses the Shapley value as the integral of partial derivatives along the main diagonal of the cube. Direct evaluation of this integral requires exponential-time computation of the partial derivatives. To improve computational efficiency, we propose a sampling-based estimation strategy within Owen’s multilinear extension framework, combining stratified sampling with coalition leverage scores. Experimental results demonstrate that the proposed method significantly reduces computational costs while maintaining explanation accuracy.

[LG-7] Variationally correct physics-informed operator learning for nonlinear partial differential equations

链接: https://arxiv.org/abs/2611.21469
作者: Yuan Qiu,Wolfgang Dahmen,Peng Chen
类目: Numerical Analysis (math.NA); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Physics-informed operator learning trains neural operators on the residual of a parametric partial differential equation (PDE) instead of labeled solution data, but its reliability depends on a variationally correct residual: one that is equivalent to the error in a norm dictated by the equation, the PDE-compliant norm. For nonlinear equations written as first-order systems, we prove that the first-order system least-squares residual is equivalent to the error in this norm near a solution branch if the linearized residual is stable and the residual derivative is Lipschitz continuous, possibly in a norm stronger than the error norm. We verify these conditions for monotone reaction, solution- and gradient-dependent diffusion, and compressible Neo-Hookean hyperelasticity. The stability mechanisms vary across these classes, as do the regions where the equivalence holds, from arbitrary bounded sets to a neighborhood of the branch. In a conforming finite element space, we show that the loss excess of a state near the finite element solution is equivalent to its squared distance from that solution. We use this loss to train a reduced basis neural operator with values in this space, with solution snapshots needed only for the basis and the loss and its symbolically derived gradient assembled by the finite element library. At inference, this loss estimates the error. Numerically, the square root of the loss excess tracks the error within factors comparable to those of the linear case, and the operator achieves higher accuracy in the PDE-compliant norm than grid-based baselines evaluated in the same finite element space.

[LG-8] Variational Monte Carlo for Quantum Excited States via Nested Low-Rank Approximation NEURIPS2026

链接: https://arxiv.org/abs/2611.21434
作者: Minchan Jeong,J. Jon Ryu,Se-Young Yun,Gregory W. Wornell
类目: Machine Learning (cs.LG); Chemical Physics (physics.chem-ph); Computational Physics (physics.comp-ph)
*备注: Accepted for NeurIPS 2026. The first two authors contributed equally. 41 pages, 6 figures, 9 tables. Code is available at this https URL

点击查看摘要

Abstract:The electronic Schrödinger equation encodes a molecule’s electronic structure and quantum properties, but solving it scales exponentially with the number of electrons. Solving this equation with neural-network variational Monte Carlo (VMC) has reached chemical accuracy on the ground state of small molecules; extending this success to excited states is constrained by two existing paradigms. Penalty-based methods need K Laplacians per step but require tuning the overlap-penalty weights against the energy gaps they aim to compute. Natural excited states VMC (NES-VMC) is penalty-free, but it combines the K ansatzes into one determinant, whose local energy requires K^2 Laplacians per step. We resolve this trade-off via low-rank approximation (LoRA) reparameterization of the rank- K variational objective. The reparameterized objective admits a VMC-compatible stochastic gradient estimator via MCMC sampling from the per-state Born densities, and a sequential nesting scheme recovers the lowest K eigenstates in eigenvalue order. The resulting algorithm, NestedLoRA-VMC, inherits a global optimality guarantee from the LoRA principle. On the first-row atoms Li through Ne at K=10 , its accuracy is comparable to that of NES-VMC and of a state-of-the-art penalty-based method.

[LG-9] Efficient Modeling of Multi-Agent Systems via Structure-Aware Physics-Informed Learning

链接: https://arxiv.org/abs/2611.21404
作者: Yulun Wu,Nicola Bastianello,Karl Henrik Johansson,Matthieu Barreau
类目: Machine Learning (cs.LG); Optimization and Control (math.OC)
*备注:

点击查看摘要

Abstract:Modeling large multi-agent systems poses challenges for trajectory inference and parameter identification, as network size and coupling increase computational demands and partial observations require reconstructing unobserved states. We propose MAS-PINN, a neural-network architecture with a tailored training algorithm for a physics-informed learning framework that leverages the network topology and the agents’ similarity to reduce computational and memory costs of simulating and identifying multi-agent systems. Its architecture represents agent trajectories through shared temporal features and agent-specific coefficients. Training alternates between learning features, fitting coupled trajectories, and estimating physical parameters. A matrix-free update exploits local interactions to fit trajectories without forming large global matrices. Experiments on Lorenz rings, Lotka–Volterra networks, and Rössler oscillators show improved trajectory and parameter accuracy over baselines, including inference in large systems and identification under partial observations or unknown topology, alongside substantial computational savings in coefficient fitting.

[LG-10] MegNIST A Benchmark for Non-Invasive Inner Speech Decoding

链接: https://arxiv.org/abs/2611.21399
作者: Teyun Kwon,Pratik Somaiya,Gereon Elvers,Francesco Mantegna,Anna Camera,Sven Braeutigam,Mark Woolrich,Oiwi Parker Jones
类目: Machine Learning (cs.LG)
*备注: 17 pages, 8 figures, 5 listings

点击查看摘要

Abstract:MegNIST is a benchmark dataset for developing and evaluating machine learning methods for inner speech decoding from magnetoencephalography (MEG). The dataset comprises 12,000 trials of a single participant imagining speaking the digits zero through nine, recorded across four sessions. Each trial contains 306-channel MEG recordings sampled at 250 Hz, preprocessed using signal space separation, filtering, and downsampling. Data are provided as pre-defined training (10,000 trials), validation (1,000 trials), and test (1,000 trials) splits, with entire experimental runs assigned to each split to preserve temporal independence. The dataset is available in both BIDS format for neuroimaging analysis and as serialised tensors for machine learning workflows. Baseline classification results demonstrate that imagined digits can be decoded significantly above chance. Standardised splits, validated baselines, example code, and a PyTorch-compatible loader in the pnpl Python library enable straightforward integration with machine learning frameworks, lowering barriers for researchers to rapidly develop and benchmark inner speech decoding methods for non-invasive brain-computer interfaces.

[LG-11] Learning What Pairs Miss for Higher-Order Multimodal Alignment

链接: https://arxiv.org/abs/2611.21372
作者: Kunyu Zhang,Shenghui Feng,Xiangjie Shi,Zeqing Wang,Yingxu Wang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Contrastive learning beyond two modalities now scores the whole tuple with multilinear products, volumes, and fusion heads, critics that can represent the joint distribution. In practice these critics still behave like bags of pairs. Their negatives break at least two of the three pairs, a single pairwise check rejects them, and when most samples can be solved by pairs the higher-order term receives almost no gradient. To solve this problem, we propose CHORD (Contrastive Higher-Order Residuals with Decoys), which trains a residual against decoys drawn from a calibrated pairwise model, negatives that the pairwise model finds plausible, so that rejecting them often needs all three modalities. Drawing decoys costs one inner-product query over a memory bank, and the optimal residual is the log-ratio between the data and the pairwise model, so the decoy loss also bounds what that model misses. On M3-Mix, where one in ten training samples needs all three modalities, our CHORD improves accuracy on those samples by 21.2 percentage points over the strongest of eight baselines. CHORD also gives the best two-to-one results on every real dataset, and its residual improves other retrievers without retraining.

[LG-12] Federated TD Learning in Heterogeneous Environments with Averag e Rewards: A Two-timescale Approach with Polyak-Ruppert Averag ing

链接: https://arxiv.org/abs/2611.21370
作者: Ankur Naskar,Gugan Thoppe,Abbasali Koochakzadeh,Vijay Gupta
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Federated Reinforcement Learning (FRL) provides a promising way to speedup training in reinforcement learning when one has access to multiple edge devices that can operate in parallel. Recently, it has been shown that even when these edge devices have access to different dynamic models—such as cars or robots with differing system configurations—an optimal convergence rate that has a linear speedup proportional to the number of devices is achievable. However, this result requires that the algorithm’s stepsize be chosen in a manner that depends on the unknown model parameters. Also, it applies only to a discounted setting, which has been argued to fit episodic tasks better than continuing control tasks. In this paper, we obtain finite-time bounds for heterogeneous FRL with average rewards. We show that optimal convergence rate and linear speedup are achievable even with a universal stepsize choice, independent of the underlying dynamics. To achieve our result, we modify the existing one-timescale FRL method to a novel two-timescale variant that additionally incorporates iterate averaging. We remark that this idea also extends to other settings already studied in the literature, such as the homogeneous FRL or the heterogeneous case with a discounted cost function.

[LG-13] EMPO: Regime-Aware Operator Learning with Locally Adapted POD Bases

链接: https://arxiv.org/abs/2611.21344
作者: Rodion Akinzhala,Alexander Terentyev,Alexey Golubev,Yury Maximov
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Operator learning approximates the solution map of a parametric partial differential equation from data, enabling fast prediction at new parameter values without repeated simulation. The dominant approach fits a single global low-rank basis, typically a proper orthogonal decomposition (POD), to all solution snapshots. This fails when the snapshots span qualitatively different solution shapes, from smooth diffusive profiles to near-discontinuous shocks. We propose TEMPO (Trajectory Expectation-Maximization Mixture of POD Operators), a two-phase framework for regime-aware operator learning. In the offline phase, TEMPO recovers latent regimes via expectation-maximization, fitting a dedicated linear or nonlinear basis to each discovered regime. In the online phase, a gating network and per-regime branch networks are trained jointly, routing new inputs to the regimes at inference. We test TEMPO on Burgers’ flow, Sod shock tubes, Coanda flow, Darcy flow, Navier-Stokes flow and the heterogeneous Helmholtz equation. With 64 total POD modes, TEMPO gives 18.61% test error on Burgers’, compared with 19.61% for global POD. It also gives lower errors on Coanda and Helmholtz. Subjects: Machine Learning (cs.LG) Cite as: arXiv:2611.21344 [cs.LG] (or arXiv:2611.21344v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.21344 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-14] Wavelet-Energy-Driven Dynamic Patching for Time Series Forecasting

链接: https://arxiv.org/abs/2611.21319
作者: Syed Hafiz Bin Salim Alsree,Stephen Roberts,Stefan Zohren
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Patch-based transformers tokenise a time series into fixed-size, uniform windows, which commits the model to a single temporal resolution even though real signals are multi-scale. The discrete wavelet transform (DWT) is a natural alternative because it decomposes a signal across scales. A growing body of work makes the wavelet filter learnable, on the assumption that a more flexible wavelet forecasts better. That assumption has rarely been tested in isolation, because a single reported improvement usually credits three things at once. These are the use of any multi-resolution decomposition, the choice of wavelet and training the filter. We introduce a forecaster that turns a learnable-DWT per-scale energy map into variable-length patches by merging adjacent windows that share a dominant scale. Three findings follow. First, most of the improvement is there before any filter is learned. A frozen wavelet beats a fixed-patch baseline on eight of the nine datasets by 1.5% to 7.8%, and on four of them it recovers all of the gain. Traffic is the one exception, where it is 1.2% worse. Second, learning the filter gives a further improvement on only some datasets. It helps on three of the nine datasets and makes Traffic 2.7% worse. Third, having variable-length patches helps, but exactly where their boundaries fall does not. A fixed-size uniform grid, even with its patch size tuned per task, is worse than our variable-length patches on 27 of 36 settings. Recomputing the boundaries from the learned filter at every step, rather than caching them once, never helps. The improvement in wavelet-based forecasting comes from the multi-resolution structure and the variable-length patches it induces, not from making the wavelet learnable. In practice, start from any valid wavelet, cache the patch boundaries once and treat training the filter as a per-dataset decision rather than a default.

[LG-15] Deep Learning for Physicists: From Linear Regression to Reasoning Models

链接: https://arxiv.org/abs/2611.21295
作者: Masanori Hanada
类目: Machine Learning (cs.LG); High Energy Physics - Lattice (hep-lat); High Energy Physics - Theory (hep-th); Computational Physics (physics.comp-ph)
*备注: 115 pages, many figures. Based on lectures given at Queen Mary University of London

点击查看摘要

Abstract:These lecture notes are based on an introductory course on neural networks and deep learning that the author taught in the School of Mathematical Sciences at Queen Mary University of London in 2023/24 and 2024/25. The author, however, is a physicist, and it shows: wherever possible, ideas are explained with physical intuition rather than heavy formalism. Readers with a background in physics or mathematics who want a self-contained, introductory route into modern machine learning may therefore find these notes particularly friendly. The notes start from maximum likelihood, least squares, linear regression and (stochastic) gradient descent. They then cover neural networks, backpropagation, the vanishing gradient problem, batch and layer normalization, overfitting and regularization, adversarial examples, and the kernel method. The second half turns to natural language processing: word embeddings, recurrent networks and LSTMs, the attention mechanism and Transformers, and large language models, including scaling laws, the history of GPT, instruction tuning, reinforcement learning from human feedback, hallucination, alignment faking, chain-of-thought reasoning and test-time scaling. The emphasis on large language models reflects requests from the students. The notes are aimed at third- and fourth-year students in the UK system. Because the course was also taken by students from other departments, very little is assumed: even basic material such as matrix multiplication is explained from scratch, so readers familiar with it can safely skip such parts. No originality is claimed; the aim is a coherent, intuitive path through the material. Developments after spring 2025 are not covered, but the basic ideas emphasized here are unlikely to go out of date soon. Comments: 115 pages, many figures. Based on lectures given at Queen Mary University of London Subjects: Machine Learning (cs.LG); High Energy Physics - Lattice (hep-lat); High Energy Physics - Theory (hep-th); Computational Physics (physics.comp-ph) Cite as: arXiv:2611.21295 [cs.LG] (or arXiv:2611.21295v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.21295 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-16] he Price of Not Planning : Parallel Sampling Depth Peaks at the Reconstruction Threshold

链接: https://arxiv.org/abs/2611.21285
作者: Chuling Wen,Weijie Liang,Jian Lu
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Masked diffusion models sample the tokens of one round independently. Planned decoders first generate a coarse plan, then fill in the tokens in parallel. We ask when a plan saves rounds, how many it saves, and how fine it must be. We study this on broadcast trees: the leaves are the tokens, the internal nodes are the plan, and the edge correlation \theta sets how strongly tokens depend on the plan. Errors are measured by the forward Kullback-Leibler divergence \varepsilon , and the lower bounds hold for every deterministic value-adaptive rule, also with learned conditionals. We show that a decoder without the plan needs of order \min\n,(n/\varepsilon)^\alpha\ rounds below the reconstruction threshold, where \alpha=\log b/(4\log(1/\theta)) rises to 1/2 at the threshold, and that the exponent stays below 1/2 above it (binary channel: 0.02\le\theta\le0.98 ). With the plan, \log_b n+1 rounds suffice. In the binary model below the threshold, a plan with units of m tokens needs of order \log_b(n/m)+\min\m,(n/\varepsilon)^\alpha\ rounds, so it gives logarithmic depth exactly when m is logarithmic. A plan therefore helps most near the threshold, and only if its units are small. Exact divergence computations in the Gaussian model with up to 2^20 leaves confirm these exponents below the threshold.

[LG-17] Heteroscedastic Gaussian Process Q-learning for Sample Efficient Experimentation

链接: https://arxiv.org/abs/2611.21249
作者: Maximilian Bloor,Max Mowbray,Calvin Tsay,Ehecatl Antonio del Rio Chanona
类目: Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注:

点击查看摘要

Abstract:Self-driving laboratories must learn to control a process from tens to hundreds of expensive experiments, too few for deep reinforcement learning (RL). We learn the Q-function directly with a Gaussian process (GP), which avoids learning the dynamics, and show that the GP posterior supplies the learning rate of Q-learning. If each Bellman target is treated as a noisy measurement of the Q-function, the learning rate that minimises the mean squared error is the Kalman gain, and applying it to every observed transition is equivalent to GP regression. Heteroscedastic Gaussian Process Q-learning (HGPQ) sets the noise variance of each target with a heteroscedastic GP, which learns how unreliable the targets are across the states and inputs. Targets from noisy operating conditions, or from regions where the value function has large gradients, then move the estimate less than targets from low-noise regions, so HGPQ learns quickly where the data are reliable without being misled where they are not. For kernel value iteration with per-target variances that are assigned when a target is collected, we prove a high-probability bound on the error of the value estimate, uniform over the explored region. Pooled over five simulated benchmarks, which include a chemical reactor and a bioprocess, HGPQ approaches the best return found by any method in fewer episodes than GP, kernel and deep RL baselines.

[LG-18] MiSTiQ: MicroScaling Trellis Quantization for Extreme KV Cache Compression

链接: https://arxiv.org/abs/2611.21216
作者: Ofir Gordon,Lior Dikstein,Paul Balanca,Idan Achituve,Douglas Orr,Hai Victor Habi
类目: Machine Learning (cs.LG)
*备注: 26 Pages, 7 Figures, 7 Tables

点击查看摘要

Abstract:Modern large language models (LLMs) increasingly support longer contexts and serve more concurrent requests, increasing both sequence lengths and batch sizes. Both trends increase KV-cache size and memory traffic during inference. KV-cache quantization reduces these costs, with recent approaches using codebook-based vector quantization to achieve extremely low bit widths. However, hardware constraints limit codebook size, and most methods reconstruct the cache in higher-precision floating-point formats for attention, limiting the computational benefits of quantization. To address these limitations, we introduce MiSTiQ, a drop-in replacement for codebook-based vector quantizers that combines trellis-coded storage with microscaling (MX) formats for low-precision computation. We combine trellis quantization with beam search to encode the KV cache over large reconstruction spaces without large codebooks. To support MX computation, we derive MX-Hash functions that generate reconstruction values directly in the target MX element format. These functions use the scale-conditioned marginal distribution of block-normalized Gaussian inputs. On Llama3 and Qwen3, MiSTiQ achieves lower WikiText perplexity and higher GSM8K accuracy than the evaluated baselines across several block-scale formats at approximately 1.25 effective bits per element. With NVFP4, it also achieves the highest LongBench average on both models.

[LG-19] Open-fMIND: A Large-Scale Open-Access fMRI Corpus for Pretraining Brain Foundation Models

链接: https://arxiv.org/abs/2611.21215
作者: Carlo Alberto Barbano,Giovanni Marraffini,Demian Wassermann
类目: Machine Learning (cs.LG); Neurons and Cognition (q-bio.NC)
*备注: 11 pages + references and appendix

点击查看摘要

Abstract:Self-supervised pretraining has transformed language and vision, but its application to functional MRI (fMRI) is constrained by data access. The largest standardized fMRI resources are governed by data use agreements and dominated by a small number of acquisition protocols, while the thousands of openly licensed datasets on OpenNeuro are individually small, heterogeneously organized, and unprocessed. We present the first release of Open-fMIND, a uniformly preprocessed, openly accessible corpus of parcellated fMRI timeseries from 322 OpenNeuro studies, comprising 11,379 subjects and 45,731 functional runs. Each run is preprocessed with fMRIPrep and parcellated under 52 atlases spanning 14 families, yielding a massive 2.4 million timeseries. Timeseries are released either denoised with our default pipeline (preproc) or as the fMRIPrep output with its confound files (quasiraw), with functional connectivity matrices, per-run provenance and motion metadata. Open-fMIND is intended as a pretraining resource for self-supervised brain representation learning. Pretraining on it yields downstream performance comparable to that obtained with UK Biobank.

[LG-20] BoBDeLaN: Bounded-Bandwidth Context Adaptation for Deep Lagrangian Robot Control

链接: https://arxiv.org/abs/2611.21209
作者: Lucas Schulze,Sucheth Shenoy,Jan Peters,Oleg Arenz
类目: Robotics (cs.RO); Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注: Submitted to the 9th Annual Learning for Dynamics Control Conference (L4DC 2027)

点击查看摘要

Abstract:Learned physics-based models provide structure for robot control, but uncertain payloads and unmodeled dynamics limit their predictive accuracy. Context-aware Deep Lagrangian Networks enable adaptation through a latent context inferred from measurement histories while retaining Lagrangian structure for each fixed context. However, this structure alone does not constrain sensitivity to measurement disturbances. We introduce BoBDeLaN, which combines a contracting, gain-bounded Recurrent Equilibrium Network context encoder with stable linear filtering to shape disturbance sensitivity across frequencies. This architecture guarantees exponential forgetting of the recurrent initialization and admits a finite-horizon, frequency-weighted bound on context perturbations, which extends to residual torque prediction under a Lipschitz continuity assumption on the context-conditioned dynamics model. We evaluate a low-pass realization within model predictive control on simulated and real hardware manipulation tasks, together with frequency-sensitivity analysis on real-robot data. Compared with an LSTM encoder using the same input filter, BoBDeLaN exhibits reduced high-frequency context sensitivity, lower residual torque prediction error, and smoother motion, with less degradation under injected disturbances.

[LG-21] Decision-Relevant Forecast Calibration for Predictive Control

链接: https://arxiv.org/abs/2611.21203
作者: Victor M. Preciado
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注: 66 pages, 7 figures

点击查看摘要

Abstract:Calibration evidence for a predictive controller need not remain valid when operating-context frequencies change. We develop an audit and certification procedure for a declared family of context-density shifts, with a frozen predictor/controller and unchanged conditional outcome law. Coverage requires calibrated tests to span density-weighted selected-plan sensitivities. A residual Gram matrix yields the minimum augmentation rank and an optimal test budget for uncovered L_2 geometry. This optimum need not minimize bias over probability-feasible conditional means. For finite contexts, an asymmetric absolute-loss dual and directional linear programs enclose the continuous worst-shift feasible bias. Intersecting simultaneous moment intervals with the conditional-mean domain gives a further certificate and a linear program for selecting an operating shift after validation. In a designed 240-context study, two spectral directions remove at least 80.8% of the feasible worst-case bias. For a separate misspecified logistic forecaster, matched sensitivity and generic corrections reduce expected cost by 2.590% and 3.239%, with the generic method winning all 200 refits. On 100,000 new episodes, both corrections certify absolute evaluation bias below the illustrative tolerance 0.04 throughout a declared shift interval, while the base certifies no shift. This outcome is highly predictable under the known synthetic law used for design. The probability-feasible certificate does not improve on the moment rectangle in this sample. The contribution is reusable evaluation and audit capability with explicit geometric and statistical limits, rather than recalibration superiority. Evidence uses synthetic finite contexts, known frequencies and committed plans; evaluation bias, oracle regret, constraint risk and feedback deployment remain distinct. Comments: 66 pages, 7 figures Subjects: Systems and Control (eess.SY); Machine Learning (cs.LG) Cite as: arXiv:2611.21203 [eess.SY] (or arXiv:2611.21203v1 [eess.SY] for this version) https://doi.org/10.48550/arXiv.2611.21203 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-22] Determining Concentration from Raman Spectroscopy with Manifold Learning and Y-shaped Conformal Autoencoders

链接: https://arxiv.org/abs/2611.21196
作者: Anna Taka,Eleni D. Koronaki,Ioannis G. Kevrekidis,Eike Cramer,Joern Viell,Alexander Mitsos
类目: Machine Learning (cs.LG)
*备注: Manuscript: 32 pages, 6 figures, 4 tables. Supplementary Information: 22 pages, 20 figures. Submitted to the Journal of Raman Spectroscopy

点击查看摘要

Abstract:Monitoring of chemical processes is crucial in chemical engineering, motivating sophisticated process analytical technology (PAT). In-line analytics are advantageous, with Raman spectroscopy being an established and reliable tool in various applications, from which many properties can be determined. We recently [Koronaki et al., AIChEJ 2024] proposed determining polymer particle size from Raman spectroscopic data, by combining Diffusion Maps (a manifold learning method) with a Y-shaped Conformal Autoencoder Neural Network. Herein, we extend this workflow to predict liquid concentrations, allowing also for multiple target predictions from the same model. We apply the methods on two datasets consisting of Raman measurements and concentration measurements: one from microgel polymerization and one from citric acid dissociation. The single-target concentration prediction results, where one target is determined from each model at a time, are accurate. Comparing against dedicated methods such as Indirect Hard Modeling (IHM) is not straightforward, due to the different validation strategies and the considered spectra range; nevertheless, it is fair to say that IHM performs slightly better than the proposed workflow. We further compare this approach with other regression methods, such as artificial neural networks, extreme gradient boosting and partial least squares, and the results show that the Conformal Autoencoders achieve higher predictive performance, demonstrating that the workflow can also be used for concentration determination. Finally we find that single-target concentration predictions are more accurate than the multiple target predictions for both case studies.

[LG-23] When Feature Representation Changes Local Explanations: A Controlled Comparison of SHAP and LIME under Numerical Discretization

链接: https://arxiv.org/abs/2611.21159
作者: Parv Patel(Indian Institute of Technology Palakkad)
类目: Machine Learning (cs.LG)
*备注: 20 pages, 7 figures, 6 tables. Code: this https URL

点击查看摘要

Abstract:Local feature attributions are computed over a feature representation chosen by an upstream data-engineering step, yet they are read as statements about the underlying quantity. We ask how discretizing one continuous feature changes local KernelSHAP and LIME explanations when the classifier is held fixed, and whether the two explainers respond differently. A random forest is trained on UCI Adult and frozen; age is collapsed onto equi-width or quantile bin representatives (k in 3,5,10) before the model receives it, and the same 120 held-out instances are explained in every condition across multiple seeds. We show analytically and numerically that for KernelSHAP this is exactly equivalent to explaining the unchanged model at the quantized instance against a quantized background, so the comparison is between two effective inputs rather than a pure explainer effect; for LIME no such equivalence holds. Each attribution change is compared with a seed-noise floor from re-running the explainer on the unchanged condition. KernelSHAP’s age attribution exceeds its noise floor in all 6 pre-specified conditions (median ratios 2.4-59, all significant after Holm correction); on the raw scale LIME’s exceeds its floor in only 2 of 6 and in the rest is no larger than its own run-to-run variation. SHAP’s normalized change exceeds LIME’s in every condition (rank-biserial 0.81-1.00), while LIME is 2.5-4.6x less consistent across repeats. Predicted classes change for only 0.8-5.8% of instances and held-out ROC-AUC never falls below 0.9097, yet 45-94% of instances show a large SHAP attribution change with no class flip. Feature-importance rankings barely move (median Spearman = 0.979, top-3 overlap 1.00). The representation change and the change in effective input cannot be separated within a frozen-model design.

[LG-24] Cap-Aware Boosting for Green Day-Ahead Electricity Price Forecasting

链接: https://arxiv.org/abs/2611.21149
作者: Shubham Pratap
类目: Machine Learning (cs.LG)
*备注: 19 pages, 7 figures. LightGBM forecasts of 15-minute IEX Green Day-Ahead prices, compared with Chronos-2

点击查看摘要

Abstract:Accurate forecasting of Green Day-Ahead Market (GDAM) electricity prices is essential for market participants in the Indian power sector. We present a LightGBM-based forecasting framework for Indian Energy Exchange GDAM prices, which clear in 96 fifteen-minute intervals per day. The proposed system explicitly models price-cap behavior by separating capped-price events from normal market conditions. A classification head estimates the probability of the regulatory price cap (Rs 10,000/MWh), while a regression head predicts uncapped prices in logarithmic space. Final forecasts are obtained through probabilistic blending of these components and are accompanied by nine quantile forecasts for uncertainty estimation. Evaluated on 257 trading days between 1 January and 14 September 2026, the model achieves a mean absolute error (MAE) of 397.9 Rs/MWh and a mean absolute percentage error (MAPE) of 9.57%, substantially outperforming a Chronos-2 baseline (583.3 MAE) and a persistence baseline using the previous day’s profile (589.1 MAE). We further introduce leak-safe variants that incorporate real-time market signals, unmatched sell volumes, and solar generation proxies. In a recent out-of-sample evaluation, the quantile-clock variant achieved the best short-term performance with an MAE of 789 Rs/MWh during 25-28 October, while the enhanced capped-price model achieved an MAE of 716 Rs/MWh across 21 evaluation days through 16 November. These results demonstrate the effectiveness of regime-aware probabilistic forecasting for high-volatility electricity markets.

[LG-25] Non-Myopic Safety Filtering with Learned Intervention Cost-to-Go

链接: https://arxiv.org/abs/2611.21036
作者: Johannes Autenrieb,Mark Spiller,Daniel Silvestre,Namhoon Cho,Jason J. Choi
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注: 19 pages, 5 figures

点击查看摘要

Abstract:Control barrier function (CBF) safety filters minimize instantaneous deviation from a nominal control command, but a sequence of pointwise-minimal interventions need not minimize cumulative intervention effort. This paper proposes a discrete-time non-myopic safety filter using a learned approximation of the optimal intervention cost-to-go over CBF-admissible policies. The approximation enters through a relaxed Bellman-residual constraint, while the original barrier condition remains hard. Feasibility is equivalent to that of the conventional CBF filter, and forward invariance remains independent of approximation accuracy under the stated assumptions. We derive a trajectory-wise performance certificate that requires neither the optimal value function nor an approximation-error bound. Under additional approximation and horizon assumptions, a suboptimality bound contains an explicit endpoint approximation-error term together with accumulated tolerance and relaxation terms. Numerical results for a collision-avoidance problem show an approximately 50% reduction in mean cumulative intervention effort relative to conventional CBF filtering over 100 initial conditions not used during training or parameter selection. The improvement is consistent across three independently trained networks, while maintaining the prescribed safety constraints and requiring less online computation than the tested predictive filter.

[LG-26] A physics-constrained emulator for time-discretised landslide runout modelling

链接: https://arxiv.org/abs/2611.21015
作者: Ye Chen,Lorenzo Nava,Maximillian Van Wyk de Vries,Fawu Wang
类目: Machine Learning (cs.LG); Geophysics (physics.geo-ph)
*备注:

点击查看摘要

Abstract:Numerical simulations of landslides are commonly used to forecast impact areas. Full physics-based models provide accurate predictions, but their computational expense can preclude their use for large parameter sweeps. To achieve efficient modelling for such cases, we developed a physics-constrained neural network to emulate full landslide runout in a time-discretised manner. This emulator is trained on over 200,000 samples, each consisting of a pair of landslide thicknesses and velocities on a given DEM with a fixed time interval. To constrain the model’s physical consistency, we define physics-based loss functions based on mass conservation and the system’s kinetic energy loss. The model demonstrates high fidelity across the test dataset, achieving root mean square errors of approximately 0.06 m for thickness and 0.11 m/s for velocity in single-step predictions. In autoregressive rollout, these errors evolved modestly to about 0.21 m and 0.09 m/s after 110 iterations, while greatly improving computational efficiency by more than 50 times.

[LG-27] Learning and Verifying Stabilizing Controllers with Structured Neural Control Lyapunov Functions

链接: https://arxiv.org/abs/2611.21009
作者: Jose Luis Ledesma Martin,Patricia Pauli
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:This work investigates the joint learning of neural controllers and control Lyapunov functions (CLFs) for nonlinear discrete-time systems. We exploit the structural properties of Input Convex Neural Networks and Polyak-Łojasiewicz Networks to guarantee positive definiteness of CLF candidates by construction, reducing the conditions requiring formal verification. We develop a mixed-integer programming framework to formally verify the remaining stability conditions and introduce an iterative region of attraction (ROA) expansion procedure to enlarge the certified ROA without retraining. Experiments on an inverted pendulum and a path-tracking system demonstrate the advantages of structured neural CLFs, particularly PLNets, and illustrate how their structural bounds influence the geometry of the certified regions. Our code is available at: this https URL.

[LG-28] FH-Sampling: Frequency-aware Hypergraph-driven Oversampling for Imbalanced Time-Series Classification

链接: https://arxiv.org/abs/2611.20991
作者: Yunrong Hao,Ruohao Dai,Zhen Liu,Guan Wang,Wenbin Pei,Qiang Zhang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Imbalanced time-series data appears commonly in diverse applications, such as medical diagnosis. When trained on imbalanced data, classifiers tend to exhibit a bias toward the majority class, making it difficult to identify minority-class instances accurately, despite their critical importance. Sampling techniques have proven effective in rebalancing imbalanced data. However, existing interpolation-based oversampling methods typically rely on pairwise interpolation, limiting their ability to capture high-order inter-sample relationships, and their inherent smoothing effects tend to suppress high-frequency components, compromising spectral fidelity. Therefore, we propose a Frequency-aware Hypergraph-driven Oversampling method (FH-Sampling) for imbalanced time-series classification. In FH-Sampling, a hypergraph is first constructed in the original time-series space to capture complex high-order inter-sample relationships, and sampling ratios are adaptively assigned to mitigate the impact of noisy samples. The discrete wavelet transform decomposes each sample into frequency components. Guided by the sample hypergraph, FH-Sampling combines high-frequency components from related samples while retaining the target sample’s low-frequency component to help preserve important frequency characteristics. Experiments on 24 imbalanced datasets demonstrate the advantages of FH-Sampling for imbalanced time-series classification. We further apply FH-Sampling to aircraft trajectory generation. Experiments on simulated aircraft trajectories show that FH-Sampling produces trajectories that better match the real distribution, highlighting its potential for practical applications.

[LG-29] VASR: Vector-Aware Assembly Planning with Equivalent Structures and Repair-Cost Feedback

链接: https://arxiv.org/abs/2611.20984
作者: Ruizhe Wang,Rui Liu,Hongliang Chen,Junyi Guo,Fangyu Wu,Yaran Chen
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Furniture assembly planning requires inferring subassembly hierarchies from individual parts and a target illustration. Existing raster-based methods do not explicitly exploit the geometric and structural information in SVG drawings. We introduce VASR, a framework that fuses features from a dedicated SVG encoder with visual representations from a pretrained vision-language model to predict assembly hierarchies. To avoid penalizing valid alternatives to a reference assembly tree, we propose Repair-Aware Hierarchical Policy Optimization (RHPO) with hierarchy-preserving equivalence rewards. RHPO further uses repair-cost feedback to favor decisions with lower downstream correction costs among candidates with comparable equivalence-aware rewards. VASR achieves the highest Equivalent-Hard F1 among the evaluated methods, outperforming GPT6-Astra by 17.69 percentage points on the IKEA-derived test set while maintaining strong performance on an independent furniture dataset.

[LG-30] STORMGUIDE-FL: Stochastic Remote Guidance for Efficient Mining of Federated Streaming Data

链接: https://arxiv.org/abs/2611.20975
作者: Manuel Röder,Frank-Michael Schleif
类目: Machine Learning (cs.LG); Databases (cs.DB); Distributed, Parallel, and Cluster Computing (cs.DC)
*备注: 21 pages, 7 figures, 4 tables

点击查看摘要

Abstract:Federated Learning enables collaborative model training across distributed edge devices (“clients”) while preserving data privacy. However, real-world deployments involve clients that observe continuous data streams from evolving environments while operating under strict resource limitations. Existing solutions still face two major challenges: local concept drift, where each client’s data distribution shifts over time, and global client drift, arising from asynchronous engagement and imbalanced contributions across heterogeneous training participants, formally described as dual client heterogeneity. We propose STORMGUIDE-FL, a federated optimization and data-mining framework that mitigates both types of drift using a server-guided pipeline for data acquisition and local training. For data acquisition, a lightweight client-side sampling strategy incrementally selects informative and diverse instances from non-stationary, non-IID local streams under bounded memory, compute, and labeling budget. For local training, the server’s remote guidance is embedded directly into local optimization as an update-level correction, counteracting client drift without discarding scarce labeled samples. We empirically verify the effectiveness of the proposed sampling strategy on federated data streams across gradual, sudden, and recurring concept drift, and demonstrate that its gains compound with the update-level drift correction. Experimental results further show improved speed and robustness of global model convergence under resource-constrained, heterogeneous conditions.

[LG-31] Same Fix Different Weights: Replicating Weight-Space Programming with Claude

链接: https://arxiv.org/abs/2611.20953
作者: Subham Kundu(TCG Crest)
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Takehi et al. (2026) report that GPT-6 Astra, given the weights of small neural networks as text, can say what they were trained to do, repair them, and write new ones without training. I re-ran the core of that study with Claude Opus 5.5 in ten pre-registered experiments, with untrained-network and permutation controls and hand audits of every sandboxed code cell. Reading was mixed: Claude named 3 of 4 Atari games from policy weights and found the measured copying heads in 15/15 character-LM runs, but its answers about training data often followed priors. Unlike the original, it identified a sycophantic Gemma 3 270M finetune from weights alone in 6/6 runs. One-shot weight edits took a failing MountainCar policy from 0% to 100% success and a hovering LunarLander policy from 0% to 85.6%; zero-shot networks beat PPO at 100k steps on 4 of 5 environments, each implementing a textbook controller. With no training or forward passes, Claude wrote a 1.48M-parameter 9x9 Go network that won 85-94% of games against three PPO opponents. Given Qwen3.5-0.8B’s weights and 10, 25 or 50 forward passes of 64 practice problems, Claude raised MATH-500 from 6.7% to 51.4%, 55.2% and 58.1% (+51.4 points at 50 passes, 95% CI [47.9, 55.0]; original 48.1%), GSM8K from 11.6% to 68.2% and IFEval from 22.2% to 54.2%. Unlike the original, which removed an internal loop direction, every edit acted at the output: it scaled the final norm, raising the effective sampling temperature, and lowered the logit of “Wait” while raising that of /think. The norm scaling alone reaches only 20.9-22.1% on MATH-500. On Gemma 3 270M, which rarely loops, no edit helped (-0.8 points), as in the original.

[LG-32] Joint Learning of Control Lyapunov Functions and Controllers via Stable-by-Design Surrogates

链接: https://arxiv.org/abs/2611.20900
作者: Patricia Pauli,Matthieu Barreau
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Training and verifying neural network-based control Lyapunov functions (CLFs) can be computationally demanding. To address this challenge, we propose a framework that jointly learns a CLF and a stabilizing controller using structured neural networks and a projection-based parameterization. Specifically, we employ Polyak-Lojasiewicz networks (PLNets), which parameterize nonconvex, positive-definite CLF candidates, and construct a stable-by-design surrogate of the closed-loop dynamics that satisfies the Lyapunov decrease condition by construction. The controller combines linear state feedback with a higher-order correction term resembling Sontag’s universal formula and is trained to align the resulting closed-loop dynamics with the stable-by-design surrogate. Unlike classical approaches based on Sontag’s formula, our framework does not require a known CLF; instead, it learns the CLF and controller simultaneously. We establish local and global stability guarantees for the learned controller and demonstrate the effectiveness of the proposed approach through numerical examples.

[LG-33] Comfort as a Constraint or as a Cost in Reinforcement Learning and Model Predictive Building Control

链接: https://arxiv.org/abs/2611.20889
作者: Joram Liebeskind,Valentin Bhend,Varsha N. Behrunani,Alisa Rupenyan
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:A building controller that keeps a zone within comfort bounds at minimum energy can treat the bounds as a soft constraint, by adding the weighted violation to its cost, or as a hard constraint. We analyze the effect of the hard and the soft formulation of the same problem and isolate the effect of the formulation from the effect of the algorithm. To this end, we measure the difference in energy use between the two formulations on the BOPTEST bestest_hydronic_heat_pump case, in model predictive control (MPC) and in reinforcement learning (RL). The soft formulations are a weighted MPC and weighted SAC and DDPG learners, each at twelve weights. The hard formulations are an MPC with an exact penalty and the Lagrangian learners SAC-Lagrangian and primal-dual DDPG. We find that the hard formulations use more energy than the soft formulations at the largest weight, but violate the comfort bounds less. For a large weight, the soft formulation approaches the behavior of the hard formulation. The frontiers of the energy-comfort trade-off of the three weighted controllers lie close to each other, and the weight selects an operating point on this frontier. Under the soft formulation, the three algorithms reach very similar operating points, although at somewhat different weights. For this HVAC problem, the formulation and the weight thus affect the energy use more than the choice of algorithm.

[LG-34] Markov-Equivalent Conditional Independence Graphs on Path Spaces

链接: https://arxiv.org/abs/2611.20877
作者: Simon Kempf,Niki Kilbertus,Georg Manten
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Causal discovery for continuous-time stochastic processes with latent confounding typically faces a trade-off between (i) approaches that are highly informative about the causal structure but are difficult to test statistically, and (ii) approaches that admit practical tests from data but yield more coarse-grained conclusions, potentially not leveraging the direction of time. We introduce W-separation, an asymmetric graphical criterion for directed mixed graphs defined as a quaternary relation that allows distinct conditioning on past and future components. This added flexibility yields a highly informative graphical independence model while preserving testability. W-separation satisfies a global Markov property for a broad class of stochastic differential equations with latent confounding, linking it to conditional independence relations among coordinate processes that are testable from data. A central result of this work is the full characterization of W-Markov equivalence classes: we show that every equivalence class admits a unique greatest element that serves as a canonical representative and a principled target for learning algorithms. In addition, we provide necessary and sufficient conditions under which directed or bidirected edges can be added without altering the induced independence model.

[LG-35] Practical Semantic Communication for General-Purpose Networks: Compression of High-Dimensional Embeddings Under Quantization

链接: https://arxiv.org/abs/2611.20863
作者: Leonardo Roque,Louis-Adrien Dufrène,Quentin Lampin,Grégoire Lefebvre,Mohamad Assaad
类目: Information Theory (cs.IT); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:This paper introduces a practical framework for semantic communication over wireless channels, focusing on the compression of high-dimensional embeddings under quantization and channel errors. Building upon a relaxed Gaussian Information Bottleneck (GIB) model, the approach provides analytical solutions for task-relevant information preservation, even with non-Gaussian targets. The framework relaxes traditional Gaussian assumptions, enabling effective compression for discrete and complex data such as text embeddings. It derives closed-form gradients for optimizing the compression matrix and establishes bounds linking continuous representations to discrete, noisy system performance. Comprehensive experiments on text classification tasks demonstrate that the proposed method outperforms task-agnostic quantization-based compression methods, maintaining high semantic fidelity at low bit budgets per embedding. The approach offers a scalable, interpretable, and resource-efficient solution for deploying semantic communication in general-purpose networks.

[LG-36] Motor Imagery Classification Using Hybrid Cross-Subject Class Mixup and Selective Attention

链接: https://arxiv.org/abs/2611.20853
作者: Jiwon Kim,Chae-Won Lee,Jong-Seok Lee
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Motor imagery (MI) enables a brain-computer interface (BCI) to decode movement intentions without overt movement. Hybrid electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) classification combines complementary signals, but inter-subject variability and the design of information exchange between modalities affect generalization to new subjects. We propose a framework that combines Cross-Subject Class Mixup (CSCM) with Selective Attention. CSCM aims to construct new training data to make the model robust against inter-subject variability. It independently selects EEG and fNIRS donor trials from other training subjects within the anchor class and mixes each modality while preserving the original label. Selective Attention provides an effective way to simultaneously model the temporal signal characteristics and fuse the two modalities. It retains causal temporal self-attention for EEG and uses a classification token (CLS) to integrate both modalities, without fNIRS attention updates or direct cross-modal temporal-token attention. Leave-one-subject-out (LOSO) evaluation across 28 subjects on a public EEG-fNIRS dataset yields a mean balanced accuracy (BA) of 76.85% and a macro-averaged F1 score (Macro-F1) of 76.58%. The proposed model achieves the highest mean performance among five reimplemented baselines. We also examine the contributions of CSCM and Selective Attention through modality-specific mixing and donor-policy comparisons, temporal-policy comparisons, interventions on direct cross-modal paths, and component ablations.

[LG-37] artist: A Python package for differentiable ray tracing and gradient-based optimization of solar tower power plant operation

链接: https://arxiv.org/abs/2611.20851
作者: Marlene Busch(1),Marie Weiel(2 and 3),Kaleb Phipps(2 and 3),Daniel Maldonado Quinto(1),Max Pargmann(1),Robert Pitz-Paal(1 and 4),Markus Götz(2 and 3) ((1) Institute of Solar Research, German Aerospace Center (DLR), Cologne, Germany, (2) Scientific Computing Center (SCC), Karlsruhe Institute of Technology (KIT), Eggenstein-Leopoldshafen, Germany, (3) Helmholtz AI, Karlsruhe, Germany, (4) Chair of Solar Technology, RWTH Aachen University, Aachen, Germany)
类目: Machine Learning (cs.LG)
*备注: 14 pages, 4 figures, abridged abstract, submitted to Machine Learning: Engineering (IOP Publishing)

点击查看摘要

Abstract:Solar tower power plants provide dispatchable, renewable energy by concentrating sunlight with thousands of mirrors, called heliostats, onto a central receiver. Efficient operation requires accurate modeling of the optical path from each heliostat to the receiver and compensation for manufacturing imperfections, kinematic misalignment, and material degradation that distort the flux distribution. Conventional ray tracers predict the flux from an assumed plant model but are limited to forward simulation. They cannot recover unknown plant parameters from measured flux nor directly optimize plant operation. We present artist, an open-source, FAIR4RS-compliant Python package implementing a fully differentiable digital twin of solar tower power plants. By making the optical pipeline, comprising heliostat surface representation, kinematics, blocking, and ray tracing, end to end differentiable, artist turns the digital twin from a passive forward model into an active physics-based tool for inverse modeling and data-driven optimization. The resulting gradient-based framework unifies tasks usually handled by separate solutions, such as reconstructing heliostat surfaces and kinematics from operational flux images without costly deflectometry, predicting irradiance at the receiver, and optimizing aim points to realize a desired receiver flux density distribution. Built on PyTorch, artist scales to real plants through GPU acceleration, a structure-of-arrays memory layout, and nested data parallelism. It interfaces with open data sources such as the PAINT database. We describe the software architecture, physical models, and optimization framework, and demonstrate the workflow on representative tasks. By integrating modeling, prediction, and optimization into a single open, interpretable package, artist provides a principled foundation for data-driven research toward autonomous solar tower operation.

[LG-38] Hierarchical Regime-Conditioned Dynamics for Spatiotemporal Graphs

链接: https://arxiv.org/abs/2611.20835
作者: Jeremy Baffou,Pascal Frossard,Dorina Thanou
类目: Machine Learning (cs.LG); Computational Geometry (cs.CG)
*备注:

点击查看摘要

Abstract:Forecasting the behavior of spatio-temporal systems often requires more than accurate predictions: models should uncover the dynamical modes governing future evolution. Existing temporal graph and latent dynamical models achieve strong forecasting performance, but their latent representations often entangle multiple dynamical patterns, limiting interpretability. Conversely, regime-inference methods explicitly model such modes but typically do not support generative forecasting over graph-structured systems. We introduce HERON, a self-supervised regime-aware generative framework that unifies graph-structured trajectory generation with regime identification. HERON factorizes the latent state into discrete system-level regimes, capturing global dynamical modes, and continuous entity-level dynamics, yielding a modular regime interface compatible with existing spatio-temporal backbones. Crucially, the inferred regime actively conditions system dynamics, enabling regime tracking and controlled simulations under alternative dynamical modes. Without regime annotations, HERON jointly learns forecasts and regime dynamics in a self-supervised manner, recovering meaningful modes across controlled synthetic and real-world datasets while maintaining competitive forecasting performance. These results open a path toward interpretable, controllable generative modeling of spatio-temporal dynamics.

[LG-39] ght PAC-Bayes Generalisation Guarantees for Large Language Model Judges

链接: https://arxiv.org/abs/2611.20818
作者: Tom A. Lamb,Philip H. S. Torr,Tim G. J. Rudner
类目: Machine Learning (cs.LG)
*备注: 43 pages, 12 figures

点击查看摘要

Abstract:Large language model (LLM) judges are increasingly used to monitor the outputs of LLM-based AI agents, yet formal guarantees on the reliability of these judges remain largely absent, limiting confidence in their use for safety-critical oversight. We consider certifying such a judge and define its risk over safety classifications of prompt–response pairs drawn from a deployed generator. We apply a PAC-Bayes framework to this setting, refining existing compression-based methodology with the use of tighter bounds and more efficient coding. We show that standard parameter-efficient fine-tuning can yield non-vacuous and even tight certificates for safety judges, for both classification and Brier risk, without holding out data for validation. We introduce LoRA-GT, which shares a single quantisation scale across layers and achieves the tightest certificates with minimal degradation in predictive performance. Across five guardrail models and four benchmarks, we find that certificate tightness is governed more by adaptation complexity than by base model scale. Finally, we show how far quantisation can be pushed before the resulting increase in risk outweighs the bound tightening gained from shorter descriptions. Together, these results provide a practical route toward certifiable scalable oversight and offer insight into how the complexity of safety adaptations shapes generalisation.

[LG-40] STA: A Split Temporal Adversary Method for Evaluating Robustness of Multimodal RL-based Driving Policies

链接: https://arxiv.org/abs/2611.20765
作者: Shenghao Xiong,Xiaolin Chang
类目: Cryptography and Security (cs.CR); Machine Learning (cs.LG); Robotics (cs.RO)
*备注: 10 pages, 6 figures, submitted for publication

点击查看摘要

Abstract:Multimodal reinforcement learning (MRL) offers a promising paradigm for autonomous driving by jointly leveraging complementary sensor observations to learn effective decision-making policies. However, its effectiveness critically depends on temporal alignment across modalities, making temporal inconsistency a fundamental vulnerability. Even subtle misalignment or latency can corrupt fused representations and ultimately lead to decision failures. Existing studies have explored adversarial attacks to induce cross-modal temporal inconsistencies, thereby enabling systematic evaluation of the temporal robustness of MRL policies. However, existing methods remain limited in generality, realism, and adaptivity. To address these limitations, we propose the Split Temporal Adversary (STA), which formulates multimodal temporal attacks as a unified decision process that jointly optimizes attack timing and modality selection under realistic policy-aware and budget-constrained settings. STA includes two mechanisms: a) a unified multimodal temporal perturbation mechanism that delays observations via historical states, removing the need for modality-specific noise; b) a split multimodal adversarial decision mechanism that jointly optimizes attack timing and modality selection through adaptive optimization while leveraging the victim policy’s clean actions to identify sensitive decision patterns. We provide a theoretical analysis of the induced distribution shift and a modality selection principle. Experiments in the CARLA simulator show that STA more than doubles the collision rate of the strongest baseline against PPO while using 17.8% fewer attack attempts, and that it generalizes to a VLM-based driving policy. These results reveal temporal consistency as a critical vulnerability, highlighting temporal robustness as essential for safe autonomous driving.

[LG-41] Who Benefits from Telemonitoring? Uplift Modelling of a Nationwide Home Blood Pressure Monitoring Programme

链接: https://arxiv.org/abs/2611.20752
作者: Connie Kou
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Home blood pressure (BP) telemonitoring programmes are being deployed at national scale, yet enrolment resources are finite and the benefit of telemonitoring is unlikely to be uniform across patients. We study the Primary Technology-Enhanced Care for Hypertension (PTEC-HT) programme, a nationwide telemonitoring programme for hypertensive patients in Singapore. Using retrospective data on 16,694 patients registered between August 2020 and May 2024, we first estimate the average treatment effect of the programme on 6-month BP control against a propensity-score-matched non-participant cohort, finding an absolute improvement of 5 percentage points in the rate of BP control (74% vs. 69%, p = 3\times10^-12 ), rising to 12 percentage points (63% vs. 50%, p = 1\times10^-15 ) among patients with uncontrolled BP at baseline. We then move beyond the average effect and apply uplift modelling to estimate individual treatment effects. Across five uplift learners evaluated with 5-fold cross-validation and Qini curves, an uplift random forest performed best on the full cohort and an S-learner on the uncontrolled-baseline subcohort, although performance bands overlapped across learners. On held-out patients, realised uplift was largest in the top predicted-uplift groups and declined across the ranking. These patients had higher baseline BP grade, more recent hypertension diagnosis and lower medication and diagnosis burden, a profile we offer as hypothesis-generating for targeted enrolment. A handcrafted rule built from these features was competitive with the learned models, outperforming all of them on the uncontrolled-baseline cohort, suggesting that uplift models are useful for discovering heterogeneity that can be distilled into simple enrolment criteria for directing limited care team capacity.

[LG-42] Calibrate Once Grow the Vocabulary: Candidate-Independent Scores for Conformal Risk Control in Open-Vocabulary Segmentation

链接: https://arxiv.org/abs/2611.20715
作者: Md Tanveer Hossain Munim,Bijoy Ahmed Saiem,Muhammad Abdullah Adnan
类目: Machine Learning (cs.LG)
*备注: 44 pages: 8 pages main text, 4 pages references, 32 pages supplementary material. 8 figures, 52 tables. Code: this https URL

点击查看摘要

Abstract:Open-vocabulary segmenters accept class names at inference. Conformal risk control can bound how often their prediction sets miss the true label, but we show that this bound breaks when names are added after calibration. On ADE20K, a threshold calibrated for 10% risk at 150 names reaches up to 15.4% risk at 845 names in released CAT-Seg models. The risk exceeds the target on every calibration split, yet the prediction sets grow far less than the target requires. In SED they instead grow larger than needed. The cause is that pruning, cross-label attention and vocabulary normalisation make each label’s score depend on the other requested labels. Scores can move either way, so a threshold recalibrated at a larger vocabulary need not hold at smaller ones. We remove this dependence without training by scoring each label against a fixed reference bank of names. A label’s score then stays the same when names are added, so one calibration keeps its guarantee as the vocabulary grows. The guarantee holds whatever names are added, even ones chosen per image after seeing it. On three benchmarks and five segmenters, the converted models meet the target at every vocabulary size we evaluate. For CAT-Seg and SED at alpha=10%, they use at most one more label per pixel than the released models recalibrated at the grown vocabulary. Treating names instead of images as calibration samples also bounds the miss rate of added names. If the annotated names are a random sample of the names in the calibration images, an added name’s expected miss rate on those images is at most the target. For harder added names, the excess is at most their gap in mean miss rate. Applied to Grounding DINO, which splits large vocabularies into chunks, the conversion also makes inference about six times faster.

[LG-43] MSAR: Next-Scale Autoregressive Forecasting for Time Series via Modular Multi-Scale Decoupling NEURIPS2026

链接: https://arxiv.org/abs/2611.20714
作者: Fanda Fan,Kuiye Ding,Liming Mao,Bingsong Wang,Yao Wang,Xiaorui Wang,Ruijie Jian,Zhipeng Liu,Luqi Gong,Zhenghua Lu,Chunjie Luo,Jianfeng Zhan
类目: Machine Learning (cs.LG)
*备注: Accepted by NeurIPS 2026 main track. 26 pages, 7 figures. Code available at this https URL

点击查看摘要

Abstract:Time series forecasting underpins critical applications in finance, energy, healthcare, and transportation. Although deep models have achieved strong results, most adopt single-scale modeling or restrict multiscale processing to the input side, causing a misalignment between multiscale inputs and single-scale outputs and limiting predictive power. We introduce the Modular Scale-wise Autoregressive Framework (MSAR), a model-agnostic design that forecasts progressively across multiple temporal resolutions. MSAR offers three advantages: (1) scale-wise aligned modeling, which disentangles heterogeneous temporal patterns by aligning inputs and outputs at each scale; (2) scale-wise autoregression, where coarse-scale predictions guide finer-scale forecasting through hierarchical information flow; and (3) a modular architecture, enabling seamless integration with diverse backbones such as CNNs, MLPs, and Transformers. Extensive experiments across a broad set of datasets and forecasting models demonstrate that MSAR achieves consistent improvements in both accuracy and inference efficiency, validating the effectiveness of scale-aligned autoregression for multiscale time series forecasting. The code is available at: this https URL.

[LG-44] Conic Structure of the Cross-Entropy Loss at the Final Layer and Its Absence in Rectifier Backbones

链接: https://arxiv.org/abs/2611.20696
作者: Omer Nezih Gerek
类目: Machine Learning (cs.LG); Signal Processing (eess.SP)
*备注: 14 pages, 2 figures, to be submitted

点击查看摘要

Abstract:Slices of the loss along lines and planes in parameter space are widely used to visualize deep networks. They often show a single basin with nearly straight walls, which raises the question of whether such a loss is a cone rounded at its apex, or a lower envelope of several cones that meet along creases. We answer this question for softmax cross-entropy at two levels of a classifier. Over the weights of an affine final layer, the loss equals a polyhedral cone plus an offset that decreases monotonically and exponentially along every ray, towards a limit in closed form. The distance at which the asymptotic wall reaches the height of the starting point, which we call the core radius, equals the mean surprisal of the classes that the direction favours, divided by the slope of the wall. Over the parameters of a rectifier backbone, the loss along a line has a kink wherever the activation pattern of a sample changes. Using the dimension theory of semialgebraic sets, we show that, along almost every line, the samples that switch at one point share an irreducible factor of their pre-activation polynomials. Under a coprimality condition, each kink therefore carries the share of one sample among n , so creases comparable to the slope do not form. Measurements on ResNet-18 and ResNet-56 trained on CIFAR-10 illustrate both results, with a test that bounds the size of any crease that sampled loss values can hide.

[LG-45] Rethinking Supervision Scaling in On-Policy Distillation: From Dense Supervision to Stochastic Horizons

链接: https://arxiv.org/abs/2611.20687
作者: Xinhao Yao,Zhenxing Dou,Yong Liu
类目: Machine Learning (cs.LG)
*备注: Work in progress

点击查看摘要

Abstract:On-policy distillation (OPD) provides dense token-level supervision along student-generated trajectories, yet the learning benefits of longer trajectories may not scale proportionally with their computational cost. We investigate how supervision utility evolves with trajectory length. Controlled experiments reveal diminishing returns from additional supervised tokens, while showing that later positions remain learnable and exhibit substantial cross-position supervision transfer. Gradient analysis further identifies shared update components across response positions and diverse position-specific residuals whose cross-rollout consistency decreases at greater depths. Motivated by these findings, we propose Stochastic Horizons On-Policy Distillation (SH-OPD), which reduces redundant supervision by randomly varying rollout horizons while preserving access to deeper response states. SH-OPD samples a shared horizon for each training batch, enabling efficient parallel generation without per-sequence horizon scheduling overhead. We characterize the resulting reweighting bias and derive a compute-reweighting trade-off under standard optimization assumptions. Experiments on mathematical reasoning and code generation demonstrate that SH-OPD achieves slightly higher mathematical reasoning accuracy at approximately 45% of the per-update cost of standard OPD, while maintaining competitive code generation performance at around 60% of the baseline cost. These results highlight the benefits of efficient supervision allocation across response depths.

[LG-46] Conditional Diffusion Models for Amortized Bayesian System Identification

链接: https://arxiv.org/abs/2611.20666
作者: Mike Hänseroth,Niklas Wahlström,Thomas B. Schön,Thomas Seel,Jan-Hendrik Ewering
类目: Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注: 8 pages, 7 figures

点击查看摘要

Abstract:Generative diffusion models rapidly solve Bayesian inverse problems, such as image restoration and inpainting, with a single trained model and without retraining for each new observation or task. Bayesian system identification, which typically involves inferring the hidden states and parameters of a dynamical system from measurements, poses a closely related inverse problem. Yet it is often addressed with computationally heavy Monte Carlo methods that must be rerun for every new measurement record. Moreover, these methods are restricted to a single inference task, and their cost grows with the state dimension and the record length. To alleviate these restrictions, we interpret Bayesian system identification in hiddem Markov models (HMMs) as a general Bayesian inverse problem, taking advantage of conditional diffusion models. Specifically, a transformer-based diffusion model is trained once on short segments of simulated trajectories to approximate the joint posterior of states and parameters. Through an attention mask induced by the Markov property and a relative time encoding, it can then be deployed on any new measurement record of custom length. At inference time, the same model can flexibly and rapidly be conditioned for joint state and parameter inference, state smoothing, parameter inference, or reconstructing missing measurements, without retraining. We demonstrate these capabilities on four benchmark systems and examine how the length of the training segments affects state and parameter inference. On records five times longer than the training segments, the model maintains its state accuracy, whereas comparison methods either show a more than tenfold error increase or require ten times as much sampling time.

[LG-47] ParGLU: Partial Gating Improves Language Modeling in MoE Transformers

链接: https://arxiv.org/abs/2611.20663
作者: Yuting Ge
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Multiplicative gating improves Transformer feed-forward networks (FFNs), but it reduces intermediate representation width under a fixed parameter budget. This trade-off is particularly relevant to Mixture-of-Experts (MoE) models, where individual experts have limited capacity. We introduce ParGLU, a partially gated FFN that combines SwiGLU and ReLU ^2 in two parallel branches. ParGLU matches the parameter count and principal matrix computation budget of conventional SwiGLU. We evaluate ParGLU in 30M, 75M, and 152M MoE language models using three random seeds per scale. ParGLU outperforms SwiGLU across all three seeds at 75M and 152M, reducing mean validation NLL by 0.0200 and 0.0249, respectively. Results at 30M are mixed. Controlled ablations show that increasing backbone width or expert width strengthens ParGLU’s relative performance. Ablation studies further show that combining SwiGLU and ReLU ^2 improves performance over either architecture alone. Token-frequency analysis reveals larger improvements on low-frequency tokens. These findings show that fully gated FFNs are not always the most effective use of an MoE expert’s parameter budget. Partial gating can improve language modeling in MoE Transformers.

[LG-48] Enhancing Mixed Oil Length Prediction through Graph Representation Learning

链接: https://arxiv.org/abs/2611.20660
作者: Xinyi Hong,Hai Shu,Zhaoyang Zhang,Lei Chen,Yanfeng Yang,Ziqi Chen
类目: Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注: 21 pages, 6 figures, 7 tables

点击查看摘要

Abstract:During the pipeline transportation of refined oil products, the inevitable formation of mixed oil segments poses considerable economic challenges to pipeline networks. Therefore, it is important to accurately predict the length of these mixed oil segments. Although previous methods have achieved relatively accurate predictions by integrating machine and deep learning algorithms, they ignore the statistical dependencies that remain among real-world mixed oil samples, which degrades model accuracy. To address this issue, we propose a novel Sample Dependencies Incorporated Graph Representation Learning (SDI-GRL) model. By representing samples and pipeline features as nodes of a graph, SDI-GRL connects dependent samples through their shared features, allowing message passing to propagate information across related samples. This graph-based framework allows our model to capture the statistical dependencies among samples absent in prior works. A case study on a mixed oil dataset demonstrates that our SDI-GRL model surpasses the strongest existing approach, reducing both the RMSE and the MAE by more than 10%.

[LG-49] Contrastive Activation-Guided Zeroth-Order Preference Optimisation for Memory-Efficient LLM Alignment NEURIPS2026

链接: https://arxiv.org/abs/2611.20635
作者: Weiyun Huang,Alessio Galatolo
类目: Machine Learning (cs.LG)
*备注: Accepted at AXIOM Workshop @ NeurIPS 2026

点击查看摘要

Abstract:Direct Preference Optimisation (DPO) simplifies reinforcement learning from human feedback, but still relies on traditional backpropagation and its associated activation-memory cost. Zeroth-order (ZO) optimisation avoids backpropagation, yet isotropic perturbations can produce noisy updates for preference objectives defined by differences between chosen and rejected responses. We introduce Contrastive Activation-guided Zeroth-Order Direct Preference Optimisation (CA-ZODPO), which constructs a low-rank search subspace from the activation discrepancy between the chosen and rejected responses in each preference pair. Perturbations are restricted to this subspace, while momentum is maintained in low-rank coordinates to preserve the memory advantages of ZO training. On Qwen2.5-1.5B with Anthropic HH-RLHF, CA-ZODPO outperforms common ZO baselines and remains competitive with first-order DPO in pairwise preference evaluation. It also reduces peak GPU memory by about 59% at a sequence length of 2048 while retaining the sequence-length scaling behaviour of zeroth-order training.

[LG-50] GPromptShield: Elevating Resilience in Graph Prompt Tuning against Adversarial Attacks

链接: https://arxiv.org/abs/2611.20621
作者: Shuhan Song,Ping Li,Ming Dun,Maolei Huang,Huawei Cao,Xiaochun Ye
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The paradigm of ``pre-training and prompt-tuning", with its effectiveness and lightweight characteristics, has rapidly spread from the language field to the graph field. Several pioneering studies have designed specialized prompt functions for diverse downstream graph tasks based on various graph pre-training strategies. These prompts concentrate on the compatibility between the pre-training pretext and downstream graph tasks, aiming to bridge the gap between them. However, designing prompts blindly to adapt to downstream tasks based on this concept neglects crucial security issues. By conducting covert attacks on downstream graph data, we find that even when the downstream task data closely matches that of the pre-training tasks, it is still feasible to generate highly misleading prompts using simple deceptive techniques. In this paper, we shift the primary focus of graph prompts from compatibility to vulnerability issues in adversarial attack scenarios. We design a highly extensible shield defense system for the prompts, which enhances their robustness from two perspectives: Direct Handling and Indirect Amplification. When downstream graph data contains unreliable biases, the former directly combats invalid information by incorporating hybrid multi-defense prompts to the input graph’s feature space, while the latter adopts a training strategy to bypass the invalid components and amplifies valid part. We provide a theoretical derivation that proves their feasibility, indicating that unbiased prompts exist under certain conditions on unreliable data. Extensive experiments across various scenarios of adversarial attacks indicate that the prompts within our defense system exhibit enhanced resilience and superiority. This paper explores a new perspective in graph prompt learning, offering a novel option for robust prompt tuning in downstream tasks.

[LG-51] DNA: A Double-Helix Hierarchical Deep Interaction Network Architecture for Industrial Digital Twin

链接: https://arxiv.org/abs/2611.20607
作者: Haoran Yang,Yanan Xu,Yinan Zhang,Wenjie Zhang,Peiyu Liu,Jiahui Xiang,Wenhai Wang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:High-fidelity modeling is fundamental to industrial digital twin, yet a trade-off persists between the physical transparency of mechanistic models and the superior predictive power of data-driven approaches. Existing hybrid methods often fail to bridge this gap, suffering from shallow integration and rigid one-shot modeling strategies that struggle with complex industrial systems. To address these issues, we propose a \textbfDouble-Helix Hierarchical Deep Interaction Network Architecture (DNA). Inspired by the double-helix structure of biological \textitDNA, our architecture intertwines mechanistic and data-driven modeling strands through two novel interaction strategies, forming a symbiotic architecture in which the two strands mutually enhance predictive capability. Furthermore, the hierarchical structure of DNA allows it to model complex systems in a multi-stage manner through two tailored deep learning frameworks. We evaluate the proposed approach on two distinct industrial systems, where DNA achieves reductions in MAE of 3.704% and 3.175%, respectively, compared to the strongest baseline. These results demonstrate the effectiveness of DNA and suggest its potential to provide new insights for the development of industrial digital twins.

[LG-52] Quantum Computing-Based Wasserstein Generative Adversarial Networks for Airfoil Parameterization

链接: https://arxiv.org/abs/2611.20602
作者: Abdur Rahman Ovy,Xiaosong Du
类目: Machine Learning (cs.LG)
*备注: Submitted to AIAA AVIATION Forum 2027

点击查看摘要

Abstract:Airfoil geometry parameterization is essential in aerodynamic design optimization. Optimization search in the large space suffers from low efficiency and modest feasible region, which could further incur failures to convergence. Generative adversarial networks (GAN), aiming to learn the statistical distribution of existing data, pave a new avenue for shape generation and airfoil parameterization. GAN manages to implicitly reduce design space by generating only smooth, realistic airfoils. Meanwhile, quantum computing has been emergingly exhibiting great potential for achieving superior performance in specific types of engineering problems but has not been introduced to airfoil parameterization. To fill this gap, this work develops quantum computing-based Wasserstein GAN (Q-WGAN) and investigates the performance for airfoil parameterization. The Q-WGAN adopts parameterized quantum circuits as the generator to promote the generative performance, while the discriminator remains classical. The developed Q-WGAN is demonstrated on a dataset of optimal airfoils parameterized by free-form deformation (FFD) control points, resulting from 992 airfoil aerodynamic design optimizations. Results show that Q-WGAN achieves 92.1% mean fitting accuracy on the optimal FFD control points and 99.2% accuracy on airfoil coordinates, comparable with the classical WGAN performance of 92.28% on FFD control points and 99.26% on airfoil coordinates. Q-WGAN achieves lower Jensen-Shannon divergence, Fréchet distance, and Wasserstein distance than the classical WGAN, indicating an improved agreement with the reference airfoil distribution. Therefore, this work presents a convincing evidence that Q-WGAN makes a viable option for parameterizing airfoils and possesses a great potential for other relevant engineering applications.

[LG-53] Calibrated Functional Prediction from Overlapping Nonlinear Observations

链接: https://arxiv.org/abs/2611.20567
作者: Peida Li
类目: Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注: 34 pages, 8 figures

点击查看摘要

Abstract:We study prediction of nonlinear audit functionals from scalar observations of a shared hidden table state. For Schur-rational operators, local Gaussian projections yield a common sparse kernel for overlapping tasks, repeated row identities, and cross-row dependence. Shared-state residual samples extend this kernel through a positive-semidefinite feature representation. Positive matrix integration and sample-value enclosures connect the original inputs to the actual saved hybrid prediction. A sorted calibration-role perturbation bound then certifies its complete finite-pool pinball risk. Independent complete-loss selection chooses the ridge, score scale, and residual weight, while task-rank calibration supplies an upper endpoint. Evaluation uses twenty new reference contexts per dataset at two task sizes, giving 400 conditions across ten benchmark populations. The mean dataset complete-loss ratios are 0.9977 to the degree-two kernel, 0.9961 to matched complete-MC shrinkage, and 0.9911 to the retrospective minimum over 174 public configurations. The last ratio has a paired reference-bootstrap interval of [0.9898,0.9927]. On 40 profiles at coefficient tolerance 10^-7 , Gamma quadrature has a geometric mean cold-time ratio of .0646 to strengthened Arb/Petras integration.

[LG-54] RoShift: Lazy Shift and Weight-aware Rotation for Activation Sparsity in LLM Inference

链接: https://arxiv.org/abs/2611.20549
作者: Geunsoo Song,Jiwon Seo,Hoeseok Yang,Youngmin Yi
类目: Machine Learning (cs.LG)
*备注: 40 pages, 10 figures, including appendices

点击查看摘要

Abstract:Activation sparsity accelerates LLM decoding by skipping unnecessary weight reads and computation. However, SiLU-based LLMs exhibit little natural sparsity, and imposing high sparsity can substantially degrade accuracy. We propose RoShift, a training-free method that transforms activations into sparsity-friendly representations. RoShift introduces lazy shift, which recenters SiLU activation distributions around zero and defers correction to the end of the FFN to obtain additional sparsity. It also reduces sparsification-induced output error through rotation that jointly accounts for activation and weight information. By fusing the corrections required for shifting and rotation, RoShift translates these gains into decoding speedup. At 50% and 60% sparsity, RoShift improves average accuracy over TEAL by 1.60 and 3.29 percentage points, respectively, across five models. It achieves up to 1.72\times decoding speedup over the dense model on an RTX 5090 at 50% sparsity and yields a better accuracy-speedup Pareto frontier than TEAL. Code is available at this https URL

[LG-55] A Temporal Convolutional Network with Multi-Head Attention Model: A Surrogate for Parallel DEVS Semiconductor Fab Simulation

链接: https://arxiv.org/abs/2611.20493
作者: Vamsi Krishna Pendyala,Hessam S. Sarjoughian,Edward J. Yellig,Bharadhwaj Balasubramaniam
类目: Machine Learning (cs.LG)
*备注: 13 pages, 11 figures, 5 tables

点击查看摘要

Abstract:Semiconductor factories process many kinds of wafers organized into lots of various sizes and configurations. Different kinds of machines, each with its own operations, process wafers for hours to days, and the wafers undergo many incremental fabrication steps (with re-entry) and flows spanning days to weeks across connected machines. Fast and accurate predictions of throughput ( TH ) and turnaround time ( TA ) dynamics for individual machines and their combinations are essential for factory operations. To use machine learning methods for this purpose, this paper presents a Temporal Convolutional Network with Multi-head Attention (TCNA) architecture. It uses simulated datasets from a benchmark Parallel DEVS (PDEVS) semiconductor fabrication factory model. The factory processes configurable wafer lots with uniform and sinusoidal wafer arrival profiles and may undergo preventive maintenance. The wafer lot size, arrival profile, and factory maintenance policy are used as static covariates. TCNA is applied to a selected set of simulated discrete-event throughput and turnaround time trajectories. The TH and TA time series predicted by TCNA, evaluated with the common MAPE , MSE , and R^2 measures, show comparable or higher accuracy relative to other model architectures, with similar parameter counts and epoch times; the TCNA model size and execution time are similar to those of other architectures such as the Temporal Fusion Transformer. Although time series predictions are less accurate than high-resolution simulations, they can be deployed at a much lower computational cost. The TCNA achieves orders-of-magnitude speedup over PDEVS simulation with an acceptable loss of accuracy, enabling frequent product scheduling, capacity planning, and maintenance planning.

[LG-56] Separating learning from search and feedback in budgeted inverse design of periodic unit cells

链接: https://arxiv.org/abs/2611.20480
作者: Shaoliang Yang,Jun Wang
类目: Computational Engineering, Finance, and Science (cs.CE); Machine Learning (cs.LG)
*备注: 49 pages, 11 figures, including Supplementary Information

点击查看摘要

Abstract:To design a periodic material, engineers choose the shape and dimensions of a repeating cell to meet target properties. With few simulations, the benefit of a learned predictor can be hard to separate from the effect of which designs the search tries and how it uses simulator feedback. We compare two ways to predict properties in a library of 80 cell architectures, with up to 30 simulator calls per target. For each architecture, both searches use a table of predicted properties at different dimensions. One fills the table with predictions from neural networks. The other fits a simple model to the existing simulations for that architecture. Both searches use the difference between the predicted and simulated properties to adjust the next design. We generated 200 targets from simulated designs after fixing the analysis plan. The two searches showed no detectable difference in how often they matched all three properties within 2% or in the mean error of their returned designs. The upper limit of the 95% confidence interval for the gain in overall success rate was 4.5 percentage points. Follow-up controls on these targets changed success from 83.5% to 36.0% with a degraded table and 96.5% with an oracle table. In separate tests, using the simulation result to adjust the next design increased success for both types of table. Results cover one library and a coarse 32^3 voxel simulator. Designers can use this comparison to judge what a learned predictor adds when the search, feedback and budget are the same.

[LG-57] On the Applicability of p-adic Numbers in Computational Modules for Neural Networks

链接: https://arxiv.org/abs/2611.20467
作者: Vladimir Evgrafov
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The applicability of p -adic arithmetic as an intermediate computational module in an otherwise unchanged machine-learning pipeline is investigated. A one-feature linear regression on the scikit-learn Diabetes dataset serves as the test model. The fitted coefficients are represented in \mathbbQ_p , predictions are computed in that field, and rational reconstruction is applied to the outputs. In the pyAdic experiments, rational reconstruction either returns no result or produces values that differ from the baseline predictions; changing computational precision can change both the magnitude and sign of the reconstructed value. In SageMath, forward conversions succeed, but the reconstructed predictions differ from the baseline. An alternative representation based on digit reversal fails to preserve ordinary addition because of carry propagation. Under the tested settings, none of the test predictions is reconstructed correctly. These results show that the tested conversion and reconstruction procedures do not provide a compatible intermediate module even for linear regression. Models defined directly over p -adic data and operations offer a direction for further investigation.

[LG-58] AdaModalX: Synchronization-Free MFM Inference for Distributed Multimodal Mobile Sensing via Progressive Data Orchestration

链接: https://arxiv.org/abs/2611.20464
作者: Xiangrui Xu,Sicong Liu,Zhiwen Yu,Zimu Zhou,Bin Guo
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Multimodal foundation models (MFMs) support generalizable inference in distributed mobile sensing applications by integrating heterogeneous data streams. However, existing MFM inference pipelines typically require complete and synchronized multimodal inputs, whereas mobile sensing streams arrive asynchronously due to network variability and heterogeneous data volumes across modalities. Waiting for all modalities increases end-to-end latency, while inference with incomplete inputs can reduce accuracy. We argue that multimodal inference should depend on the utility of available information rather than modality completeness. Based on this principle, we propose AdaModalX, a progressive multimodal inference framework that decouples inference execution from modality synchronization through runtime utility-aware orchestration. At each device, importance-aware data uploading uses a lightweight probe distilled from the edge-side MFM to allow reliable samples to exit locally while uncertain samples continue uploading. At the edge, on-demand MFM activation constructs candidates from fully received samples and, when needed, imputed features derived from partially received delayed inputs. A lightweight contribution assessor built on frozen MFM encoders estimates candidate utility and activates the MFM once the candidate meets its corresponding utility threshold. Experiments on four mobile sensing applications show that AdaModalX improves the accuracy-latency trade-off, achieving up to 26.97% accuracy improvement and up to 64.75% end-to-end latency reduction.

[LG-59] XMimic and XSoma: Learning Humanoid Motion Tracking and Locomotion in Under a Minute

链接: https://arxiv.org/abs/2611.20452
作者: Nico Bohlinger,Jan Peters
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We present XMimic and XSoma, which learn humanoid motion tracking and locomotion on rough terrain in under a minute of training. Both follow SuperMimic, which learns a world model of the robot and optimizes the policy with analytic gradients through short rollouts imagined in this model. Existing implementations of this method do not make use of larger GPUs, as every training iteration is a long chain of small GPU operations that run one after another. XMimic and XSoma therefore shorten this chain and run the steps of every iteration at the same time on up to four GPUs. We compile the complete training loop with JAX, including the MuJoCo Warp simulator. For locomotion, XSoma leaves only the dynamics of the robot to the world model and computes everything else exactly, including the terrain height under the feet. It learns with a horizon of one second and without a critic. On four H100 GPUs, XMimic learns to track all 40 LAFAN1 clips with the Unitree G1 in 0.86 minutes, 3.7 times faster than BlinkMimic on one GPU. XSoma reaches the tracking target that Holosoma’s tests set for FastSAC in 0.77 minutes, 12.3 times faster than FastSAC on one GPU, while needing only an eighth of its environment steps.

[LG-60] Generative AI-Driven Parametric Design Optimization Framework for Vehicle Closure Systems

链接: https://arxiv.org/abs/2611.20451
作者: Dongeon Lee,Yurim Lee,Jiyong Kim,Sanghyeon Kim,Sanghyun Lee,Seungyeon Shin,Jeahong Kim,Namwoo Kang
类目: Machine Learning (cs.LG)
*备注: 26 pages, 15 figures

点击查看摘要

Abstract:This study presents a generative AI-driven parametric design optimization framework integrating performance-oriented generation with editable computer-aided design (CAD) and downstream computer-aided engineering (CAE) analysis for industrial vehicle closure systems. The framework is demonstrated on an automotive tailgate assembly with complex geometry, coupled components, and competing performance and feasibility requirements. A production-oriented CATIA model is restructured into a reference-based parametric template with a 13-dimensional design representation. An automated CAD/CAE workflow generates parametric models, performs structural analysis, and constructs a database of mass, maximum displacement, and maximum von Mises stress. Gaussian process regression surrogates predict performance, while diffusion and flow matching models generate design candidates. The design problem minimizes mass and maximum displacement, with stress assessed as an engineering feasibility constraint. Performance-guided generation using surrogate guidance and performance-aware training using a performance-aware loss incorporate engineering objectives, with sensitivity to guidance scale and loss weight evaluated. Generative search is compared with NSGA-II under matched evaluation budgets using a common surrogate model. Selected Pareto-region candidates are reconstructed in CATIA and re-evaluated through native CAE; all verified candidates are successfully regenerated and analyzed. Relative to the baseline, the knee-point design reduces mass by 7.1% and maximum displacement by 29.4%, while extreme verified solutions achieve maximum reductions of 10.3% and 37.3%, respectively. The results demonstrate performance-oriented multi-objective design exploration compatible with editable industrial CAD and native engineering verification.

[LG-61] Shortcut Learning in Feature-Based Algorithm Selection

链接: https://arxiv.org/abs/2611.20440
作者: Shuhei Tanaka,Shoichiro Tanaka,Toshiharu Hatanaka
类目: Neural and Evolutionary Computing (cs.NE); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Landscape features give learning models information about the objective function for algorithm selection, algorithm configuration, and performance prediction. Should these features capture more information? Not always, because a model may rely on information that the algorithms ignore. This phenomenon is called shortcut learning, and in-distribution accuracy can hide it. We propose the Rubber Ruler Test. It transforms the objective values of test instances with random strictly increasing transformations, reruns the algorithms, and measures prediction performance. Comparison-based algorithms such as CMA-ES perform identically after the transformation, so for them any drop reveals shortcut learning. On 10-dimensional BBOB, a model for two comparison-based algorithms learned shortcuts even from normalized objective values. The R^2 of a model using all ELA features dropped from 0.85 to 0.23, below 0.45 with invariant features only. Rank-based features reduced the drop when algorithms search using ranks. Data augmentation with transformed training instances improved prediction on transformed instances for every portfolio we studied, but lowered R^2 on original instances by up to 0.31. These results suggest that models should rely only on information that the algorithms use. We recommend reporting the Rubber Ruler Test alongside in-distribution accuracy and using it to choose feature extraction and data augmentation.

[LG-62] EEG Can Help EEG: Improving Clinical Scalp EEG Analysis through Intracranial Priors KDD2027

链接: https://arxiv.org/abs/2611.20435
作者: Xianglong Bao,Yi Ding,Zengxiang Li,Nan Liu,Kewen Wang,Zhe Wang,Ming Jin
类目: Machine Learning (cs.LG); Signal Processing (eess.SP); Neurons and Cognition (q-bio.NC)
*备注: Accepted at KDD 2027

点击查看摘要

Abstract:Scalp electroencephalography (EEG) is a cornerstone of clinical brain monitoring, making the quality of its signal fundamental to downstream tasks and models. However, that signal is heavily contaminated by non-neural interference and blurred by volume conduction. Existing methods operate almost exclusively within the scalp modality, so they inherit the contamination they would need to overcome. In practice, intracranial EEG (iEEG) observes the neural activity with far higher fidelity, yet it is invasive and therefore scarce, and the two modalities are rarely recorded as pairs at scale, share no common channel space, and offer no clean ground truth to directly supervise a mapping. Motivated by the neuronal synchronization mechanism shared by the two modalities, we propose a reprogramming approach that fits a signal-to-signal model on iEEG alone and freezes it as a cross-modal prior. We introduce ReBrain, a framework to transfer an iEEG-learned prior to scalp EEG, which re-expresses any scalp signal through that prior and can be inserted upstream of any downstream model with no pairing or fine-tuning. ReBrain consistently improves host models across recordings from multiple hospitals and across architecture families, and our analysis traces this gain to genuine neural structure carried by iEEG rather than to the architecture itself. This work provides evidence that a scarce clean modality can serve as a transferable prior for an abundant noisy one, offering a general recipe for cross-modal transfer among biosignals. Our code is available at this https URL.

[LG-63] Fuzzy Explainable Continual Hierarchical Reinforcement Learning for Robotic Navigation

链接: https://arxiv.org/abs/2611.20427
作者: Peng Qin,Tao Zhao
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Continual robot navigation requires reinforcement learning policies to continuously adapt to new environments while retaining previously acquired knowledge. Meanwhile, explainable reinforcement learning aims to learn meaningful behavioral representations through policy interaction. However, these objectives are inherently conflicting, as continual learning emphasizes policy stability whereas explainable behavior learning requires adaptive exploration. Existing methods rarely address both simultaneously. To resolve this issue, this paper proposes a Fuzzy Explainable Continual Hierarchical Reinforcement Learning (FECHRL) method. A two-stage behavior cloning strategy is introduced to balance rapid policy adaptation and explainable behavior evolution by progressively transitioning from strong imitation guidance to autonomous reinforcement learning. Furthermore, a fuzzy critic based on a Takagi–Sugeno fuzzy inference system is proposed to alleviate value extrapolation errors during the transition between imitation-guided learning and autonomous exploration, thereby improving continual training stability. Experimental results on sequential multi-scenario navigation tasks demonstrate that the proposed method effectively unifies continual learning capability and policy explainability within a single reinforcement learning framework, achieving stable knowledge retention and meaningful behavioral explanation.

[LG-64] Reasoning Again with a Quarter of the Experts: Retry First Then Train

链接: https://arxiv.org/abs/2611.20419
作者: Anik Jha
类目: Machine Learning (cs.LG)
*备注: 12 pages, 5 figures, 4 tables

点击查看摘要

Abstract:Removing three quarters of the routed experts from a 122B-parameter reasoning mixture-of-experts model makes it fit in a 12 GiB file, and breaks it in a specific way: on LiveCodeBench it enters verbatim reasoning loops that run to the token limit on 44 of 120 development problems, where the unpruned model at the same 2-bit quantization loops on none. We study how much of that loss can be recovered without changing the kept experts’ quantized bytes, in three steps of increasing cost. On development problems, a sampling penalty recovers part of it. Stopping a reply when it starts to loop and drawing again recovers more on top of it, with no weight change. Our method, B, adds rank-8 trained corrections to the kept experts, loaded as an adapter, and restores two experts per layer; it is fitted layer by layer to the unpruned model’s expert block on the pruned model’s own activations. In a pre-registered comparison on 130 easy and medium LiveCodeBench problems that no part of the study had touched, with both arms given the same sampling penalty and the same stop-and-redraw policy, B solves 116 problems against 104 (+9.2 pp, 95% CI [+3.1, +15.4]; 15 problems gained, 3 lost). A training-free weight edit aimed at the loops, tested the same way earlier, did not help. B removes more wrong answers than loops, and costs 1.7x the wall time. On two older benchmarks used as regression checks, B stays inside a pre-set 2 pp tolerance but shows no gain (HumanEval+ +0.6 pp, MBPP+ -1.9 pp). The evidence is one model, one quantization and one benchmark family; we state what it does and does not show.

[LG-65] Integrated GPU Core for Mamba-3

链接: https://arxiv.org/abs/2611.20409
作者: Shiva Raja,William Desueza,Ajay Joshi
类目: Hardware Architecture (cs.AR); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Mamba-3, the latest Mamba state-space model (SSM), enables efficient long-context modeling, matching Mamba-2 perplexity with half the state size and increasing decoding arithmetic intensity. However, its multi-input multi-output (MIMO) computation, exponential-trapezoidal discretization, and rotary positional embeddings (RoPE) introduce dense, element-wise, and sequential operations. Despite kernel fusion, conventional GPUs execute these operations inefficiently. Our characterization of Mamba-3 hybrid models reveals numerous kernel launches, substantial data movement, and increased register pressure that reduces warp occupancy and memory-latency hiding. We propose RescuCore, a specialized GPU execution unit for Mamba-3 inference that retains other hybrid model layers on existing GPU resources. Integrated alongside CUDA Cores and Tensor Cores within each Streaming Multiprocessor (SM) sub-partition, RescuCore shares the register file, shared memory, and warp schedulers. Dense projections remain on conventional GPU resources, while RescuCore pipelines discretization, normalization, RoPE, MIMO computation, recurrent state updates, and output gating to reduce kernel fragmentation and intermediate data movement. We select BF16 after evaluating multiple numerical formats, preserving model accuracy with an estimated area overhead of 5.9% of an A100 die for state size 32 and MIMO rank 4. We evaluate RescuCore using cycle-accurate Accel-Sim simulation, RTL synthesis, and synthesis-informed AccelWattch power modeling. Compared with conventional GPU execution, RescuCore achieves 1.4-9.4x speedup and 1.4-5.0x energy reduction during prefill, and 1.3-2.8x speedup and 1.2-2.6x energy reduction during autoregressive decoding. These results demonstrate the potential of GPU-integrated specialized execution for SSM-based hybrid models. Subjects: Hardware Architecture (cs.AR); Machine Learning (cs.LG) Cite as: arXiv:2611.20409 [cs.AR] (or arXiv:2611.20409v1 [cs.AR] for this version) https://doi.org/10.48550/arXiv.2611.20409 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-66] Latent Feedback Is a Compute and Data Amplifier

链接: https://arxiv.org/abs/2611.20404
作者: Benhao Huang,Chaitanya Dwivedi,Xuezhe Ma,Alexander Moreno,Hector Liu
类目: Machine Learning (cs.LG)
*备注: 23 pages. Code: this https URL

点击查看摘要

Abstract:When high-quality data runs out, a common practice is to train on it for several epochs, but each repeat is worth less than the last. Latent feedback, which feeds each token’s top-layer state into the next token’s input, is another way to spend that compute, yet it has not been tested against a vanilla Transformer that spends the same compute on more epochs. We run this test across a ladder of model sizes, unique data and epochs, tuning both models at every size and amount of unique data. Latent feedback makes every FLOP and every token count for more except when unique data is scarce relative to model size, a regime that today’s models are far from. We find that latent feedback serves as an amplifier: with unique data fixed, it reaches a Transformer’s loss with less compute; with epochs fixed, it reaches that loss with less unique data. Furthermore, both savings grow with unique data per parameter and with repetition, and downstream tasks agree. To go beyond our grid, we fit one scaling law to both models, in which latent feedback enters as a Transformer with more parameters; it predicts that both savings keep growing with unique data and repetition. Fitted on models from 60M to 271M parameters, the law predicts the losses at a held-out 778M size to within 0.3% and the gap between the two models to within 0.05%. At most, latent feedback saves 48% of the compute and 47% of the unique data. When data, not compute, is the limit, latent feedback is the better use of a FLOP for a model of a given size.

[LG-67] Architectural Principles for Parameter-Efficient Unitary Transformations in Coherent Optical Neural Networks

链接: https://arxiv.org/abs/2611.20390
作者: Kazuo Aoyama,Hiroshi Sawada,Shota Kita,Akihiko Shinya,Masaya Notomi
类目: Machine Learning (cs.LG); Hardware Architecture (cs.AR); Applied Physics (physics.app-ph); Optics (physics.optics)
*备注: 28 pages, 28 figures

点击查看摘要

Abstract:Coherent optical neural networks based on interferometric architectures offer a promising route toward high-throughput, energy-efficient computation by exploiting the multiple degrees of freedom of light, including spatial, temporal, and spectral modes. However, practical implementations are constrained by the number of tunable parameters due to fabrication complexity, device footprint, and control overhead. In such parameter-constrained settings, achieving high expressive power remains a central challenge. We identify two key architectural requirements for maintaining high expressive power under these constraints: (i) statistical uniformity in the implemented transformations and (ii) strong mixing capability per parameter, which enables rapid and unbiased redistribution of information across modes. To realize these requirements, we develop a logarithmic-depth butterfly-topology architecture whose fundamental building block is a single phase-shifter–beam-splitter pair, in contrast to conventional rectangular mesh architectures based on Mach-Zehnder interferometers. Computational evaluations show that, particularly at restricted parameter counts, the proposed architecture achieves lower matrix-reconstruction errors and higher classification accuracies than conventional meshes in both feedforward and recurrent networks. We further quantify these design principles using graph-theoretic measures, expected squared moduli of transformation-matrix entries, and matrix-reconstruction error. The resulting framework provides a quantitative and physically interpretable basis for parameter-efficient optical neural networks.

[LG-68] Halting in Recurrent Sequence Models NEURIPS2026

链接: https://arxiv.org/abs/2611.20377
作者: Simon Dräger
类目: Machine Learning (cs.LG)
*备注: 16 pages, 7 figures, 2 tables. Accepted at the NeurIPS 2026 Workshop on Interpretability as a Science

点击查看摘要

Abstract:Recurrent sequence models, such as recurrent neural networks (RNNs) and state space models (SSMs), are often trained to emit an end-of-sequence (EOS) token when their output sequence is complete. During training, EOS is treated as another token without a special loss term or incentive. Although this training paradigm is widely used, it is not understood how models decide to output EOS. We study a minimal version of the EOS halting problem in which networks are trained to emit a given input sequence an instructed number of times, followed by EOS. We train tanh RNNs and linear SSMs on this task, finding that the RNNs encode the number of tokens left to copy in a linear subspace, of which two principal components capture most of the variation. Hidden states with the same remaining count converge to a shared representation regardless of their instructed repeat count. By modulating the hidden state using an affine transformation at the boundary between copied sequences, we shift the RNN’s state to the representation shared by states with one fewer copy remaining. The unchanged network emits EOS one copy earlier while reproducing the shortened target sequence in 94.4% of new sequences, revealing that this representation causally determines the stopping time. In the SSMs, the difference between trajectories with different repeat counts is separated across two-dimensional rotation planes whose contributions to the EOS comparison cancel during copying and align near the stopping time. Rotating the planes independently removes most of the intended stopping while preserving the copied content, which shows that the stopping time depends on the planes’ relative phases. These results identify two candidate primitives of EOS halting in recurrent models: a shared encoding of the number of remaining tokens in tanh RNNs and phase alignment across fixed rotations in linear SSMs.

[LG-69] HoloGene: Learning to Lift Sliced Spatial Transcriptomics to Holistic 3D Gene Fields NEURIPS2026

链接: https://arxiv.org/abs/2611.20368
作者: Yumin Zheng,Qingtian Zhu,Ziyan Zhu,Kailu Song,Yinqiang Zheng,Jun Ding
类目: Machine Learning (cs.LG)
*备注: NeurIPS 2026

点击查看摘要

Abstract:Spatial transcriptomics (ST) provides spatially resolved gene expression measurements, but most 3D ST datasets are acquired as independently processed 2D sections. Computational approaches span slice registration and integration, explicit or generative interpolation, and coordinate-field modeling. Within the last family, implicit neural representations (INRs) offer a continuous, coordinate-queryable parameterization, but standard isotropic coordinate encodings do not explicitly reflect that serial-section measurements are dense in-plane and sparse along the axial direction. We present HoloGene, an anisotropic conditional INR framework for estimating continuous 3D gene-expression fields from pre-registered ST sections. Its central contribution is an acquisition-aware factorization that couples dense in-plane synthesis with sparse axial evolution using established GAE, FiLM, pseudo-label supervision, and CORAL components. HoloGene compresses high-dimensional expression profiles into a graph-autoencoder latent manifold, uses pseudo-label-derived signature conditioning to provide coarse intra-section context, and applies depth-dependent FiLM modulation to model axial variation. A signature-stratified correlation-alignment regularizer encourages cross-slice covariance alignment in latent features. Using held-out measured sections as cross-sectional references, HoloGene improves reconstruction accuracy and biological preservation metrics compared with general-purpose INRs and adapted spatial-transcriptomics baselines. These results support anisotropic conditional modeling as a practical strategy for estimating 3D expression fields from discontinuous ST sections. Code is available at this https URL.

[LG-70] Settling the Support Complexity of Correlated Equilibria

链接: https://arxiv.org/abs/2611.20362
作者: Ashkan Soleymani,Patrick Jaillet,Gabriele Farina,Akbar Rafiey
类目: Computer Science and Game Theory (cs.GT); Machine Learning (cs.LG)
*备注: Innovations in Theoretical Computer Science (ITCS), 2027

点击查看摘要

Abstract:We study the support complexity of approximate correlated equilibria (CE) in two-player games, that is, the minimum number of pure action profiles that must be assigned positive probability to approximate a correlated equilibrium of a two-player matrix game. For \varepsilon -approximate coarse correlated equilibria, constant approximation is possible with support O(\log N/\varepsilon^2) in N\times N games. For correlated equilibria, the best known upper bound was O(\log^2 N/\varepsilon^4) [Babichenko et al., 2014, 2017], but whether this bound was tight remained open. We establish a matching lower bound, resolving the open question of Ko and Schvartzman [2017]. Specifically, we show that, already for two-player zero-sum games, for every 1\leq k\leq N^1/2 the worst-case correlated-equilibrium gap achievable by a distribution supported on at most k pure profiles is \Theta\left( \min\left\1,\frac\sqrt\log Nk^1/4\right\ \right), which implies a \Theta(\log^2 N/\varepsilon^4) worst-case support bound. The proof builds on a probabilistic-method construction based on random sign matrices. Comments: Innovations in Theoretical Computer Science (ITCS), 2027 Subjects: Computer Science and Game Theory (cs.GT); Machine Learning (cs.LG) Cite as: arXiv:2611.20362 [cs.GT] (or arXiv:2611.20362v1 [cs.GT] for this version) https://doi.org/10.48550/arXiv.2611.20362 Focus to learn more arXiv-issued DOI via DataCite (pending registration) Submission history From: Akbar Rafiey [view email] [v1] Mon, 16 Nov 2026 00:41:56 UTC (15 KB)

[LG-71] Flow Matching for Sampled-Data Control of Angular Distributions

链接: https://arxiv.org/abs/2611.20360
作者: Xin Mao,Anqi Dong,Hongteng Xu,Kenan Zhang
类目: Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注:

点击查看摘要

Abstract:Synchronization and clustering in oscillator networks depend on the joint distribution of the phases. Flow Matching can steer this distribution by learning an angular velocity field from shortest arcs between paired configurations. Each arc is traversed at constant speed, yet the resulting field can vary along its own trajectories. In this article, we study how to control angular transport on the circle and product torus when each controller evaluation selects a rate to hold until the next update. Using the same local velocity samples, Flow Matching learns the instantaneous angular rate and MeanFlow learns its average along the reference trajectory over the control interval. Under direct angular actuation, holding the exact average gives the reference displacement. We relate the displacement error of holding the instantaneous rate and the training residual of MeanFlow to the conditional covariance of the teacher velocities. Experiments with 100 phases show that endpoint pairing affects the comparison. MeanFlow gives smaller one-step errors under minibatch pairing, while permutation alignment reverses the ranking for exchangeable phase clusters.

[LG-72] EPI-SST: Renewal-Structured State-Space Transformers for Epidemic Forecasting

链接: https://arxiv.org/abs/2611.20352
作者: Yiyun Chen,Zewen Liu,Shengpu Tang,Max S. Y. Lau,Wei Jin
类目: Machine Learning (cs.LG)
*备注: 13 pages, 4 figures, 12 tables

点击查看摘要

Abstract:Epidemic forecasting supports public health decisions ranging from capacity planning to intervention timing, for which both point accuracy and calibrated uncertainty are essential. Although neural models have advanced point forecasting, producing reliable probabilistic predictions over multiple future horizons remains challenging. Renewal models encode epidemic propagation, while neural forecasters learn flexible multi-horizon representations; integrating these strengths within a probabilistic future rollout remains underexplored. We propose EPI-SST (Epidemic State Space Transformer), a probabilistic forecaster centered on an evidence-driven renewal transition. At each horizon, neural evidence produces a bounded update to an internal multiplier and advances a finite-memory renewal state that generates the future mean trajectory. A separate heteroscedastic Gaussian head controls predictive scale using a validation-selected log-variance penalty. Across four epidemic surveillance benchmarks, EPI-SST achieves the lowest pooled RMSE on every dataset, reducing RMSE by up to 14.2 percent relative to the strongest baseline. It also achieves the lowest continuous ranked probability score (CRPS) at horizons h = 4, 5, 6 on every dataset, with reductions of up to 43 percent relative to the strongest probabilistic baseline. During the abnormal 2019-2020 influenza season, EPI-SST reduces CRPS by 16.9 percent and calibration error by 78.9 percent; a three-seed replication preserves its distributional lead. Ablation studies and frozen-path interventions show that the renewal transition, attention conditioning, horizon coupling, and uncertainty-scale control make complementary contributions to predictive performance.

[LG-73] Spectral Landmarks: Alignment of Landmark Kernel Translates with the Leading Eigenfunctions of a Kernel Operator

链接: https://arxiv.org/abs/2611.20336
作者: Jeanie Schreiber
类目: Numerical Analysis (math.NA); Machine Learning (cs.LG); Functional Analysis (math.FA); Metric Geometry (math.MG)
*备注:

点击查看摘要

Abstract:We study the problem of choosing m landmark points so that the span of the kernel translates is as close as possible to the leading s -dimensional eigenspace of a positive-definite kernel. The natural measure is the Hilbert-Schmidt operator discrepancy between the reproducing-kernel-Hilbert-space (RKHS) projectors onto the two subspaces. We give a closed form spectral alignment function in terms of the truncated Mercer kernel and relate it to the classical integrated power function (Nyström trace error) through a two-sided sandwich bound. Minimizing the function by gradient descent on the landmark coordinates produces landmark sets that are evenly spaced. Theoretical results show that optimal configurations are those whose eigenfunction sampling matrix is a scaled isometry. On homogeneous spaces these are exactly the minimal equal-weight cubature rules for products of leading eigenfunctions, and for every m\ge s they are the configurations whose sampled eigenfunctions form a unit-norm tight frame. Experiments on \mathbbS^1 and \mathbbS^2 agree with the theory. On the circle, with s\le2m and a moderate bandwidth, every random start converges to equispaced landmarks, and on the sphere the optimizer recovers the regular tetrahedron, octahedron, and icosahedron from random starts. Finally, we analyze the numerical conditioning of the problem.

[LG-74] Ontology-Compiled Control Barrier Functions for Heterogeneous Robot Teams

链接: https://arxiv.org/abs/2611.20306
作者: Peng Wu,Seyede Fatemeh Ghoreishi,Nathan Bastian,Mohsen Imani,Tian Lan,Mahdi Imani
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Safety filters for heterogeneous robot teams fail because of missing constraints, overestimated braking capabilities and conflicts between individually satisfiable constraints. We present OntoCBF, which compiles a domain ontology with a class taxonomy, typed relations, safety axioms and a dynamics profile per robot class into a control barrier function filter with one quadratic program per agent. Description-logic entailment supplies the constraints of any team online, including those that follow from chains of relations and from classes defined by what a robot carries, so the existing safety axioms cover new robot classes, payload classes and combinations. We prove that a subclass may reuse the braking capability of its parent whenever it can produce every velocity of the parent, a property separate from subsumption. For box-bounded double integrators this property is also necessary, and every loss of safety then begins with a conflict. Once three robots interact, the safe set of pairwise barriers contains states from which a collision is unavoidable, and such states persist in larger teams under any fixed capability reduction. In a simulated warehouse where humans ignore the robots, a filter written for the design classes violates safety in 89-100 % of the episodes with unseen classes. In all scenarios, including one with 32 robots, OntoCBF violates safety in 20-70 % of the episodes, each time after a conflict, with a median deepest intrusion of 10-17 % of the keep-out distance. A multi-agent PPO policy trained with the filter in the loop lowers the rate of infeasible quadratic programs, nearly all of them conflicts, by 18-41 %.

[LG-75] Scaling large language model-assisted line-of-therapy adjudication with human-in-the-loop evidence-calibrated selection

链接: https://arxiv.org/abs/2611.20259
作者: Olatunji Oni,Sanhita Sengupta,Marc De Benedetti,Nathanial Eddy
类目: Machine Learning (cs.LG); Quantitative Methods (q-bio.QM); Applications (stat.AP)
*备注: 12 pages, 4 figures

点击查看摘要

Abstract:Large language models (LLMs) have enabled the ability to scale clinical adjudication tasks traditionally dependent on exhaustive manual review. Line-of-therapy (LoT) adjudication, a major bottleneck in oncology cohort construction, is a prime example. Line boundaries must be inferred from longitudinal treatment histories according to nuanced, study-specific definitions, at every regimen transition for every patient. LLMs can apply this adjudication at scale, but their reliability varies across patients. Treatment histories differ in complexity, and ambiguous or atypical patterns depend on expert clinical judgment. Reliable automation thus requires verified support for model assignments and the ability to identify cases that warrant manual review. Here, we present evidence-calibrated selection (ECS), a human-in-the-loop framework for scaling LLM-assisted LoT adjudication. ECS derives confidence from observed clinician adjudications, prioritizes patients with the least-supported assignments for follow-up review, and returns completed adjudications to the model as precedent. As the evidence base grows, ECS continuously calibrates confidence to estimate residual error, thus enabling more efficient use of clinician time. In simulations on 136 clinician adjudicated patients with multiple myeloma, ECS achieved a residual error of 5% after review of 44.6% of patients, compared to 62.2% when review was directed by LLM self-reported confidence and 81.9% under random sampling. These results support a scalable approach to study-specific LoT adjudication, using clinician evidence to inform automated predictions, minimize error, and reduce review burden.

[LG-76] ght Bounds for Non-Monotone Submodular Maximization under a Cardinality Constraint with Constant Recourse

链接: https://arxiv.org/abs/2611.20257
作者: Vaneet Aggarwal
类目: Data Structures and Algorithms (cs.DS); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We study the maximization of a nonnegative submodular function under a cardinality constraint when elements arrive one at a time and an algorithm must report a feasible set after every arrival. The recourse of an algorithm is the largest number of insertions and deletions that it makes to the reported set at a single arrival. For every \epsilon0 , we give an algorithm with recourse O(\epsilon^-2) whose reported set has, at every time, expected value at least 1/4-\epsilon times the largest value of a feasible set of arrived elements. The algorithm evaluates the function polynomially often, and only on sets of arrived elements. We also show that no algorithm with constant recourse achieves a ratio above 1/4 , even if it knows the function and the arrival order in advance and has unlimited computation, and that any fixed ratio above 1/4 requires recourse that grows as a positive power of the cardinality budget. Thus 1/4 is the optimal limiting ratio with constant recourse. The previous best ratio with constant recourse was 3-2\sqrt2\approx0.172 . Our algorithm assigns selection probabilities of at most one half and maintains, as elements arrive, an inequality bounding the value of every feasible set by a multiple of the expected value of a random set drawn with these probabilities; the reported set moves to a rounding of these probabilities through gradual replacements. The upper bound uses a weighted directed graph whose vertices arrive in groups of decreasing sizes.

[LG-77] Deep Sequence Model for Genome Wide Discovery of Coding and Regulatory Element Signatures

链接: https://arxiv.org/abs/2611.20255
作者: Rayhaneh Shabani Nia,Ali Karkehabadi
类目: Machine Learning (cs.LG); Genomics (q-bio.GN)
*备注: 4 pages, 3 figures, 1 table. Accepted at CVIS 2025. A shorter 3-page version was published in the Journal of Computational Vision and Imaging Systems, Vol. 11, No. 1 (2026)

点击查看摘要

Abstract:Interpreting deep neural networks for genomic sequence classification remains challenging despite strong predictive performance. We develop a lightweight CNN with post-training gradient-based analysis to identify which sequence positions drive coding versus intergenomic classification. Applied to the standardized demo_coding_vs_intergenomic_seqs dataset, our model achieves 91.87% validation accuracy while revealing interpretable patterns: nine hot spots with consistently high importance, clear class-specific separation at positions 20-100 (coding) and 150-190 (intergenomic), and a strong mean-variance correlation (r = 0.530) indicating robust discriminative features. Gradient-based importance analysis shows that the model implicitly learns biologically meaningful sequence distinctions without explicit annotations. This work demonstrates that neural network interpretability and accuracy can coexist, providing a framework for understanding genomic sequence classification and enabling biology-driven hypothesis generation.

[LG-78] Cache: Serving Multi-Round Diffusion Image Editing with Compressed Activation Caches

链接: https://arxiv.org/abs/2611.20246
作者: Yuyang Tian,Desen Sun,Shuncheng Jie,Lishan Yang,Sihang Liu
类目: Distributed, Parallel, and Cluster Computing (cs.DC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Multi-round diffusion image editing can reduce computation by reusing intermediate activations, but retaining these activations across editing rounds introduces substantial memory and bandwidth overheads. Compression can reduce these overheads, but efficient cache construction is challenging: compressing only after editing requires buffering over 100 GB of activations per request. Moreover, the compressed cache must support selective reuse, as future masked edits require only a subset of the cached activations. It also exposes a scheduling challenge, as cached edits are faster than initial generations, causing naive policies to severely delay either initial generations or editing requests. In this paper, we aim to address both challenges through compression-informed serving. We develop TileCache, a serving system for multi-round diffusion editing. It compresses reusable activations and incorporates a scheduler that balances both edits and initial generation requests. Compared with the state-of-the-art image editing baseline, TileCache sustains over 2x the request arrival rate at 90% SLO attainment. It completes edits 2.4-4.5x and initial generations 1.3-2.2x faster than the baseline, while preserving image quality.

[LG-79] UrbanROI: Causal Modeling for Regional ROI Forecasting under POI Entry KDD2027

链接: https://arxiv.org/abs/2611.20180
作者: Zhaoqi Zhang,Miao Xie,Linyou Cai,Qianlong Xie,Xingxing Wang,Siqiang Luo,Gao Cong
类目: Machine Learning (cs.LG)
*备注: Accepted by KDD 2027

点击查看摘要

Abstract:As urban commercial environments continue to evolve rapidly, accurately modeling the economic impact of point of interest (POI) entry is essential for supporting data-driven advertising allocation, merchant onboarding, and regional business decision-making. Existing return on investment (ROI) prediction methods mainly focus on individual ads or merchants, overlooking the interaction effects among POIs and failing to capture the causal influence of structural changes in urban commercial ecosystems. In this paper, we introduce a novel research problem – regional ROI forecasting under POI entry, which aims to predict the future ROI trajectory of a commercial region after the introduction of a new POI by modeling temporal ROI evolution and competition-driven interactions among surrounding POIs. To address these challenges, we propose UrbanROI, an intervention-aware causal forecasting framework for region-level economic outcome prediction. UrbanROI leverages an Interaction-aware Regional Graph Module to explicitly model competitive relationships among POIs based on user request contexts, spatial proximity, and request-level commercial signals, and constructs structurally aligned treatment and control graphs to simulate factual and counterfactual regional states. Extensive experiments on large-scale real-world POI and advertising datasets demonstrate that UrbanROI consistently outperforms strong baselines in post-entry regional ROI forecasting. The results demonstrate its effectiveness in temporal economic modeling, inter-POI interaction representation, and intervention-aware forecasting, providing support for regional economic analysis and business decision-making in dynamic urban commercial systems.

[LG-80] Weights Caches: LLM s Can Program Tabular Foundation Models

链接: https://arxiv.org/abs/2611.20174
作者: Shi Bin Hoo,Magnus Buehler,Omar Swelam,Frank Hutter,Lennart Purucker
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Neural networks are typically programmed through optimization, rather than by explicitly specifying their computations. Recent work on weight-space programming suggests that large language models (LLMs) can instead reason about and manipulate neural-network weights directly. In this work, we show that LLMs can directly program statistical computations in tabular foundation models (TFMs), whose parameters enable inference across different datasets. We investigate three complementary settings using GPT-6 Astra, without gradient-based optimization. First, Astra directly programs nanoTabPFN from zero-initialized weights, achieving a mean ROC AUC of 0.849 across 20 TabArena datasets, compared to 0.876 for a pretrained baseline. Second, Astra diagnoses and repairs a signed-product regression failure in pretrained TabPFN-3.5, restoring near-perfect predictions on held-out synthetic datasets while preserving performance on TabArena-Lite, with only a 0.08% average relative increase in error. Third, it writes KV-cache-editing programs that steer TabPFN-3.5’s predictions on unseen synthetic datasets, demonstrating transformations such as reversing linear trends and converting sine waves into approximate square waves, without modifying model weights. Together, our findings suggest that LLM-driven programming could complement gradient-based optimization across TFMs’ development and usage, while offering a new way to investigate and intervene in their internal computations.

[LG-81] Kernel-Guided Learning of Lifted Sparse Polynomial Models of Nonlinear Dynamical Systems

链接: https://arxiv.org/abs/2611.20157
作者: Mohammad Itani,Timm Faulwasser
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Polynomial models of dynamical systems enable constructive analysis via sum-of-squares programming. Yet, many physical systems involve non-polynomial nonlinearities which for certain classes of systems may be inferred from the underlying physics even when the exact dynamics are unknown. We show that fitting lifted polynomial models directly to lifted data is problematic, since the exact lifting constraints render the lifted derivatives linearly dependent. Instead, we learn the dynamics in the extended state, i.e., the state together with the known nonlinearities, and obtain the dynamics of the nonlinearities via the exact lifting relations. To avoid the combinatorially large libraries that sparse regression requires when the polynomial degrees are unknown, we propose KINDy, which selects the degree of each component by kernel ridge regression before applying sparse regression to the correspondingly reduced library. We prove that, for sufficiently small noise and regularization, provided the data satisfy a rank condition, the selected degree never underestimates the true degree. On the cart-pole system, KINDy recovers the true degrees at low noise levels and achieves speedups of up to 8.5 over sparse regression on the full library, while in almost all cases yielding models that are at least as sparse and have lower or equal test errors.

[LG-82] Symmetry Fingerprinting with Geometric Learnable Invariants Transformers

链接: https://arxiv.org/abs/2611.20150
作者: Bart M.N. Smets,Gijs Bellaard,Pieter Wils
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Geometric deep learning has made available a large variety of structurally equivariant neural network architectures, meaning a particular transformation of their input causes a predictable transformation of their output. But structural equivariance is a very strong inductive bias that can limit performance if chosen poorly. In contrast, choosing an architecture with low geometric bias necessitates data augmentation to obtain correct behavior; extending the training process. We propose a novel neural network architecture that is soft equivariant; we select a set of symmetry groups and during training the network will become equivariant only to those symmetry groups that are appropriate given the training data. After training we can inspect the network to obtain a symmetry fingerprint that will tell us exactly which symmetries ended up being relevant and to what degree. We show how our architecture and the fingerprinting process work in the case of an image classification task.

[LG-83] Cleanly-DINO: Efficiently Learning Visual Task Spaces for Robotics

链接: https://arxiv.org/abs/2611.20137
作者: Adrian Röfer,Melika Shekarriz,José Arce y de la Borbolla,Iman Nematollahi,Abhinav Valada
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注: 11 pages, 5 figures, accepted at COMPOGEN Workshop CoRL 2026, Austin, USA

点击查看摘要

Abstract:Learning manipulation tasks from few demonstrations requires abstractions that capture the changes relevant to task completion, such as object poses in rigid-body manipulation. However, tasks such as cleaning are defined by changes in visual state rather than object pose, making the choice of task representation less clear. In this work, we propose Cleanly-DINO to learn visual task spaces from few demonstrations. Given a mask of a task-relevant region and frozen ViT features, Cleanly-DINO fits a projection to isolate a task space and fit a progression metric to it. We train the model on diverse cleaning examples and evaluate its ability to localize incomplete task regions and transfer across visual domains. We also study its utility for robot learning by using it as an observation encoder and reward signal in a reinforcement learning setting. Finally, we deploy our fitted model within a real-world robotic wiping pipeline to evaluate its use in guiding task execution.

[LG-84] QA-CMIF:Accurate and Lightweight Cross-Modal Latent Imputation and Fusion for Fault-Tolerant sEMG-Inertial Hand Gesture Recognition

链接: https://arxiv.org/abs/2611.20124
作者: Xinyu Tao,Shuang Huang,Yeyun Cai,Fang Deng
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Fusing surface electromyography (sEMG) with inertial measurements is widely used for myoelectric prosthesis control and wearable human-machine interfaces. In daily wear, however, socket micro-displacement, electrode lift-off, and wire failures often corrupt an unknown subset of sEMG channels through burst artifacts, lift-off shocks, or open-circuit dropouts. Existing fusion networks mix all channels in the first convolutional layer, so a few corrupted electrodes contaminate the entire latent representation, whereas gate-only attenuation of corrupted channels discards the co-activation cues needed to separate fine-grained gestures. We propose QA-CMIF, a quality-aware cross-modal latent imputation and fusion framework that encodes each sEMG channel independently, scores its reliability with a soft gate, and protects intact channels via a confidence-weighted residual bypass, while a cross-attention imputer repairs degraded channels in latent space using the inertial stream and remaining reliable sEMG channels under a bad-channel alignment loss. Across within-subject evaluations on 29 participants from Ninapro DB3, DB2, and DB5 spanning 17-52 gesture classes, QA-CMIF outperformed the gate-only baseline under 50% channel corruption by 1.29-2.65 percentage points, denoted pp, across all three 17-class cohorts with paired-t Holm-adjusted p0.002, widened this margin to 2.59-4.15 pp on the 40- and 52-class tasks, and surpassed corruption-augmented conventional fusion by 4.61-19.85 pp. Requiring only 0.175-0.185M parameters and 2.45-2.63M multiply-accumulate operations across the 2000 Hz and 200 Hz setups, QA-CMIF saves 46%-55% in parameters and 72%-81% in compute over recent baselines with a 1.44 ms 99th-percentile single-thread CPU latency.

[LG-85] High-Throughput Serving of Nemotron 3.5 Transcribe with Regime-Aware Megakernels and Kernel Fusion

链接: https://arxiv.org/abs/2611.20095
作者: Aditya Chatterjee,Sanjay Singh Chauhan,Mayank Jain,Arushi Dhingra,Sahin Chowdhury,Rahul Mittal
类目: Performance (cs.PF); Distributed, Parallel, and Cluster Computing (cs.DC); Machine Learning (cs.LG); Audio and Speech Processing (eess.AS)
*备注:

点击查看摘要

Abstract:Nemotron 3.5 Transcribe ranks first among open-weight models on the Hugging Face Open ASR leaderboards. Serving it efficiently takes more than fast operators: continuous batching changes the amount and mix of work at every iteration, and the best fusion granularity changes with it. We present a serving engine that, for every pass, chooses between persistent megakernels and chains of fused kernels at crossovers measured offline. On an H100 SXM GPU, throughput of our accuracy-preserving serving peaks at 4,345x real time, 1.39x the peak of tuned vLLM, and is up to 4.39x faster at low concurrency; on B200 it peaks at 6,367x, 1.64x vLLM at the highest concurrency we measured. Word error rate (WER) is 1.91 to 1.93% against 1.88 to 1.93% for vLLM. Within a 400 ms budget for p95 end-to-end completion time, it serves 8x as many concurrent requests. Choosing per pass beats either level alone: 23% higher throughput than always using megakernels at 128 concurrent requests, and 9% higher than always using fused kernels at 8. Opt-in switches, including greedy MoE layer skipping, add 13 to 16% throughput on B200 for offline jobs that can accept higher WER.

[LG-86] Does Learning Order Explain Forgetting? The Lie Bracket as an Early-Warning Signal for Catastrophic Forgetting Without Old Data

链接: https://arxiv.org/abs/2611.20087
作者: Ephraim Kofi Ayi Kwapong
类目: Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注: 14 pages, 4 figures. Code and results: this https URL

点击查看摘要

Abstract:A network trained on task A and then task B ends at different weights than one trained on B and then A. To leading order this gap is set by the Lie bracket of the two tasks’ gradient fields, and Dherin (2023) conjectured that a non-vanishing bracket may be linked to catastrophic forgetting. I test this conjecture empirically. Across Split MNIST, Split Fashion-MNIST and Split CIFAR-10, the forgetting of A is almost exactly the difference in A’s loss between the two training orders (Spearman rho = 0.96-1.00). I then ask whether the bracket can predict that forgetting early in training on B and without any of A’s data, using only a small summary stored when A finishes: its gradient and its top 20 Hessian eigenpairs. When the old task is well learned and forgetting is slow (task-incremental), this stored bracket ranks forgetting with rho = 0.70-0.81, against 0.33-0.58 for the penalty of Elastic Weight Consolidation, and matches or exceeds re-evaluating A on its own data. When forgetting is fast (domain-incremental), the stored summary goes stale and the EWC penalty is as good or better. The bracket fails outright when the old task is only half learned; I trace this to a residual gradient on A an order of magnitude larger, which makes first-order (slope) damage matter. The bracket is therefore a useful early-warning signal for the quiet kind of forgetting, under a clear condition.

[LG-87] Decode Early Verify Late: Streaming Offline Speech Recognizers Without Accuracy Loss

链接: https://arxiv.org/abs/2611.20064
作者: Doyeob Kim
类目: Machine Learning (cs.LG)
*备注: 8 pages, 2 figures, 5 tables

点击查看摘要

Abstract:Voice agents need accurate transcripts soon after a user stops speaking. We propose Decode Early, Verify Late (DEVL), a training-free framework for streaming inference with offline ASR models. The server keeps partial transcripts as drafts and verifies the latest against the complete audio in one forward pass at turn end. Under stated input and decoding conditions, the final equals the model’s greedy offline output in exact arithmetic. We adapt DEVL to four model families. In our five-recognizer comparison, 95-100% of final transcripts are byte-identical to the offline transcripts, WER stays within 0.13 points of offline, and single-stream time to final falls 1.5-7.3x. Serving Qwen3-ASR-1.7B on one A100, DEVL returns finals a median 22-26 ms after turn end at up to 64 streams, 2.7-3.9x sooner than the official streaming mode. It is the only system we tested that meets a 100 ms p99 target, up to 80 streams on Pipecat-STT and 64 on WildASR.

[LG-88] Communication-Efficient Distributionally Robust Federated Averag ing with Multi-Source Data

链接: https://arxiv.org/abs/2611.20060
作者: Yingzhu Liu,Zhongkui Li,Pengcheng You,Ashish Cherukuri
类目: Machine Learning (cs.LG); Optimization and Control (math.OC)
*备注:

点击查看摘要

Abstract:Federated learning trains a shared model from distributed client data. The deployment risk is typically specified through two approximations: each unknown local distribution is replaced by its empirical distribution, and the unknown deployment mixture across clients is replaced by empirical or prescribed training weights. Existing robust methods usually address one approximation at a time through local distributional ambiguity sets or worst-case client mixtures. However, joint treatment of both approximations remains underexplored. To address this gap, we construct a global ambiguity set formed by admissible mixtures of client-wise Wasserstein ambiguity sets, yielding a distributionally robust optimization formulation. The construction allows client-dependent ambiguity radii, preserves client-wise separability of the objective function, and yields a high-probability out-of-sample guarantee for the resulting robust problem. We then introduce a penalty-based approximation of this formulation and develop a communication-efficient algorithm called BiDRO-FedAvg, for which we establish a convergence guarantee. We validate the proposed framework through a market-demand prediction simulation.

[LG-89] MOBO-CAPS: Multi-Objective Bayesian Optimization with Cardinality-Aware Pareto Selection

链接: https://arxiv.org/abs/2611.19999
作者: Hanyang Wang,Juergen Branke,Matthias Poloczek
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Multi-objective Bayesian optimization (MOBO) traditionally aims to approximate the entire Pareto front, but, due to resource constraints, real-world deployment is often limited to testing a small, representative subset of solutions. Hence current methods first approximate the full front before selecting solutions, potentially wasting costly evaluations. We formalize this problem as Multi-objective Bayesian Optimization with Cardinality-aware Pareto Selection (MOBO-CAPS), where the goal is to identify a set of exactly M solutions with high hypervolume. We propose two practical cardinality-aware acquisition strategies: REHVI, a replacement-aware approximation to expected hypervolume improvement, and CHO-UCB, which constructs an optimistic M-point set by optimizing a collective UCB-based hypervolume objective. We prove a sublinear simple regret bound for the batched version of CHO-UCB, and empirically show that cardinality-aware acquisition design improves over standard MOBO followed by post-hoc subset selection across nine benchmarks.

[LG-90] When Zero Is Competitive: A Descriptive Pilot of Transformer and Gate-Graph Surrogates on Random-Angle Quantum Circuits

链接: https://arxiv.org/abs/2611.19985
作者: Mahesh Reddy Pagadala
类目: Machine Learning (cs.LG)
*备注: 9 pages, 1 figure. Descriptive pilot study. Exact experimental evidence archive currently private

点击查看摘要

Abstract:We examine supervised classical prediction of all one-qubit Z and two-qubit ZZ expectation values on noiseless, layered 4-10-qubit RX/RY/RZ-CNOT circuits. In a 63-cell stratified workload with disjoint circuit-template tests, we evaluate one transformer encoder and gate-dependency graph neural networks against literal-zero and training-derived constant predictors. On the Arm K training labels, 44.89% of scalar targets have absolute magnitude below 0.05. The held-out-template zero predictor has equal-instance mean absolute error 0.193399, and none of the evaluated transformer checkpoints or final graph-network endpoints achieves a lower score. A separate product-state positive control achieves MAE 0.007755 versus 0.514811 for zero, but does not explain the failure on the entangled workload. The difference between zero and a training-fitted mean constant under training MSE is only 0.14%, a descriptive point estimate without uncertainty analysis. The transformer’s 815,233-parameter source identity is locally verified; detailed packed-input semantics have not been independently audited. These results constitute a descriptive pilot, not a causal explanation, formal equivalence test, or impossibility claim. Comparable circuits can be regenerated from the workload specification, but exact frozen datasets and model-run evidence remain private.

[LG-91] Practice Is Not Free: Measuring the Embodied Cost of Reset-Free Robot Practice

链接: https://arxiv.org/abs/2611.19979
作者: Vedant Telrandhe,Sushil Sharma
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注: 16 pages, 6 figures. Code, data, logs and videos: this https URL

点击查看摘要

Abstract:When a robot practices without human resets, every attempt still costs wall-clock time, actuator wear, and sometimes a person’s attention, yet the schedulers that decide what to practice next treat attempts as free. We present an open, low-cost testbed for measuring these costs: a self-built SO-101 arm runs two demonstration-trained ACT policies that reset each other, with a vision-based success detector and a ledger that records the time, joint travel, and human interventions of every attempt. We formulate cost-aware practice scheduling, in which task selection is rewarded by measured competence gain net of embodied cost, and specify in advance the comparison that will test it. Its first two conditions, run on hardware for 51 minutes of robot time, show that a fixed retry limit did not reduce the share of execution time lost to failures (36%, against 27% for a feasibility rule; not significant) and saved no robot time. The logs suggest why: failures showed no sign of clustering, so a retry right after a failure was as likely to succeed as any other attempt. The ledger also shows that most of the retry rule’s extra human time came from cubes knocked out of reach rather than from the rule itself, and a cost-aware rule replayed on the logs recovers the sensible choice without a hand-set limit. We also report lessons from building the system, including a coverage effect in which concentrating demonstrations in a smaller workspace raised success there from 8/20 to 10/10.

[LG-92] he structure of variational autoencoder latent spaces and its consequences for representation and optimization

链接: https://arxiv.org/abs/2611.19972
作者: Sergei V. Kalinin
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Variational autoencoders (VAEs) are widely used to build low-dimensional representations of scientific data, with the expectation that regularization yields smooth coordinates for interpolation, generation and optimization. However, applications in microscopy reveal domains, narrow transitions and rare-state islands in VAE latent spaces that smoothness alone does not explain. We show that, for prototype data with squared-error reconstruction and an optimal decoder, the latent-space objective has a direct correspondence with the Flory-Huggins theory of a regular solution, and use it to analyze how the objective, posterior distributions and latent dimensionality organize the representation. Data families occupy domains separated by finite-width walls that broaden with increasing regularization until class separation is lost through posterior collapse. Class centroids behave like particles in a harmonic trap: the prior confines them, and structural dissimilarity sets a soft repulsion that determines which classes become neighbors and how far apart they sit, with separations growing only logarithmically with dissimilarity. Rare states can remain unresolved, form small islands, accumulate within walls or develop intervening domains, depending on their frequency and similarity to common states. The posterior covariance can align domains with the latent axes, but axis alignment does not imply separation of physical factors. Rotations can produce rings and encoder cuts that reflect data symmetry, while insufficient dimensionality can lead to suppression, mixing or folding of additional factors. Comparisons with flow models, deep kernel learning (DKL) and hybrid VAE-DKL models show how these structures change when reconstruction and prediction play different roles. We discuss implications for interpreting data, discovering rare states, generating intermediate states and optimizing properties.

[LG-93] Evaluating Near-Perfect Accuracy Claims in Machine-Learning-Based Leak Detection for Two-Phase Pipelines: A Condition-Blocked Audit

链接: https://arxiv.org/abs/2611.19965
作者: Issah N. Suleiman,Sormeh Serpoosh,Nadine Elkholy,Hicham Ferroudji,Mohammad Azizur Rahman,Matthew Hamilton
类目: Machine Learning (cs.LG)
*备注: 8 figures, 8 tables. Analysis code archived at this https URL

点击查看摘要

Abstract:Reliable leak detection in gas-liquid two-phase pipelines is important for process safety, yet machine-learning studies frequently report near-perfect performance using evaluation protocols that may not represent deployment to unseen operating conditions. This study audits a pressure-based leak-detection model using condition-blocked leave-one-condition-out validation, in which the classifier is tested on a flow condition excluded from training. Accuracy decreased from 93.3% under within-condition evaluation to 59.1% across fifteen unseen-condition tests. Ablation analysis showed that most of this difference arose from the data-splitting protocol, with a smaller contribution from encoding operating conditions as model inputs. Under condition-blocked evaluation, the false-alarm and missed-detection rates increased to 34.1% and 10.2%, respectively, demonstrating that near-perfect reported accuracy did not translate into reliable operation at unseen flow conditions. The reduction persisted across all evaluated classifiers, indicating that it was caused by the evaluation design rather than a particular model architecture. Grouped evaluation of an independent acoustic pipeline dataset produced a comparable reduction toward chance-level performance. These findings show that validation units must represent the conditions encountered during deployment. Condition-blocked evaluation and explicit reporting of operational error rates are therefore necessary before machine-learning leak detectors can support process-safety decisions.

[LG-94] Discovering Programmatic Policies from Reinforcement Learning-Based Traffic Signal Controllers

链接: https://arxiv.org/abs/2611.19951
作者: Lindong Xie,Yang Zhang,Beiyu Song,Zeren Xing,Flora D. Salim,Edward Chung
类目: Emerging Technologies (cs.ET); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Traffic signal control (TSC) is critical for mitigating urban congestion. Deep reinforcement learning (DRL) has achieved strong performance by learning adaptive policies from traffic observations, but learned neural controllers are often opaque, structurally complex, and difficult to generalize across scenarios. We propose PPL-TSC, a programmatic policy learning framework that distills compact and human-readable TSC policies from DRL teachers. PPL-TSC formulates policy learning as an evolutionary search over phase utility functions (PUFs), which score candidate signal phases and select the highest-utility phase for execution. Each generation consists of four steps: policy generation, policy review, policy calibration, and policy evaluation. In the generation step, a large language model (LLM) acts as a semantics-aware evolutionary operator to propose and refine candidate PUFs. Generated PUFs are screened through explicit checks for syntactic validity, mathematical soundness, and consistency with TSC principles, with invalid ones repaired through iterative feedback. Valid PUFs are then calibrated to improve behavioral fidelity to the DRL teacher, and evaluated using a joint score combining teacher fidelity and traffic-control performance to select candidates for the next generation. Extensive experiments across diverse traffic scenarios and representative DRL-based controllers show that PPL-TSC produces compact and human-readable policies, achieves competitive performance for most teacher families, improves transfer generalization, and substantially reduces computational cost. Code is available at this https URL.

[LG-95] CoSyDelay: Cognitive Symbolic Regression for Traffic Delay Estimation

链接: https://arxiv.org/abs/2611.19946
作者: Lindong Xie,Xiexin Zou,Yang Zhang,Beiyu Song,Zeren Xing,Edward Chung
类目: Emerging Technologies (cs.ET); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Trustworthy estimation of traffic delays at signalized intersections is critical for modern transportation systems, as it supports proactive traffic control for congestion mitigation. Existing analytical methods often rely on idealized assumptions that limit their accuracy under complex traffic conditions, whereas data-driven methods may achieve strong estimation performance but usually lack transparent structures, which reduces confidence in their practical application. To address these limitations, we propose CoSyDelay, a cognitive symbolic regression (SR) framework for automatically deriving explicit and physically consistent empirical delay estimation expressions (DEEs). Specifically, CoSyDelay decomposes intersection-level delay estimation into turning-movement-level subproblems, thereby reducing the complexity of the symbolic search. During DEE learning, CoSyDelay constructs domain-informed symbolic inputs from traffic variables. Large language models (LLMs) are employed as symbolic search operators to iteratively generate and refine candidate expression structures, whose movement-specific coefficients are subsequently calibrated using approach-level delay labels. A composite fitness function evaluates the learned DEEs jointly in terms of estimation accuracy and compliance with specified traffic-delay principles. Experimental results on multiple signalized intersections demonstrate that the resulting expressions achieve competitive estimation performance while maintaining physical consistency, thereby supporting trustworthy delay estimation. Source code and datasets are available at this https URL.

[LG-96] AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data

链接: https://arxiv.org/abs/2611.19937
作者: Annan Wang,Yuchen Xie,Yu Cen,Leon Liangyu Chen,Qianyu Julie Zhu,Xin Li,Tianyi Wei,Weisi Lin
类目: Machine Learning (cs.LG); Trading and Market Microstructure (q-fin.TR)
*备注: Project Page: this https URL

点击查看摘要

Abstract:Real-world portfolio management requires converting noisy market observations into coordinated, simultaneous allocations across multiple assets under shared capital constraints. Advances in forecasting models and alpha discovery help identify opportunities. However, traditional pipelines often decouple signal forecasting from portfolio construction, potentially misaligning learning with allocation and execution objectives. We present AllocFlow, an end-to-end flow-matching framework that directly generates joint portfolio allocations from raw multi-asset market histories. To capture dependencies among portfolio positions, AllocFlow employs a cross-asset flow Transformer conditioned on temporal market representations to generate continuous, signed position weights across all instruments jointly. To address the sample inefficiency and instability of learning continuous allocation policies from scratch, we establish a two-stage training scheme: strategy-supervised imitation shapes a structured distribution of candidate portfolios, followed by group-relative reinforcement learning (RL) that refines joint decisions using portfolio-level financial rewards and a frozen policy reference. On hourly cryptocurrency perpetual futures, RL refinement raises the annualized net return of the imitation policy from 27.79% to 51.38% and its Calmar ratio from 0.955 to 4.043 , surpassing the teacher strategy on every metric, and AllocFlow reaches a Calmar ratio of 4.90 on six further months of unseen 2026 data without retraining.

[LG-97] Delayed Heuristic-Behavior-Target Coupling for Off-Policy Reinforcement Learning in Inventory Control

链接: https://arxiv.org/abs/2611.19917
作者: Enhui Li(1),Yuan Cheng(1),Peiting Han(1) ((1) Xi’an Jiaotong University)
类目: Machine Learning (cs.LG)
*备注: 75 pages, 9 figures

点击查看摘要

Abstract:Many reinforcement learning methods tie data collection to the task policy, while coordinating independently optimized behavior and target policies remains challenging. Policies learned by these methods can perform unreliably in long-horizon tasks under uncertainty, such as inventory control, where failures can incur extreme costs. We introduce Heuristic Behavior Policy Optimization (HBPO), an off-policy framework that routes heuristic guidance through persistent behavior learning, coupled with target-policy, updating separately with residual-based advantage estimation and emphatic weighting. Under matched environment interaction budgets on DSIS and its risk-augmented variant DSIS-RB, HBPO recovers many high-cost initial policies and improves most non-collapse comparison groups. Cold starts achieve competitive costs without final cost collapse across all tested seeds. We identify a transmission chain in the delayed heuristic-behavior-target coupling as a key driver of recovery and improvement. Generation-level diagnostics trace heuristic uptake and subsequent selective target responses, while component ablations support the contribution of this transmission mechanism to policy recovery and improvement.

[LG-98] “Dont Leave the Task Unfinished”: Persistence-Driven Violations under Quiet Oversight NEURIPS2026

链接: https://arxiv.org/abs/2611.19911
作者: Seunghyun Lee,Jaehui Hwang,Dongyoon Han,Sangdoo Yun
类目: Machine Learning (cs.LG)
*备注: 10 pages, 4 figures. NeurIPS 2026 Workshop on Dynamic Alignment on Human-AI Coupled System (BiAlign)

点击查看摘要

Abstract:LLM agents are increasingly given long, multi-step tasks to carry out, with the user reviewing little of what happens in between. Persistence, the tendency to keep working until a task is complete, is installed on purpose to improve performance on long tasks. Long-horizon tasks are typically complex, and the user’s instruction can fall short of what the work actually requires, or conflict with the constraints the agent encounters in the task. We study what persistence does on long-horizon operations tasks that cannot be solved without breaking a documented constraint. We find that the agent pursues completion to the point of violating the constraint, and that a critically engaged user lowers the violation rate. Persistence takes violations from 29.2% to 58.3% across four models. Violations stay high wherever oversight goes quiet, whether the user is silent, assenting, or delegating with expressed trust. Only a questioning user, who asks for the basis of each action before allowing the next, brings the rate back to roughly the level measured without persistence. Practitioners should treat high persistence as a setting that carries a safety cost, and should recognize that the cost appears when oversight goes quiet. Safety here is better cast as a property of the human-AI pair than of the model alone. Project page: this https URL

[LG-99] Max-Min Throughput in RSMA-Aided SWIPT for Sustainable Internet of Underground Things: A Deep Reinforcement Learning Approach

链接: https://arxiv.org/abs/2611.19865
作者: Kaiqiang Lin,Mohamed-Slim Alouini
类目: ystems and Control (eess.SY); Machine Learning (cs.LG); Signal Processing (eess.SP)
*备注: 15 pages, 7 figures, and 3 tables

点击查看摘要

Abstract:Unmanned aerial vehicle (UAV)-enabled wireless-powered underground sensor networks (WPUSNs) allow underground devices (UDs) to harvest radio-frequency energy from a UAV via wireless energy transfer (WET) techniques, enabling sustainable Internet of underground things (IoUT) applications. This paper integrates rate-splitting multiple access (RSMA)-aided simultaneous wireless information and power transfer (SWIPT) into UAV-enabled WPUSNs to deliver information and energy to UDs. Within this framework, we investigate a max-min fairness problem that jointly optimizes the UAV transmit power, power allocation and message splitting, power-splitting ratios, and time allocation, aiming to maximize the minimum uplink throughput across UDs subject to per-UD downlink throughput requirements and a UAV energy-consumption constraint. To address this non-convex problem under dynamic underground environments, we develop an off-policy deep reinforcement learning approach based on the primal-dual truncated quantile critics (PD-TQC), where a TQC-based reward critic evaluates the fairness objective, cost critics estimate constraint costs, and an actor learns the resource-control policy with adaptive dual-variable updates. Simulations using in-situ soil volumetric water content measurements from a center-pivot irrigation farm demonstrate that PD-TQC achieves the highest annual average minimum uplink throughput compared with four benchmarks while satisfying the per-frame constraints in over 96% of the evaluated frames, and that this throughput advantage is preserved under different Rician factors, numbers of UDs, and downlink throughput thresholds.

[LG-100] Where Is Rank-1 DiM Refusal Intervenable in Multimodal LM Derivatives?

链接: https://arxiv.org/abs/2611.19826
作者: Chansu Han,Seok-Hwan Choi,Takeshi Takahashi
类目: Machine Learning (cs.LG); Cryptography and Security (cs.CR)
*备注:

点击查看摘要

Abstract:Open-weight model hubs now ship not only chat checkpoints but a long tail of derivatives—fine-tunes, safety guards, vision–language models (VLMs), audio wrappers, and generative heads—yet residual-stream safety audits still often intervene at a single site, usually the text decoder T . We ask a localization question under one shared mean / rank-1 difference-in-means (DiM) extract = intervene protocol, with matched-norm random and empty-rate controls: does causal chat attack lift concentrate at T , or do perceptual encoders, projectors, and fusion paths ( E / P / F ) supply a comparable bypass? On the mixed-input settings we stress-test—HADES adversarial images paired with harmful text prompts, and spoken AdvBench via text-to-speech (TTS) or silence—text DiM at T raises attack success rate (ASR) while the same mean DiM at E / P / F does not; attacks track harmful text rather than image or audio content. Margin-rank and leverage probes on a Vision leaf show that this recipe is extremal at T and near chance at P / E , which pressures a weak-extractor reading of those flat cells for the DiM family we study. Catalog breadth and mapper-free base \to derivative transfer when residual widths match provide supporting coverage under the same gates; generative U-Nets are a separate band under generation judges. The primary claim is this recipe-scoped T vs.\ non-text intervenability asymmetry—intervention accessibility under mean / rank-1 DiM, not a claim that refusal is absent or fully represented outside T (text vs.\ perceptual contrasts and pooling differ by site). On VLGuard Unsafe prompts ( n=200 ; Llama-Guard), the same T vs.\ P / E directionality holds at modest absolute ASR ( 0.030\to0.150 at T ). An intervention-oriented Class \times site ledger records where this protocol still hooks as supporting coverage, not as a universal multimodal bypass.

[LG-101] Foundation Pareto Flow Policy for Multi-Objective Reinforcement Learning

链接: https://arxiv.org/abs/2611.19825
作者: Zhanjiang Yang,Lijun Sun,Yueming Li,Meng Li,Yijun Yang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The overarching goal of reinforcement learning (RL) is to efficiently optimize the agent’s policy for various environments, tasks, dynamics, and even user preferences, which often consist of multiple and potentially conflicting rewards/objectives. This necessitates the learning of Pareto-optimal policies that can navigate different trade-offs among objectives. However, existing multi-objective RL (MORL) methods, whether trained online or offline, are typically limited to task- or objective-specific policy learning and do not address generalization of learned trade-offs to unseen tasks/objectives. In this work, we propose the Foundation Pareto Flow Policy (FP2) for offline MORL in continuous robotic control, inspired by the pre-mid-post-training paradigm underlying recent foundation models. Specifically, we instantiate FP2 as a multi-modal diffusion transformer equipped with a likelihood-free flow matching loss, enabling unified, scalable, and instructable learning of diverse trade-off trajectories. We further develop a three-stage training pipeline to gradually optimize the policy: (1) pre-training on a large-scale MORL dataset, (2) mid-training on a carefully curated high-quality dataset, and (3) post-training on few collected trajectory data from new tasks and trade-offs. Empirical results demonstrate that FP2 outperforms previous state-of-the-art methods on most tasks of the D4MORL benchmark, achieving superior zero-shot interpolation and extrapolation on complex Pareto manifolds as well as efficient few-shot adaptation of the learned trade-off prior to new continuous-control tasks. Code is available at this https URL.

[LG-102] Conditional Optimal Transport and Triangular Maps for Imitation Learning

链接: https://arxiv.org/abs/2611.19815
作者: Mohammad Al-Jarrah,Yuhang Mei,Karen Leung,Bamdad Hosseini,Amirhossein Taghvaei
类目: Machine Learning (cs.LG)
*备注: 38 pages, 5 figures

点击查看摘要

Abstract:We introduce a unified formulation of imitation learning from demonstrations as a conditional sampling problem solved by triangular transport: a policy is the second block of a triangular map that transports a reference distribution to the conditional distribution of the action chunk given the observation history. Unlike in Diffusion Policy, where it is a fixed Gaussian, the reference becomes a design variable; we take it to be the empirical distribution of the demonstrated action chunks and introduce three instances of the formulation: an optimal transport policy (OTP), a flow matching policy (FMP), and a stochastic interpolant policy (SIP). On the Push-T and robomimic Square tasks, the three policies are competitive in task performance with Diffusion Policy and Consistency Policy while using considerably smaller networks and shorter training, and the informative reference accounts for a substantial part of their performance when few network evaluations are used. The three instances offer different trade-offs rather than a single best method: OTP generates an action chunk with a single evaluation of the smallest network (0.23 million parameters on Push-T) and is suitable for settings in which inference time and model size matter, whereas SIP attains the highest scores among the proposed policies at a higher cost.

[LG-103] Barron Approximation and locally optimal weight densities for shallow neural networks

链接: https://arxiv.org/abs/2611.19798
作者: G. Welper
类目: Machine Learning (cs.LG); Numerical Analysis (math.NA)
*备注:

点击查看摘要

Abstract:Mean field theories provide optimization results for shallow neural networks by analyzing the weight distribution in infinite width limits. Corresponding results for finite sized networks are obtained by particle approximations, for which sharp quantitative bounds are still an open problem. In this paper, we consider a modified mean field loss, which allows a more fine grained control over finite sized networks. We prove convexity and equidistribution properties, which directly lead to Barron type approximation results for networks sampled from local loss minimizers. We demonstrate that particle approximations of the new loss function naturally lead to gradient descent methods with dropout type regularization. Subjects: Machine Learning (cs.LG); Numerical Analysis (math.NA) MSC classes: 41A46, 65K10, 68T07 Cite as: arXiv:2611.19798 [cs.LG] (or arXiv:2611.19798v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19798 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-104] he Sense Paradigm for World Models

链接: https://arxiv.org/abs/2611.19796
作者: Haidong Hu
类目: Machine Learning (cs.LG)
*备注: 11 pages, 4 figures, 4 tables

点击查看摘要

Abstract:Latent world models predict the future in an embedding space rather than in pixels. An encoder trained only to predict collapses and maps every observation to the same point, so existing models add a mechanism against collapse, such as a target network, a variance penalty or a distribution penalty. These mechanisms keep the embedding from collapsing, but they leave its distances without meaning, and the model cannot tell a reversible transition from an irreversible one. We propose Sense, a single cross-entropy loss over the directed distance d(x\to y)=\sum_i \mathrmReLU(y_i-x_i) that replaces these mechanisms. The loss avoids collapse on its own, and since the encoder no longer needs the predictor, the predictor is fitted afterwards to fixed embeddings by plain regression. On four pixel environments, the return distance learned by Sense detects irreversible transitions with an AUC between 0.95 and 1.00, where LeWorldModel stays between 0.39 and 0.59. On TwoRoom, the navigation benchmark of LeWorldModel, planning with Sense reaches the goal in 100% of episodes against 91% for LeWorldModel. As the goal moves farther away, the planning failure rate of every baseline grows, while that of Sense stays near zero in Cup, Cliff and Sokoban and is the lowest of all methods in Pendulum.

[LG-105] How Should Parallel Langevin Chains Share Noise? NEURIPS2026

链接: https://arxiv.org/abs/2611.19784
作者: Ryan Farell
类目: Machine Learning (cs.LG)
*备注: 48 pages, 6 figures, 36 tables. NeurIPS 2026 (poster). Code: this https URL

点击查看摘要

Abstract:Parallel Langevin sampling usually draws an array of independent Gaussian vectors, one row per chain. The average row drives the ensemble mean, with variance 1/K per coordinate for K chains. Yet the unadjusted Langevin algorithm (ULA) specifies only each row’s distribution. Subtracting the column means and rescaling, as in classical antithetic sampling, makes that variance zero. Each chain keeps its ordinary ULA path law for any finite run. For quadratic targets we derive the joint covariance of linear estimates computed from the ensemble mean, including coefficient estimates, time averages and predictions. The identity separates contributions from initial states and Gaussian draws. The coupling changes the second through one scalar, which is zero whenever the draws sum to zero across chains. For nonquadratic targets we bound the remaining variance. Centering has a price: for a quadratic form in the mean-subtracted Gaussian state, the ensemble average has variance 1/(K-1) relative to one chain, compared with 1/K under independent draws. On three tempered logistic posteriors, eight centered chains reduced the variance of the posterior-mean estimate, summed over coordinates, by two to three orders of magnitude compared with eight independent chains. They also beat 128 independent chains on variance and on variance times sampling time. Under Metropolis correction the reduction shrank to about three- to eightfold: rejected draws no longer balance accepted ones. The measured variance ratios were consistent with a leading-order prediction from rejection frequency and noise energy.

[LG-106] PISG: Constraint-Aligned Signal Amplification for Diffusion-Based Combinatorial Optimization NEURIPS2026

链接: https://arxiv.org/abs/2611.19768
作者: Yuming Zhang,Xianchen Zhou,Hongxia Wang
类目: Machine Learning (cs.LG); Optimization and Control (math.OC)
*备注: Accepted at NeurIPS 2026 (poster). 32 pages. Code: this https URL

点击查看摘要

Abstract:Diffusion-based solvers have become a prominent approach to neural combinatorial optimization. At inference time, they assign each decision variable a prediction score. A greedy decoder ranks candidates by these scores and checks hard constraints to ensure feasibility. Top-ranked candidates should be both promising and jointly feasible, making prediction-constraint alignment critical. When this alignment is weak, greedy decoding skips constraint-violating candidates and falls back to lower-ranked alternatives, degrading solution quality. Existing test-time methods improve search, sampling, or decoding, but do not directly strengthen this alignment before decoding. We observe that trained denoisers already carry constraint-aligned structural signals. Building on this observation, we propose Perturbed Instance-Structure Guidance (PISG), a training-free method that amplifies these signals at test time. At each denoising step, PISG constructs a structure-degraded prediction and extrapolates the unperturbed prediction away from it. The resulting guided prediction improves candidate rankings before constrained decoding. Signal analysis shows that this guidance concentrates on constraint-critical variables rather than uniformly rescaling scores. Under greedy decoding and without downstream refinement, PISG improves all 16 TSP solver-scale pairs across DIFUSCO, T2T, FastT2T, and StruDiCO, and all four MIS-ER solvers, with MIS-ER solved-cost gains up to +4.8%. The gains persist on most large-scale settings after 2-opt, and a preliminary study extends PISG to a Transformer-based solver. These results identify prediction-constraint alignment as an effective test-time enhancement direction for diffusion-based combinatorial optimization. Code is available at this https URL.

[LG-107] Sample-Efficient Ensemble Reinforcement Learning at Scale

链接: https://arxiv.org/abs/2611.19733
作者: Yoav Feigenbaum,Dmitrii Krylov,Roy Fox
类目: Machine Learning (cs.LG)
*备注: 10 pages, 3 figures

点击查看摘要

Abstract:Ensemble methods such as MeanQ have been shown to improve the sample efficiency of relatively small model-free deep reinforcement learning (RL) architectures, including DQN and similarly sized variants such as Rainbow. However, to our knowledge, the effects of ensembling on large-scale architectures such as Bigger Better Faster (BBF) have not been studied. In this work, we present MeanQ-BBF, a hybrid model-free algorithm that combines the principles of MeanQ and BBF. Through statistically rigorous evaluation on the Atari 100K benchmark, we show that in contrast to naive scaling of the network size of BBF, which degrades performance, MeanQ-BBF is able to leverage the increase in model capacity to improve performance over BBF. Moreover, MeanQ-BBF does this in a computationally efficient way by organizing the added parameters into largely independent learners. Finally, we discuss the performance-cost tradeoff in the context of practical sample-efficient RL.

[LG-108] Found but Not Fixed : From Software Vulnerability Discovery to Fixes with Iterative LLM Pipelines

链接: https://arxiv.org/abs/2611.19719
作者: Youssef Gamal,Mohamad Fakih,Halima Bouzidi,Rahul Dharmaji,Gustavo Quiros Araya,Oluwatosin Ogundare,Mst Ayesha Siddika,Mohammad Abdullah Al Faruque
类目: oftware Engineering (cs.SE); Cryptography and Security (cs.CR); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Software vulnerabilities persist as a critical threat despite the deployment of sophisticated detection tools and AI assistants. While detection has advanced, the subsequent code repair phase remains a primary bottleneck; manual intervention is inherently slow, costly, and prone to introducing secondary regressions and issues. In this paper, we investigate whether an iterative, LLM-based pipeline can provide a robust solution for the automated repair of vulnerable functions in real-world code. We propose a generalized framework that leverages the reasoning capabilities of Large Language Models (LLMs) through a multi-stage feedback loop. We benchmark this approach against state-of-the-art models, including GPT-4o, GPT-5.1, and Gemini 3 Flash, demonstrating that such a pipeline can achieve high-quality code repair. To support the broader security community, we provide our evaluation framework, fine-tuned weights, and comprehensive experimental dataset on our website: this https URL

[LG-109] Multiaccuracy as an Error Benchmark for Learning with Distribution Shift

链接: https://arxiv.org/abs/2611.19701
作者: Konstantinos Stavropoulos
类目: Machine Learning (cs.LG); Data Structures and Algorithms (cs.DS)
*备注:

点击查看摘要

Abstract:Agnostic learning has a well-developed theory of computational and statistical complexity. For learning with distribution shift, even the choice of an appropriate error benchmark remains challenging. Testable learning with distribution shift (TDS) was introduced to develop computational foundations for this setting: using labeled training data and unlabeled test data, the learner must either produce a predictor with low test error or detect distribution shift. However, standard agnostic TDS guarantees allow a constant-factor amplification of the optimum error even when the training and test distributions coincide. They therefore do not directly recover the guarantee of classical agnostic learning. We introduce shift multiaccuracy, a stronger error benchmark inspired by multiaccuracy from algorithmic fairness. This yields a TDS formulation that becomes equivalent to multiaccuracy when there is no shift. Using a recent nontrivial reduction from agnostic learning to multiaccurate prediction from prior work, we obtain the first TDS framework that formally generalizes agnostic learning through an efficient reduction preserving the training marginal. The framework retains existing algorithmic upper bounds and inherits computational lower bounds from agnostic learning. Moreover, we establish a direct connection between our model and identity testing, yielding the first statistical characterization of agnostic TDS learning, in terms of Rademacher complexity. More broadly, we argue that shift multiaccuracy might be of independent interest in unsupervised domain adaptation, as it leads to sharper out-of-distribution generalization guarantees, and can be minimized directly, since its definition does not involve the unobserved test labels. Subjects: Machine Learning (cs.LG); Data Structures and Algorithms (cs.DS) Cite as: arXiv:2611.19701 [cs.LG] (or arXiv:2611.19701v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19701 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-110] Elicited Adaptation: Auditable Localized Fairness via Pairwise Queries

链接: https://arxiv.org/abs/2611.19686
作者: Shrey Shah
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:How much must a predictor’s scores change when a community’s elicited fairness standard changes? Similarity and tolerance judgments define protected comparisons and acceptable score gaps. We study minimum uniform score movement on a fixed pool of n profiles. For scores satisfying a source standard, we characterize the exact cost of retaining that standard during adaptation: runs of consecutive target constraints along shortest paths determine the optimal amplification factor. Its sharp worst case is \lfloor n/2\rfloor , and a full cost curve describes finite target-only repair budgets. Committing before the target is known can cost n-1 times the worst repair after revelation, even within one fixed convex ball of two-dimensional Mahalanobis metrics. Positive edge-budget floors cap both factors; bounded geometry also limits the commitment cost. At two interior targets of the construction attaining the n-1 bound, similarity queries alone yield a feasible 2 -approximate repair for tolerance \eps\le1/[2(n-1)] , reducing movement relative to common repair by a factor linear in n . Feasible (1+\alpha) -optimal repair on this family has worst-case tolerance-query complexity \Theta(\alpha^-2n^2\log(1/\delta)) under a fixed-slope logistic model. The analysis separates the information needed for a substantial adaptation benefit from the precision needed for nearly optimal movement. Elicitation studies and training comparisons examine tolerance choice and the relation between numerical compliance and predictive utility.

[LG-111] RIFT: Reward Improvement Field Transfer for Cross-Domain Offline RL

链接: https://arxiv.org/abs/2611.19681
作者: Jinhee Won,Mardavij Roozbehani,Munther Dahleh
类目: Machine Learning (cs.LG)
*备注: 39 pages, 4 figures, 17 tables (10 pages main text; remainder references and appendices)

点击查看摘要

Abstract:Learning from limited target data is a central challenge in cross-domain offline reinforcement learning. Abundant source data can help, but dynamics differences can make direct transition reuse unreliable. We propose Reward Improvement Field Transfer (RIFT), which instead learns how low- and high-reward source transitions differ. RIFT encodes source and target transitions in a shared latent space, pairs low- and high-reward source transitions using optimal transport, and fits an affine field to their displacements. It applies this field to low-reward target anchors and decodes the results to augment the target dataset. The approach builds on the hypothesis that related domains can share useful low-to-high reward structure despite different dynamics. We bound synthesis error under shared class geometry, identify conditions favoring target anchoring over source-high transplant, and derive a conditional connection to downstream return. Experiments on dense-reward locomotion tasks under gravity, morphology, and friction shifts show improved aggregate performance over existing baselines.

[LG-112] Attacker Diversity Reduction in Adaptive LLM Red Teaming

链接: https://arxiv.org/abs/2611.19680
作者: Chris Tava
类目: Cryptography and Security (cs.CR); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Reinforcement learning can make automated red-teaming more successful while narrowing the attacks it produces. Co-training an attacker and a defender with group-relative policy optimization (GRPO) has been reported to reduce attacker diversity. This experiment reproduces attack diversity reduction using three runs with a Qwen2.5-14B attacker and a Phi-4-mini defender trained by direct preference optimization (DPO). The attacker receives a success-only reward and uses a length-unbiased loss, bounded updates, warm-up and ratio clipping. The attacker is evaluated on 26 AdvBench behaviors. Training ended after 12 rounds, following a decision based on development checkpoints, so the reported intervals are nominal. Attack Diversity Retention (ADR) compares the within-behavior of the trained attacker spread with the untrained attacker. In order for an attack diversity to be consider as having been reduced, a criterion requires the ADR to be below a 0.75 target with a 95% confidence interval together with continued attacker effectiveness. Therefore, if an attacker narrows its attack vector by 25% consistently its considered less diverse. After 24 attacker and 12 defender LORA-based updates, ADR for the three runs were 0.63, 0.61 and 0.63, respectively. Each run met the criterion at round 12; the earlier checks at rounds 4 and 8 fell short. Despite attacks being less diverse, attack success continued to improve as the attacker was updated after each training run.

[LG-113] Score-Based Stochastic Closure Modeling via Offline Indirect Learning and Online Data Assimilation

链接: https://arxiv.org/abs/2611.19676
作者: Xinghao Dong,Huchen Yang,Jin-Long Wu
类目: Machine Learning (cs.LG); Computational Physics (physics.comp-ph); Fluid Dynamics (physics.flu-dyn)
*备注:

点击查看摘要

Abstract:Stochastic closure modeling is an emerging area in reduced-order modeling of large and complex (e.g., multiscale, multi-physics) systems, for which the computationally affordable reduced-order physical descriptions (e.g., digital twins for additive manufacturing, world models for AI agents) in numerical simulations inevitably lead to information losses that cannot be well addressed by classical deterministic closures. However, learning stochastic closures typically requires closure data that are difficult to obtain from measurements or often prohibitively expensive to compute using first-principles simulations. In this work, we develop a stochastic closure modeling framework that builds on score-based diffusion models, leveraging indirect offline training and online data assimilation without requiring the direct data of closures. Numerical experiments on two-dimensional turbulent flows demonstrate that indirect offline learning captures key closure statistics and improves predictions of system state. In addition, combining the learned closure with known physical dynamics leads to better generalization capability across different regimes than purely data-driven models that directly predict system state. Compared to standard data assimilation methods with a Gaussian assumption, the posterior-score assimilation better represents non-Gaussian features of the learned closure, and its combination with state-level data assimilation further improves prediction of the system state. The proposed framework enables stochastic closure modeling in realistic and challenging scenarios of complex dynamical systems in mechanics and engineering problems, where non-Gaussian statistics of model-form uncertainties are non-negligible (e.g., due to the nonlinearities of the system dynamics and/or the observations) and true closure data are not available, for both offline learning and online assimilation.

[LG-114] Mesoscale Neural Operators: Learning Nonlinear Constitutive Laws from Molecular Dynamics for Scalable Fracture Predictions

链接: https://arxiv.org/abs/2611.19658
作者: Yue Yu,Jihong Wang,Stewart Silling,Ziwei Yang
类目: Machine Learning (cs.LG); Materials Science (cond-mat.mtrl-sci)
*备注:

点击查看摘要

Abstract:Molecular dynamics (MD) simulations can capture the nanoscale physics of fracture at the atomistic level, but their computational cost prohibits predictions at the length scales and time scales needed for many applications. Machine learning approaches, on the other hand, have been proposed as heuristic surrogates for material fracture, yet they cannot guarantee the physical consistency of the predicted response. To address this shortcoming, we propose the mesoscale neural operator (MNO): a data-driven nonlocal constitutive model that is learned directly from MD loading and displacement data. After training, it can be used in continuum fracture simulations on much larger domains and grid sizes. The key idea of the MNO is to exploit the built-in characteristic length of a nonlocal model, called the horizon, to travel across scales. Physical consistency across scales is obtained by building two structural ingredients into the MNO: a normalized kernel that renders the strain energy density invariant under horizon rescaling, and a rescaled critical bond stretch that keeps the energy release rate consistent across scales. Through tests on synthetic data, we show that the MNO and its learning workflow indeed discover the nonlocal continuum multiscale model in the form of the nonlocal kernel and the nonlinear constitutive relation. We then apply the framework to coarse-grained MD data and find that in a test problem, the learned model matches the MD strain energy density and critical energy release rate within 3%. It also predicts the peak transmitted forces within 6% in out-of-distribution tension tests. As a further demonstration, the model is applied to a plate with 4000 times the area of the training specimens and containing an initial defect. In this problem, the model reproduces the main features of dynamic brittle fracture, including branch and cascading branch networks.

[LG-115] With the Right Representation Classical Machine Learning Rivals Deep Models for Olfactory Mixture Similarity

链接: https://arxiv.org/abs/2611.19650
作者: Ze Dong,Yi Wang,Ambuj Tewari
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Modeling olfactory mixtures remains a critical challenge in machine olfaction because the task requires reasoning over relationships between sets of odorants rather than over individual molecules in isolation. However, prior comparisons between deep and classical approaches also varied how the relationship between the two mixtures was represented, since the classical baselines primarily relied on weak representations that described each mixture independently and provided little direct signal about how the mixtures differed. In this work, we revisit classical learning using explicit symmetric comparison features on unordered pairs of mixtures, evaluating these representations with XGBoost and kernel ridge regression (KRR). We find that XGBoost with explicit comparison features achieves performance competitive with specialized deep mixture models on the main benchmark. Ablation studies show that coordinate-wise absolute differences alone provide strong predictive performance, while combining them with sums and products yields further improvements. KRR also benefits substantially from explicit comparison features, with further gains from appropriate kernel choice and preprocessing. Overall, these results suggest that explicitly encoding how two mixtures differ is an important ingredient for classical machine learning in low-data olfactory mixture prediction.

[LG-116] Multi-Objective Causal Bandits: Minimal Intervention Space and Policy-Level Learning NEURIPS26

链接: https://arxiv.org/abs/2611.19634
作者: Muhammad Qasim Elahi,Mahsa Ghasemi,Murat Kocaoglu
类目: Machine Learning (cs.LG)
*备注: To appear in Proceedings of NeurIPS 26

点击查看摘要

Abstract:Many decision-making problems involve multiple objectives, where actions affect several performance metrics, fairness criteria, or safety constraints. We study multi-objective causal bandits with multiple reward nodes, where actions correspond to interventions on subsets of variables in a causal graph and performance is evaluated through Pareto optimality. Existing approaches to multi-objective bandits often rely on pairwise Pareto dominance or optimize a fixed scalarization. However, pairwise comparisons alone can fail to identify actions dominated by randomized policies, whose expected reward vectors lie in the convex hull of action-level rewards. Thus, an intervention may be suboptimal even if no single intervention dominates it, because it can be dominated by a randomized policy over other interventions. We formalize this policy-level phenomenon and show that it has consequences for both causal action-space reduction and online learning. First, we provide a necessary-and-sufficient graphical characterization of possibly Pareto- optimal minimal intervention sets. Our characterization yields a minimal, sound, and complete candidate intervention family from the graph alone, correcting prior formulations that may include intervention sets that are never Pareto-optimal under any compatible structural causal model. Second, over this reduced intervention space, we design an online UCB-style algorithm that eliminates actions using dominance tests against convex combinations of other actions, and prove logarithmic gap-dependent Pareto regret. Finally, we study constrained multi-objective causal bandits, where feasibility and optimality may be realized only with randomized policies, and develop an online UCB-style constrained bandit algorithm with sublinear regret and constraint-violation guarantees.

[LG-117] Before the Word: Name Information in Description-Derived States of Small Language Models With and Without Naming Failure

链接: https://arxiv.org/abs/2611.19609
作者: Fernando Cardenas Piepereit
类目: Machine Learning (cs.LG)
*备注: 22 pages, 9 tables. Code and data: this https URL

点击查看摘要

Abstract:The tip-of-the-tongue (TOT) state, knowing an object but not its name, motivates this study as an analogy only. We ask whether small language models form a name-predictive state from a name-free description, and whether it differs when naming fails but description selection succeeds. In an internally pre-registered study (a frozen document, no external registry) of Qwen2-0.5B, SmolLM2-1.7B and Phi-3-mini with 200 items, we combine a similarity read-out, a Patchscopes-style read-out and activation patching with matched controls. H1 (name information in the state) was supported in all nine model-by-description-type combinations; the agreement rule was met by the similarity measure together with patching, while Patchscopes stayed near chance. Post-hoc robustness checks after internal review (logit lens, single-position patch, manipulation check) temper this reading: name information in middle layers is clearest under the multi-position patch and the similarity measure, and weak under single-position patching at 50% depth. H2’ (description-paraphrase consistency of name-free states not lower in failure items) is mixed: the SmolLM2 deficit test is significant under the pre-registered specification but not robust; Phi-3 non-inferiority holds across margins and a word-overlap covariate, although its bootstrap interval includes zero; Qwen2-0.5B is inconclusive. A capacity explanation remains untested. We report vehicle-level signals only, not meaning.

[LG-118] Explicit Regularization for Flow Matching

链接: https://arxiv.org/abs/2611.19587
作者: Lars Ruthotto,Eran Treister
类目: Machine Learning (cs.LG); Numerical Analysis (math.NA)
*备注: 40 pages (main text pp. 1-24, references pp. 24-26, appendices A-C pp. 27-40), 4 figures, 11 tables

点击查看摘要

Abstract:Learning the underlying probability distribution from samples is a core problem in statistics and applied mathematics. Generative models based on diffusion and flow matching learn a velocity field whose flow transports a known distribution to the data distribution. They produce novel, realistic samples, although the empirical learning problem is ill posed near the data distribution. Its minimizer transports the samples onto the training data, and small perturbations of the data change the velocity field strongly near the end of the flow. The observed generalization is often attributed to implicit regularization by the network parameterization, early stopping of the training, and the time integration during sampling. Explicit regularization has advanced the theory and computation of many inverse problems and has recently been proposed to prevent memorization in generative models. With a regularizer on the magnitude of the velocity, however, the field approaches the memorizing velocity unless the regularization parameter grows to infinity fast enough, and it then stays bounded by a constant that is small only for large parameters. We show that penalizing the derivatives of order k of the velocity renders the D -dimensional empirical problem well posed uniformly in time if 2k D and the training points are in general position. Our experiments in one and two dimensions discretize the velocity with splines and use direct solvers for the regularized least-squares problems to exclude optimization error. Refining the discretization in space and time changes the error of the Hessian-regularized flows only marginally, whereas without the regularizer the error tends to that of memorization. These flows generalize, with errors comparable to or lower than those of well-tuned kernel density estimates.

[LG-119] SLiDE: Structured Linear Dynamics for Forecasting with Exogenous Inputs NEURIPS2026

链接: https://arxiv.org/abs/2611.19586
作者: Sebastian Pütz,Theodore Glavas,Benjamin Schäfer,Boris N. Oreshkin,Mark Coates
类目: Machine Learning (cs.LG)
*备注: Accepted to NeurIPS 2026

点击查看摘要

Abstract:Accurate time series forecasting with exogenous inputs is critical across domains, including energy and retail, yet modern deep learning models often overfit to correlations that do not generalize under shifts in these inputs. We propose SLiDE, a Koopman-inspired architecture for forecasting with exogenous inputs that imposes structure on latent temporal dynamics. SLiDE approximates nonlinear system evolution using a learned linear operator in a latent space, combining (i) a history encoder that reconstructs the current latent state from past targets and inputs with (ii) a shared linear rollout driven by future exogenous variables. By enforcing a common linear recurrence across encoding and prediction, SLiDE aligns how past information is encoded with how future states evolve, reducing parameterization and preventing reliance on spurious correlations in the input history. Empirically, SLiDE achieves state-of-the-art accuracy on real-world benchmarks, including electricity price forecasting and retail demand, while maintaining a computational footprint comparable to lightweight MLP architectures. SLiDE is especially effective at generalizing to shifted exogenous inputs, reducing MSE by 20% on electricity price samples with out-of-range exogenous values and MAE by 30% in a synthetic transfer setting where the same dynamical system is driven by an unseen exogenous process. Ablations confirm the importance of both the structured history encoder and the linear latent rollout, suggesting that structured latent linear dynamics provide a useful inductive bias for forecasting with exogenous inputs.

[LG-120] Bellm an Residual Bracketing: Resolving the Sample Complexity of Discounted and Averag e-Reward RL

链接: https://arxiv.org/abs/2611.19566
作者: Matthew Zurek,Yudong Chen
类目: Machine Learning (cs.LG); Information Theory (cs.IT); Optimization and Control (math.OC); Probability (math.PR); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:We study the sample complexity of the empirical planning or plug-in approach for reinforcement learning (RL) with a generative model. We show the discounted plug-in approach finds an \varepsilon -optimal policy with O(SA (1-\gamma)^-3 \varepsilon^-2 \log(SA/\delta)) samples, where S, A, \gamma, \delta are the state and action sizes, discount factor, and failure probability, respectively. In the average-reward setting, we show that discounted plug-in and direct (undiscounted) plug-in achieve O(SAH\varepsilon^-2\log(SA/\delta)) and O(SAD\varepsilon^-2\log(SA/\delta)) complexities, respectively, where H upper-bounds the optimal bias span and D is the diameter. Combining with our lower bounds for nonadaptive generative models, these results resolve the minimax sample complexity of discounted RL and average-reward RL — in both the span- and diameter-based classes — up to constant factors. Our main technical contribution is a novel Bellman residual bracketing technique which improves the efficiency of prior absorbing-MDP-based concentration analyses.

[LG-121] Random Cut Forests Games and Diffusion over Approximately Nested Anisotropic Local Geometry

链接: https://arxiv.org/abs/2611.19548
作者: Sudipto Guha
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Just as the data in a dynamic stream changes, questions asked of the stream also change in intent: what to hold constant, which counterfactuals to consider, and so forth. Energy-based learning has been celebrated for its uniform approach to losses and estimation, but interpretability remains challenging. Tree-based methods, on the other hand, offer high interpretability but often rely on bespoke, disjointed estimators. Neither paradigm readily supports changing questions on streaming data. In this paper we take a step towards combining energy, interpretability, and streaming. Random cut forests retain an evolving geometry of the data. We study this geometry as the basis for diffusion dynamics, prediction, explanation, and games. We show that several existing scores are escape probabilities of random walks on the trees. The induced affinity is realized exactly by a sparse electrical network on each tree, giving a single energy calculus, a Hamiltonian, maintained along one path per update. This connects an estimate, its supporting evidence, and the response to intervention through the same boxes, gaps, masses, and paths. Different questions correspond to different clamp decisions on the energy. Fixing observed values gives harmonic prediction with Gaussian uncertainty. Fixing a partition gives normalized cuts whose cost is attributed exactly to oriented faces. Releasing clamps gives coalition games, in which collusive displacement measures the field response. With noise, the same energy gives diffusion dynamics whose relaxation rates are read from the boxes. Axis-parallel cuts localize at a rate set by affine rather than ambient dimension, and this control passes to the averaged energy. Across a forest the boxes are only approximately nested; common reference observations align the trees. A rotating star with added noise illustrates the measurements. Subjects: Machine Learning (cs.LG) Cite as: arXiv:2611.19548 [cs.LG] (or arXiv:2611.19548v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19548 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-122] Probing SAM Perturbations Through Statistical Replay NEURIPS2026

链接: https://arxiv.org/abs/2611.19480
作者: Ting Han,Jens Kleesiek,Michael Kamp
类目: Machine Learning (cs.LG)
*备注: Accepted at NeurIPS 2026 Workshop: AXIOM - Foundations of Efficient Deep Learning

点击查看摘要

Abstract:Sharpness Aware Minimization (SAM) updates model parameters using gradients evaluated at adversarially perturbed weights, but it remains unclear which aspects of this procedure are sufficient to reproduce its generalization behavior. In our ResNet 18 experiments on CIFAR 10, simple curvature regularizers based on the Hessian trace and largest eigenvalue remain below SAM in validation accuracy, motivating a direct examination of its perturbations. We introduce statistical replay, a diagnostic that summarizes perturbations from reference SAM trajectories using elementwise means and standard deviations over local windows and constructs new perturbations during independent SGD training. Pure replay uses only these summaries, whereas mixed replay adds isotropic Gaussian noise. Pure replay does not reproduce SAM’s validation behavior, while mixed replay and pure isotropic perturbations can approach its validation performance at selected scales without reproducing its training dynamics. An equal radius random direction within SAM performs similarly to SGD, and source, temporal, and layer interventions further suggest that replay behavior depends on perturbation construction. Together, these preliminary results weaken several simple explanations of SAM while leaving an important part of its mechanism unexplained.

[LG-123] Grokking Can Be Steered: Selecting Desirable Solutions By Choice Of Learning Rule

链接: https://arxiv.org/abs/2611.19477
作者: Pascal Sager,Benjamin Grewe,Pau Vilimelis Aceituno,Thilo Stadelmann
类目: Machine Learning (cs.LG)
*备注: 40 pages, 6 figures

点击查看摘要

Abstract:During grokking, a network transitions from memorization to generalization, yet what shapes the neural implementation (circuit) of the generalizing solution and specifically the number of participating neurons (its size) remains unclear. We find that the learning rule, the combination of credit assignment algorithm and optimizer, controls this: On modular addition, every learning rule we study that groks recovers the same Fourier-based algorithm, yet implemented by circuits that differ in size. Why? A neuron stays in use when its credit is large relative to the non-gradient term of its own update, and this single condition unifies three routes to unused neurons: the coupled weight decay, Adam’s \epsilon and the decay of a trainable per-neuron gain. Equalizing the magnitude of credit assigned to each neuron, or scaling each neuron’s weight decay by its credit, expands the effective circuit size, measured from the outgoing weights, from 60% to 99% of the layer’s neurons at similar training and test accuracy. We vary which neurons share one equalization factor, from the whole layer to individual neurons, and find that the circuit stays concentrated for groups of eight or more and fully distributes only when every neuron has its own factor. Credit magnitudes are usually set by the normalization of standard optimizers: Nero, which normalizes per neuron, distributes the computation, whereas Adam and seven others, including RMSProp, Muon, and Shampoo, use other normalization schemes and retain small, concentrated circuits. Distributed circuits tolerate neuron deletion better, while concentrated circuits can be pruned further and transfer about twice as fast to a task that recombines the same features, largely by recruiting their unused neurons. Learning rule choice thus steers circuit size, trading robustness against compressibility and transfer speed without changing the learned algorithm.

[LG-124] Successive Convexification in Learned World Models for Safe Offline Planning

链接: https://arxiv.org/abs/2611.19466
作者: Samuel Younger
类目: Machine Learning (cs.LG); Robotics (cs.RO); Systems and Control (eess.SY)
*备注: 28 pages, 13 figures. Under review at L4DC 2027

点击查看摘要

Abstract:World models learned from offline data allow an agent to plan without further interaction with the environment, but the sampling-based planners used with them handle safety constraints only by scoring or rejecting sampled action sequences. We instead pose safe planning in the latent space of a learned world model as a successive convex program (SCP), in which a learned discrete-time control barrier function is imposed as a constraint at every step of the horizon. SCP requires smooth dynamics, so we build the networks from smoothed MaxMin activations and spectrally normalised layers, and bound the resulting linearisation error in terms of the Lipschitz bound and the smoothing width. As SCP is a local method, we start it from a sampled plan. On two Safety-Gymnasium navigation tasks with offline data, this planner has fewer violations than MPPI and CEM drawing 128 times as many samples at equal wall-clock time in our main comparison, though not significantly, and earns greater return at longer horizons, and is much safer than SCP without a sampled start. SCP converged more often on Lipschitz-bounded models. Unlike sampling, it also reports at each step whether its plan is a feasible, first-order optimal KKT point of the planning problem in the model, or that no feasible plan was found nearby. Our implementation is publicly available at this https URL.

[LG-125] Predicting Fast and Slow: SSD Already Knows Which Guess Is Wrong

链接: https://arxiv.org/abs/2611.19457
作者: Philipp Bogdan
类目: Operating Systems (cs.OS); Machine Learning (cs.LG); Performance (cs.PF)
*备注: 3 pages, 3 figures

点击查看摘要

Abstract:Heimdall, a learned controller, calls slow on 52% of an SSD’s reads when 30% are slow. Uncertainty estimators, the remedy prescribed for its wrong guesses, cannot find them: nine of them do no better than a coin. The drive can. On one trace, 2% of the reads of 4 to 8 KiB that found less than 4 KiB outstanding while the drive was answering came back slow, and Heimdall called 96% of them slow. A 4,096-entry table of how the drive’s recent reads came back, untrained and 21 ns a decision, ranks the wrong guesses at a median AUROC of 0.68 on 57 held-out traces, where the nine sit at 0.46 to 0.52, and it predicts slow reads better than Heimdall itself.

[LG-126] Efficient and Accurate Causal Validation of Training Data Attribution

链接: https://arxiv.org/abs/2611.19451
作者: Simon Vary,Alireza Mousavi-Hosseini,Roger Grosse
类目: Machine Learning (cs.LG); Methodology (stat.ME); Machine Learning (stat.ML)
*备注: 20 pages, 9 figures

点击查看摘要

Abstract:Causal validation of training data attribution (TDA) methods is costly, since evaluation involves retraining a model on many different subsets of the original training set. For example, the Linear Datamodeling Score (LDS), which has become a standard metric for evaluating TDA methods, measures the correlation between TDA predictions and expected retrained model outputs. In this paper, we reduce the amount of computation needed to estimate the LDS along three axes: the number of sampled training subsets, the number of retraining seeds per subset, and the number of models in the attribution ensemble. Each of these factors induces a form of bias in the LDS estimation, and removing these sources of bias significantly reduces the number of retrainings needed to match the same accuracy in LDS estimation. On ResNet-9/CIFAR-10, finite-subset debiasing achieves 3.2% relative RMSE with as few as 4 subsets, and 32 subsets match the accuracy of ordinary Spearman using 96 subsets, measured against a 300-subset reference. On ResNet-9/CIFAR-10, BERT/QNLI and GPT-2/WikiText-2, estimating the training-noise variance allows us to reproduce the standard five-seed LDS at the same accuracy with 2 – 5\times fewer retrainings.

[LG-127] Architectures and KV Replay for Depth-Recurrent Language Models NEURIPS2026

链接: https://arxiv.org/abs/2611.19446
作者: Lucia Licakova,Niklas Schmidinger,Lukas Hauzenberger,Sebastian Böck,Sepp Hochreiter
类目: Machine Learning (cs.LG)
*备注: NeurIPS 2026 - LIGHT Workshop

点击查看摘要

Abstract:Depth recurrence increases effective model depth by repeatedly applying a fixed computation block. However, sharing attention parameters does not share the key-value (KV) states generated at different recurrence iterations, which causes memory growth with recurrence depth during autoregressive decoding. We study how to reduce this memory cost through architectural and cache-sharing approaches within a common depth-recurrent framework. We first study a looped hybrid that repeats a stack of Transformer layers, with recurrent mLSTM or Gated DeltaNet blocks replacing most attention layers and periodic Transformer attention anchors retained. We then restrict recurrence to a dedicated recurrent workspace in a Parcae-style architecture and test its stabilising adapter as a wrapper for the looped hybrid. Finally, we train a looped Transformer with recurrent KV replay, where only selected recurrence iterations write KV states that later iterations reuse. This trains the model to operate with the same shared KV topology used during inference. We find that looped hybrids require sufficiently frequent attention anchors, with performance depending strongly on the recurrent mixer. For cross-iteration KV sharing, recurrent KV replay allows multiple recurrence iterations to reuse the same KV states while retaining performance close to the full looped Transformer with separate KV caches for each iteration.

[LG-128] Federated versus Centralized PV Capacity-Factor Forecasting: A Non-Inferiority Study Across Seven Stations in North Macedonia

链接: https://arxiv.org/abs/2611.19421
作者: Amar Kumar Mandal,Yuvraj Chaudhary,Dimitar Kitanovski,Eftim Zdravevski
类目: Machine Learning (cs.LG)
*备注: Submitted to Energy, Sustainability and Society

点击查看摘要

Abstract:Background: Photovoltaic (PV) forecasts inform energy commitments by PV operators and aggregators and guide charge-discharge scheduling in solar-plus-storage systems. Federated learning enables collaborative model training through model updates when local operators cannot pool raw operational data. This study evaluated whether federated training was non-inferior to centralized training for PV capacity-factor prediction across seven stations in North Macedonia. A common multilayer perceptron was used in an offline forecasting comparison with a 24-hour timestamp lag. Exploratory validation compared local, centralized, Federated Averaging, and Federated Proximal (FedProx) training. FedProx (E = 1, mu = 1) was selected for a confirmatory comparison under a protocol fixed before test evaluation. The comparison used past-only offline inputs, five paired training seeds, macro-averaged mean absolute error (MAE), a prespecified non-inferiority margin of 0.005, and a paired day bootstrap with 10,000 repetitions over 49 test dates. Results: FedProx achieved a macro-MAE of 0.128397, compared with 0.133405 for centralized training. The paired difference (FedProx minus centralized) was -0.005008 (95% confidence interval: [-0.009926, -0.000537]). The upper confidence bound was below the prespecified margin, satisfying the non-inferiority criterion. All five seed-specific differences favored FedProx, and its mean MAE was lower at six of the seven stations. Post hoc calendar sensitivity analyses did not change the conclusion. Conclusions: These findings support further evaluation of federated forecasting for collaboration under data-pooling constraints. The study demonstrates predictive non-inferiority within the evaluated setting. Benefits for energy commitments and battery scheduling remain potential applications requiring operational evaluation. Comments: Submitted to Energy, Sustainability and Society Subjects: Machine Learning (cs.LG) Cite as: arXiv:2611.19421 [cs.LG] (or arXiv:2611.19421v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.19421 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-129] st-Time Constraint-Aware Diffusion Predictive Control

链接: https://arxiv.org/abs/2611.19380
作者: Braghadeesh Lakshminarayanan,Ojas Phadake,Qi Chen,Dominik Baumann
类目: Machine Learning (cs.LG)
*备注: 17 pages, 2 figures

点击查看摘要

Abstract:Diffusion models can learn multimodal distributions over finite-horizon state-input sequences, providing a way to generate candidates for predictive control. However, sequences generated by a pretrained model may violate safety constraints introduced at deployment, for example when obstacles are added or repositioned. Retraining whenever these constraints change limits the usefulness of learned predictive controllers. In this paper, we propose Test-Time Constraint-Aware Diffusion Predictive Control (TCA-DPC), an algorithm for adapting the sampling distribution of a pretrained diffusion controller to newly introduced safety constraints without retraining. We treat the pretrained model as a prior over finite-horizon sequences and define an adapted distribution through exponential tilting based on a constraint-violation penalty. The tilt strength is adjusted online using the fraction of generated sequences that fail constraint verification. We approximately sample the adapted distribution using the Metropolis-adjusted Langevin algorithm in the Gaussian noise space of the pretrained diffusion sampler.

[LG-130] β-number and Its Applications to Data Science

链接: https://arxiv.org/abs/2611.19379
作者: Pengxu Wen,Ziwei Nie,Liwen Zou,Ying Cao,Xiaoping Yang
类目: Machine Learning (cs.LG)
*备注: 26 pages, 15 figures, including appendices

点击查看摘要

Abstract:In this paper, we introduce a novel metric, \beta -number, designed to measure the " p -th order d -dimensional oscillation degree" of a set. For a given set K in the Euclidean space \mathbbR^n , \beta -number is computed by identifying an optimal d -dimensional plane L that minimizes the average p -th order distance of all points in K to L . This metric serves as a versatile tool in data science, functioning both as a dimensionality reduction technique to find the optimal p -th order d -dimensional subspace and as a measure to characterize data oscillation for classification. We conduct a thorough exploration of \beta -number through theoretical analysis and numerical experiments. Initially, we establish the upper and lower bounds of \beta -number for any set K , along with its scaling law. We then elucidate the relationships between \beta -number and its related parameters. A key finding of this study is that in discrete settings, when p=2 , the d -dimensional plane determined by \beta -number coincides with that obtained via Principal Component Analysis (PCA), positioning \beta -number as a generalization of PCA. Furthermore, we propose two algorithms for calculating \beta -number in both continuous and discrete settings. We validate our theoretical results through a series of numerical experiments in low-dimensional cases. Finally, we demonstrate the potential of \beta -number in practical data analysis through a dimensionality reduction task and an image classification task.

[LG-131] Enforcing Ellipsoidal Constraints in Neural Networks with a Polynomial Safety Layer

链接: https://arxiv.org/abs/2611.19366
作者: Rolf Findeisen,Sebastian Hirt,Hoang Hai Nguyen,Grace Li Zhang,Hendrik Alsmeier
类目: ystems and Control (eess.SY); Machine Learning (cs.LG); Optimization and Control (math.OC)
*备注: 13 pages, 6 figures

点击查看摘要

Abstract:Often the output of a learned mapping must satisfy prescribed constraints. In the control of autonomous systems, for example, the chosen action must keep the system within a collision-free region. We consider the case in which a proposed output is corrected by projection onto a safety ellipsoid. Computing this projection amounts to solving a quadratically constrained quadratic program (QCQP). We construct a neural-network safety layer by unrolling a feasibility-preserving polynomial penalty iteration. The unrolled structure allows the loss gradient to propagate from the feasible output through the safety layer to the network parameters during training. We show how to choose the layer parameters so that every iterate is feasible and give a computable bound on its distance to the exact projection at any prescribed depth. We illustrate the method by training a LiDAR-based policy to control an unmanned aerial vehicle. A privileged model predictive control expert uses the map and the measured vehicle state to generate the training data. In the simulated narrow mazes, the filtered policy avoided wall contact at every tested stage of training, including at initialization, when every unfiltered flight hit a wall. A representative build of the learned policy and its safety layer for an Arm Cortex-M4F microcontroller occupies about 36 KiB of flash, with a RAM budget below 5 KiB.

[LG-132] Gradient-Based Co-Design of Systems and Reinforcement Learning Policies with Differentiable Simulators

链接: https://arxiv.org/abs/2611.19329
作者: Akhil S. Anand,Xun Shen,Marc Peter Deisenroth,Sebastien Gros
类目: Machine Learning (cs.LG); Systems and Control (eess.SY); Optimization and Control (math.OC)
*备注: 31 pages, 5 figures, 3 tables

点击查看摘要

Abstract:Co-design is the problem of jointly optimizing the design of a system and the policy that operates it. Neither can be chosen independently: a system design is only as good as the policy trained for it, and the optimal policy changes with the design. When that policy is learned by reinforcement learning (RL), evaluating a candidate design requires a full policy training run, and the design is usually optimized by black-box search such as Bayesian optimization (BO) or evolution strategies (ES). These methods scale poorly with the design dimension and ignore the differentiable structure of modern simulators and, in particular, the sensitivity of the optimal policy with respect to the design. We instead derive a gradient-based co-design method, where the design gradient decomposes into a direct effect of changing the design at a fixed policy and an indirect effect through the design-dependent optimal policy. The indirect effect requires the sensitivity of the converged policy to the design, which we obtain by applying the implicit function theorem to the stationarity condition of the policy gradient. Based on this sensitivity, we develop a proof-of-concept co-design algorithm with stochastic value gradient (SVG)-style policy optimization, prove that it converges to an approximately stationary design, and demonstrate on a differentiable networked-system benchmark with high-dimensional design spaces that it consistently outperforms BO and ES.

[LG-133] A Backdoor Trigger Is Legible in Fine-Tune Weights

链接: https://arxiv.org/abs/2611.19304
作者: Snigdha Chandan Khilar
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:A backdoored language model behaves normally until a trigger phrase appears in its input. Every published method for recovering such a trigger runs the model. We show the trigger is legible in the weights alone: a static statistic reads a fine-tuned checkpoint with no forward pass of the suspect, no data, no labelled population of poisoned models and no prior knowledge that a backdoor exists, and names one of the tokens of the trigger phrase as the single highest-ranked candidate in the vocabulary.

[LG-134] Does Lifestyle Information Improve Hypertension Prediction? A Leakage-Controlled Evaluation on NHANES ALT ML4H

链接: https://arxiv.org/abs/2611.19302
作者: Uriel Benymon
类目: Machine Learning (cs.LG)
*备注: Accepted to the Machine Learning for Health (ML4H) 2026 Symposium, Findings Track. Code: this https URL

点击查看摘要

Abstract:Does readily available lifestyle information improve machine-learning prediction of hypertension beyond routine demographic and anthropometric data? On NHANES 2017-March 2020, five model families (Logistic Regression, Random Forest, Gradient Boosting, Bagged Decision Trees, and XGBoost) were evaluated with fold-local preprocessing and paired bootstrap confidence intervals. Adding physical activity, alcohol use, smoking, and sleep information showed no consistent detectable improvement in held-out discrimination: every paired ROC-AUC confidence interval included zero, including under a stricter measured-blood-pressure phenotype and without synthetic oversampling. Mean cross-model SHAP rank agreement fell from 0.873 to 0.806 over all encoded columns, but this decline was largely attenuated over the 19 baseline columns shared by both arms (0.873 vs. 0.863), and likewise under both robustness checks; attribution-agreement comparisons across feature sets of different sizes therefore require a matched comparison space. In exploratory analysis of untreated adults (measured-blood-pressure outcome), all five models showed small positive differences, two with unadjusted intervals excluding zero, meriting confirmation. These results do not imply that lifestyle factors are clinically unimportant; within this cohort, the evaluated lifestyle block added little incremental discrimination beyond an already-informative baseline.

[LG-135] NetBreaker: Benchmarking Network Dismantling across Objectives Scales and Toplogoies

链接: https://arxiv.org/abs/2611.19298
作者: Wenchuan Yang,Tianci Bu,Suoyi Tan,Changjun Fan,Zijie Zhou,Xin Lu
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Network dismantling seeks to fragment a network with as few node removals as possible, which is widely applied in vulnerability analysis, epidemic containment, and infrastructure protection. Although learning-based methods have become increasingly prominent, their progress remains difficult to assess because existing studies differ in method coverage, datasets, evaluation criteria, and experimental protocols. Thus, we introduce NetBreaker, a comprehensive benchmark designed for the learning era of network dismantling. NetBreaker organizes 30 methods (18 non-learning and 12 learning-based) under a common taxonomy and evaluates them on 185 networks from eight unified domains, ranging from 14 to 1.69 million nodes. To enable fair comparison, heterogeneous method outputs are converted into node removal sequences and assessed under a unified execution protocol. The evaluation covers four complementary dimensions: target-based effectiveness, trajectory-based effectiveness, structural diagnostics, and applicability computational efficiency. Extensive experiments show that current learningbased methods do not consistently outperform strong non-learning approaches, while exposing substantial differences in effectiveness, scalability, and efficiency across network conditions and evaluation perspectives. NetBreaker provides a reproducible basis for measuring progress and developing reliable, scalable, and computationally efficient network dismantling methods.

[LG-136] Where Mixture Recognition Fails: An Ingredient-Level Audit of Unseen Combinations in SmellNet

链接: https://arxiv.org/abs/2611.19297
作者: Zirui Liu,Zheming Zhu
类目: Machine Learning (cs.LG)
*备注: 6 pages, 4 tables, 2 figures

点击查看摘要

Abstract:An electronic nose must name every ingredient of a mixture, including combinations it has never measured. SmellNet, a public benchmark for this task, scores ingredients with a hit rate that is told how many each sample contains. We audit what this protocol hides. We hold out ingredient combinations from the SmellNet training data, evaluate four architectures (MLP, CNN, LSTM and a Transformer encoder) without revealing the ingredient count, and repeat the evaluation once on the official unseen split. Under the count-informed protocol, 90-96% of mixtures receive an ingredient that is absent, on our folds and on the official split alike. On our folds, every model identifies one ingredient reliably but the second rarely, and it completes the pair with a familiar training partner about twice as often as chance. On the official split, even the first-ranked ingredient is correct in only 29-46% of binary mixtures, often because a frequent training ingredient that is absent from every test pair is ranked first, and on its combinations that are truly absent from training no model names both ingredients in more than 1.2% of them. The failures fall on the same ingredients for every architecture, and a frequency correction only moves errors between ingredients. Mixture recognition is therefore limited by ingredients the models cannot separate and by a preference for frequent ingredients, not by the architecture, and it should be evaluated with set-level errors and without oracle cardinality.

[LG-137] ReCALM: A Residual-enhanced Crowdsourced Analytic Learning Mechanism for Distributed Continual Service Construction

链接: https://arxiv.org/abs/2611.19276
作者: Kejia Fan,Yajiang Huang,Jianheng Tang,Jinfeng Xu,Huiping Zhuang,Anfeng Liu,Houbing Herbert Song,Yunhuai Liu
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Crowdsourcing has emerged as a promising paradigm for constructing AI services by harnessing the collective capabilities of distributed workers. As consumers’ demands continually evolve, service models must acquire new knowledge while retaining previously learned information. However, due to the inherited reliance on gradients for backpropagation, existing approaches typically suffer from catastrophic forgetting during continual model updates on heterogeneous data. Recent studies have introduced analytic learning, a gradient-free paradigm that achieves invariance through recursive closed-form solutions. Nevertheless, existing analytic methods are typically constrained by single-layer linear networks, which tend to suffer from underfitting due to their inability to perform deep representation learning. To address this issue, we propose ReCALM, a Residual-enhanced Crowdsourced Analytic Learning Mechanism that achieves deep nonlinear invariance for distributed continual service construction. Inspired by the success of ResNet, we develop gradient-free residual blocks for ReCALM with closed-form solutions. To implement distributed continual training for ReCALM, we further devise a cross-recursive strategy via least squares. Both theoretical and experimental analyses verify ReCALM’s dual superiority in heterogeneity invariance and deep representation. With 20 residual layers, ReCALM improves accuracy by 8.85%-9.83% over the single-layer analytic baseline, while reducing computational costs by over 94% compared with gradient-based baselines.

[LG-138] Physics-Informed Causal Reasoning for Nonstationary Robot Dynamics

链接: https://arxiv.org/abs/2611.19251
作者: Jungjin Lee,Se Jin Kim,Sung Kwon On,Insoon Yang
类目: Robotics (cs.RO); Machine Learning (cs.LG)
*备注: 22 pages, 6 figures. Project page: this https URL

点击查看摘要

Abstract:Reasoning about what an action will cause requires internalizing how a system behaves. As conditions seldom stay fixed, this must often happen within a few interactions. Yet systems that look identical in a snapshot can respond differently to the same action, depending on physical conditions the snapshot hides. Context-based models capture this separation by summarizing recent transitions into a latent context, but do not specify how that context should change the equations of motion. Physics-informed neural networks build those equations into the model, but describe only a single system and cannot recognize when it has changed. We present Physics-Informed Causal Reasoning (PICR), which bridges the two. Only a context, re-inferred from recent interactions, is updated; it conditions a fixed physics-informed network that determines the robot’s inertia and potential energy, so the robot adapts to new physical conditions from its own experience without retraining or a hand-specified robot model. On a Franka Panda arm in NVIDIA Isaac Sim, PICR predicts motion under changing and unseen payloads, reducing error on the latter by 1.6-3.8x over baselines, and extends to counterfactual queries under alternative torques. Our analysis shows that the context, inferred solely from interaction, encodes payload physical properties, tracks their changes, and governs the arm’s response. The gravity and inertia terms induced by the context further demonstrate its physical validity, each agreeing with the simulator at the level of individual Lagrangian components.

[LG-139] Learning Graph-Conditioned Predictors in Weight Space

链接: https://arxiv.org/abs/2611.19250
作者: Zhewei Chen,Hao Zhu,Ahad N. Zehmakan
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The same node attributes can imply different labels in different structural contexts. Yet fast graphless models typically treat topology as additional evidence supplied to a globally shared predictor. This assumes that structure augments the available information, rather than determining how node attributes should be interpreted. We introduce SCOPE (Structure-COnditioned PrEdictor), in which a small generator maps each node’s cached structural fingerprint to coefficients over a shared low-rank basis, producing a context-specific update to a global weight matrix. The base predictor, basis, and generator remain shared across all nodes, while the resulting transformation adapts how attributes are processed in each relational environment. Once fingerprints are cached, inference requires only a forward pass and no online message passing. This design combines context-specific specialization with global parameter sharing, compact low-rank adaptation, and message-passing-free inference. On two large OGB node-classification graphs, conditioning the weights with raw positional signals yields clear gains over supplying the same signals through input concatenation. Further results under a structure-transductive held-out-node protocol and in teacher-free graph classification show that the mechanism extends beyond a particular distillation setting. These findings establish graph structure not only as predictive evidence, but as context that determines the predictor used to interpret that evidence.

[LG-140] Continuous Lower-Limb Joint-Angle Estimation from sEMG and IMU: The Conditional Value of Cross-Activity and Cross-Subject Data under Update-Matched Controls

链接: https://arxiv.org/abs/2611.19238
作者: Jiaying Xu(Conamur),Guangyuan Chen(Institute of Neuroscience, Chinese Academy of Sciences),Buyun Gao(Conamur),Zhen Dong(University of California, Santa Barbara)
类目: Robotics (cs.RO); Machine Learning (cs.LG); Signal Processing (eess.SP)
*备注: 8 pages, 5 figures, 3 tables. Submitted to IEEE Robotics and Automation Letters (RA-L)

点击查看摘要

Abstract:Intent-driven exoskeleton control needs a continuous, causal estimate of the wearer’s joint kinematics and an answer to a second question: whose data should train the estimator for a new user? Borrowing the user’s recordings of other activities, or other users’ recordings of the same activity, is usually evaluated as “more data” versus “less data,” which confounds the information a source carries with the optimizer updates it adds and hides negative transfer behind averages. We separate these effects for hip, knee, and ankle angle estimation from high-density sEMG and an IMU across six locomotion activities on all ten participants of a public dataset. A compact residual causal TCN-GRU decoder is frozen, and an update-matched replay control prices every added source against the same number of updates on the user’s own data. A user’s own other activities raise R^2 from 0.722 to 0.788, of which +0.048 is information beyond replay; stair descent is a reliably harmful donor and sit-stand-walk a reliably beneficial one. Other users’ same-activity data help while the user has only the target activity ( G_\mathrmT = +0.053 ) and half as much once own other-activity data exist ( G_\mathrmW = +0.028 ); at full calibration, donor value collapses after one to two target records. Per-target gains from the two sources correlate, and leave-one-participant-out gates recover most of the cross-subject value only at low calibration. Data value is hierarchical, state-dependent, and directional; the bottleneck is deciding, without test labels, when to transfer.

[LG-141] Generalization Testing of Grid Foundation Models for Power Flow and Optimal Power Flow

链接: https://arxiv.org/abs/2611.19194
作者: Ayushi Jolotia,Parikshit Pareek
类目: ystems and Control (eess.SY); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:There is growing interest in foundation models (FMs) for power grids, particularly for power flow and optimal power flow. Their usefulness depends on generalization accuracy, which existing works test on their own benchmarks. We propose a three-axis framework to test their generalization limits: changed admittance matrices of training grids, load levels, and unseen grids. We test GENCO, LUMINA and GridSFM against the DC baseline, alongside their reported limitations, and discuss theoretical reasons for their failures. Experiments reveal that adding a line degrades all FMs more than removing one, and longer lines eventually make LUMINA and fine-tuned GENCO worse than DC. All three degrade beyond their training load range, where the loading-voltage curve steepens. On unseen grids, LUMINA degrades sharply even inside its training range, and GridSFM fails on the grid farthest from it. Thus, beyond training conditions, FMs behave more like meta-models requiring test-time fine-tuning data and compute.

[LG-142] ImpedanceMimic: Learned Impedance as Action Space for Force-Accurate Motion Imitation

链接: https://arxiv.org/abs/2611.19153
作者: Matthias Jammot,Pembe Gizem Özdil
类目: Robotics (cs.RO); Graphics (cs.GR); Machine Learning (cs.LG)
*备注: 39 pages, 13 figures, 28 tables. Project page: this https URL

点击查看摘要

Abstract:Recovering the forces behind a motion is ill-posed: different controllers can reproduce the same movement while generating different contact forces. Physics-based motion imitation resolves this ambiguity through simulation, but relies on fixed-gain proportional-derivative (PD) control. We show that fixed gains impose a trade-off between motion and force tracking, as contact and swing place conflicting demands on joint stiffness and damping. We introduce ImpedanceMimic, a reinforcement-learning policy that modulates per-joint stiffness, damping, and position setpoints from motion and force measurements, but requires only motion at inference. Across 129 participants and six activities, ImpedanceMimic outperforms matched fixed-gain PD under every integration scheme, direct-torque, muscle-actuated and existing imitation controllers in kinematic and force fidelity, approaching a dedicated motion-to-force regressor. By expressing gains as multiples of inertia-scaled priors, a single policy controls subject-specific biomechanical models without explicit anthropometric input and transfers zero-shot to unseen humans and a humanoid robot. The learned impedance forms a low-dimensional representation of motor state, shared across bodies, that generalizes to withheld activities: stiffness scales with load, and damping anticipates unloading.

[LG-143] A predictive per-engine energy roofline of the Apple Neural Engine

链接: https://arxiv.org/abs/2611.19101
作者: Spencer H. Bryngelson
类目: Computational Engineering, Finance, and Science (cs.CE); Hardware Architecture (cs.AR); Machine Learning (cs.LG); Performance (cs.PF)
*备注: 40 pages, 5 figures, 36 tables (including a 28-table supplement as appendices). Accepted to ACM Transactions on Architecture and Code Optimization (TACO). Code and data: this https URL

点击查看摘要

Abstract:The Apple Neural Engine (ANE) is in over a billion devices, but is characterized indirectly via CoreML. Using a CoreML-free fused-graph dispatch path, we measure the ANE, along with the Apple GPU and CPU on the same device, with per-engine energy accounting. The consumer-available roofline does not predict the ANE in practice. A critical correction is that the ANE sees two effective memory bandwidths: a slow activation path and a fast weight-stream path. Two roofline ridges follow, and the weight-stream ridge is below that of the GPU. The ANE single-operand compute peak decreases instead of plateauing. Its weight stream is byte-bound in fp16 but limited by the dequantizer’s element rate in int8. As a result, int8 compression saves energy at unchanged throughput and an int4 lookup-table stream raises throughput and efficiency at a cost of 2.55 points of ImageNet top-1. The corrected roofline identifies the lower-energy ANE paths and bounds it from below. A constructed device map reproduces across four unique ANE generations. Results show that operators should be placed by arithmetic intensity and fused for optimal performance.

[LG-144] DynaDFlash: Dynamic Feature Projection for Block Diffusion Speculators

链接: https://arxiv.org/abs/2611.19085
作者: Aadi Kulshrestha,Reese Chong
类目: Machine Learning (cs.LG)
*备注: Code and trained models available at this https URL

点击查看摘要

Abstract:DFlash trains a parallel draft model on hidden features from its target model. Inspired by the previous work of DFlare that introduced per layer combinations of target hidden states, we test whether context dependent weighting of these states helps improve acceptance length of block diffusion speculators. We trained both a DFlash baseline for the Qwen 3 of block size 16, and a dynamic model that computes gates for each layer to affect the magnitude of hidden features before projection. While mean acceptance length between the dynamic model and the DFlash baseline didn’t significantly differ, we notice that the dynamic model excels for drafting in the context of math. In an evaluation of 700 prompts from the MetaMathQA and Orca-Math datasets, the mean accepted length improves by 5%. We also found this lead to a 4.1% higher throughput for a subset of the same dataset in the context of speculative decoding.

[LG-145] Using machine learning approaches to classify false-positives generated by static analysis tools - An empirical study using small datasets

链接: https://arxiv.org/abs/2611.19079
作者: Lakmal Deshapriya,Sherlock A. Licorish,Brendon J. Woodford
类目: oftware Engineering (cs.SE); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Static Code Analysis (SCA) tools are vital for software development to reduce the cost and time required for manual code reviews. It was established in our previous study that the best tools reported in the software engineering community return high false-positive and false-negative rates when used for code analysis. Several approaches have been taken in existing studies to identify false-positive alarms generated by SCA tools, in helping to enhance the utility of these tools. However, existing studies reveal several limitations, chiefly among which is the tendency to focus on a limited number of alarm types or failing to evaluate a large breadth of Machine-Learning (ML) techniques comprehensively. This study has addressed this opportunity. We identified the best Code Representation Learning (CRL) techniques and ML algorithms for identifying false-positives, where results show that performance outcomes vary based on the characteristics of the training datasets and the goal of the model building. Given the evidence established in this study, we recommend further investigations targeting specific configurations of advanced CRLs and ML Algorithms, including simple ML models.

[LG-146] A Fast Algorithm for the Familywise Loss

链接: https://arxiv.org/abs/2611.19065
作者: Leelavathi Raja,Amirreza Eshraghi,Vinay Prabhakar,Mark Blacher,Han Bao,Yutong Wang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The familywise (FW) loss for multiclass classification, introduced by Duchi, Ruan, and Khosravi, achieves universal equivalence to the 0-1 loss, and thus has stronger theoretical properties than cross entropy (CE). Prior work, however, left open whether FW is as computationally efficient as CE. First, on the theoretical side, we prove that the FW gradient can be computed in O(K) time, where K is the number of classes. Second, on the practical side, we develop optimized kernels for computing the forward and backward passes of FW which outperform CE in terms of speed on both CPU and GPU.

[LG-147] Certify Then Select: Viability-Aware Runtime Steering of Frozen Diffusion Policies

链接: https://arxiv.org/abs/2611.19047
作者: Angelo Moroncelli,Loris Roveda
类目: Robotics (cs.RO); Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注: Under review

点击查看摘要

Abstract:Diffusion policies, which generate short action sequences by denoising random noise, are a leading way to learn robot skills from demonstrations. Once deployed, they meet constraints the demonstrations never contained, and retraining for each is impractical. A safety filter edits the policy’s action to satisfy them, but the edit ignores the task, and a large edit can leave the policy where it cannot continue. We propose CAST (Certified Action Selection at Test time): draw a few action chunks from the frozen policy, certify each with a barrier condition, and execute the certified chunk that a learned value ensemble ranks highest. We prove that a certified step is safe whatever the critic does; that, whenever a certified sample exists, the executed chunk stays within log N nats of the policy conditioned on feasibility, while a safety-filter projection can be singular to it; and that an exact closed-loop identity splits the gain of ranking into availability, spread and the critic’s target. Certifying the next chunk is not enough: a generative controller can have safe actions available now and none later. We characterise the safe control authority of a frozen policy under finite sampling, and CAST-V preserves it through a policy-conditioned viability, estimated with a model under a finite-sample guarantee or learned from the controller’s own logs (CAST-VL). On robomimic manipulation, a planar quadrotor, a MuJoCo humanoid and a unicycle, with frozen policies, CAST raises joint task-and-constraint success above the output filter in nearly every setting; certifying before ranking is decisive where a filter’s edit is large, and viability removes most dead ends that one-step certification cannot see. Project page: this https URL.

[LG-148] Advancing the Calibration-Compute Frontier of LLM Classifiers via Asymmetric Duos

链接: https://arxiv.org/abs/2611.19020
作者: Berke Akturk,Ali Can Sahin,Rasit Cem Solatay,Emanuel Sommer,Yawei Li,David Rügamer
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:The importance of reliable uncertainty estimation in large language models (LLMs) has become increasingly pronounced, yet modern models remain poorly calibrated and frequently overconfident. In this study, we address this by proposing an efficient method for LLM classifier calibration using the Asymmetric Duos framework. Our approach combines a strong base model with a smaller sidekick model via logit-space fusion, utilizing global scaling parameters optimized on a validation set. Across diverse benchmarks, this Asymmetric LLM Duo consistently achieves best or second-best performance on expected calibration error and negative log-likelihood, without degrading predictive accuracy. More importantly, these gains are realized at a fraction of the computational cost of more complex baselines, placing our method on the calibration-compute Pareto frontier. We further demonstrate that our method acts orthogonal to various established calibration-aware techniques, meaning it can be easily combined with them to consistently boost their performance. Finally, our analysis of the logit-space behavior illustrates how asymmetric fusion successfully mitigates base model overconfidence. Overall, our findings demonstrate that the benefits of an asymmetric fusion successfully transfer to the language domain, providing a simple, highly cost-efficient approach to improving uncertainty estimation in LLM classifiers.

[LG-149] Revealing the Biases of Large Language Models By Inferring Prior Distributions

链接: https://arxiv.org/abs/2611.18996
作者: Joshua Ryo Cheng,Thomas L. Griffiths,Jake C. Snell
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Large Language Models (LLMs) have been known to exhibit a wide range of social biases including name-nationality associations, occupation-gender associations and racial stereotypes. Existing approaches to evaluating these biases typically involve measuring surface-level behavior where a model is fed a dataset of prompts and its outputs are scored relative to ground-truth answers. Inspired by prior elicitation in cognitive science, we propose a different perspective that reconstructs latent bias as a prior underlying the model’s behavior. We formulate the problem as a Bayesian inverse problem where, given prompt-response observations on a bias benchmark dataset, we infer a low-dimensional latent distribution that summarizes the model’s stereotyping tendencies. This paper develops the mathematical foundation, a practical reconstruction algorithm, and experiments using the Bias Benchmark for QA (BBQ) benchmark to infer the prior distributions of LLMs over responses in ambiguous social contexts.

[LG-150] Bilateral Log-Linear Attention

链接: https://arxiv.org/abs/2611.18974
作者: Kazi Ashik Islam,Guojing Cong
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Sequence models represent past context at different resolutions, using token-level keys and values, fixed-size recurrent states, or hierarchies of compressed summaries. Full attention supports strong associative recall but requires quadratic work during training and prefill and a linearly growing key-value cache during inference. Linear attention and state-space models reduce these costs through fixed-size recurrent memory, but this compression can limit recall. Log-linear attention offers a middle ground by maintaining logarithmically many state matrices. However, its hierarchy favors finer resolution near the current query while compressing earlier context into coarser summaries. To complement this allocation of resolution, we introduce bilateral log-linear attention, which pairs the existing hierarchy with a second hierarchy anchored at the sequence origin. By combining their reads, the two branches provide finer access to both early and recent context while remaining causal. We derive recurrent and chunkwise parallel formulations for the added branch, preserving log-linear training and prefill work, logarithmic decoding work per token, and logarithmic persistent state storage at fixed feature dimensions and chunk size. In multi-query associative recall, this design raises mean overall test accuracy from 76.37% to 96.12% over forward log-linear attention across multiple training seeds. For pretrained language models under approximate parameter matching, it raises the mean score on six selected RULER needle-in-a-haystack tasks from 26.97% to 50.03% at a context length of 6144 tokens. In the same setting, bilateral attention also outperforms forward log-linear attention on seven of eight downstream retrieval tasks, including question answering and information extraction.

[LG-151] Contextual Non-Convex Bandit Optimization with Two-Point Feedback

链接: https://arxiv.org/abs/2611.18965
作者: Khaled Eldowa,Pierre Gaillard
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We study a non-stochastic contextual bandit problem with Lipschitz losses. At every round, the learner observes the context then picks two actions, observes the difference between their losses, and incurs their mean. The aim is to minimize the regret against a generic non-parametric comparator class of policies mapping contexts to actions; for example, Lipschitz policies and Hölder-smooth policies. We show how relying on a hierarchical discretization technique allows us to improve upon the regret rates achievable under one-point feedback, and even recover the full-information rate in some cases; notably, for sufficiently smooth policy classes or low-dimensional instances. Additionally, for Lipschitz policies, we present an algorithm whose per-round computational complexity is only logarithmic in the time horizon.

[LG-152] Supercharging Olmo-core for Efficient and Scalable MoE Training WWW

链接: https://arxiv.org/abs/2611.18956
作者: Tianhua Tao,Akshita Bhagia,Dirk Groeneveld,Pete Walsh,Tyler Romero,Yashas Samaga,Jacob Morrison,Iz Beltagy,Taira Anderson,Noah A. Smith,Hannaneh Hajishirzi
类目: Distributed, Parallel, and Cluster Computing (cs.DC); Machine Learning (cs.LG)
*备注: 169 pages. Web version and video: this https URL

点击查看摘要

Abstract:Mixture-of-Experts (MoE) models add parameters without proportionally increasing per-token compute. However, training cost also depends on bytes moved, memory held, and kernels launched, and many of those costs follow total parameter size rather than active compute. On infrastructure built for dense models, this mismatch is costly: in early baselines, throughput fell as stored expert parameter size grew, even though the useful work per token stayed nearly fixed. To close this gap, we present a redesigned training stack in Olmo-core, built for efficient, scalable MoE training. We replace the Fully Sharded Data Parallel (FSDP) of earlier Olmo-core versions with an architecture built on Distributed Data Parallel (DDP), and add Expert Parallelism (EP), Pipeline Parallelism (PP), a distributed optimizer, and activation recompute. We propose NVSHMEM-based rowwise EP, which writes each token directly into its expert’s buffer, avoiding dispatch and combine permutation overhead. We make EP fully synchronization-free by keeping routing metadata on the GPU and using device-scheduled grouped GEMMs. Closely integrated MXFP8 cuts compute, activation-memory, and transport costs, and topology-agnostic checkpoints let the parallel layout change between runs. On NVL8 B300 nodes, the stack supports training configurations from 12.9 billion to 1.2 trillion total parameters on up to 512 GPUs, reaching 858 useful-model TFLOP/s/GPU at 1.2 trillion with MXFP8 and per-layer recompute; an experimental capacity test with the optional DeepEP v2 backend reaches 2.38 trillion. We also explain each design choice and share findings for other training systems; for example, we identify a failure pattern we call Token Gerrymandering, in which the MoE router can learn to hack the load-balancing loss, and find that overlapping communication with computation can slow the overlapped kernels by more than it hides.

[LG-153] OverclusterDRG: a Deterministic LP-Free Heuristic for Robust k-Center

链接: https://arxiv.org/abs/2611.18931
作者: Rudra Bhardwaj,Rajni Dabas,Neelima Gupta
类目: Machine Learning (cs.LG)
*备注: 11 pages, 7 figures, 8 tables

点击查看摘要

Abstract:The k -center problem is a fundamental formulation for metric clustering. However, its min-max objective is highly sensitive to outliers, motivating the Robust k -Center (R k C) formulation which permits excluding up to z points as outliers. Existing algorithms for R k C suffer from scalability issues: LP rounding relies on fractional relaxations that scale poorly, known combinatorial approximation algorithms are computationally expensive due to radius guessing, and scalable randomized heuristics require exponential trial multipliers to guarantee success. We present OverclusterDRG, a deterministic, LP-free heuristic for R k C. The algorithm leverages the theoretical insight of farthest-point oversampling – adding exactly z extra candidates to the center set to absorb outliers – followed by a forward-greedy initialization and a 1 -swap local search. Evaluations across real-world datasets show that OverclusterDRG attains a mean approximation ratio of 1.147 against fractional LP lower bounds on the UCI Adult dataset, outperforming state-of-the-art baselines in both clustering quality and speed. By avoiding subset-guessing and exponential trial multipliers, it scales smoothly to n=10^5 points where LP solvers and theoretical algorithms become intractable. Comments: 11 pages, 7 figures, 8 tables Subjects: Machine Learning (cs.LG) Cite as: arXiv:2611.18931 [cs.LG] (or arXiv:2611.18931v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.18931 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-154] Privacy-Preserving Clustering Algorithm and Parameter Selection for Collaborative Population Stratification Using Distributed PCA

链接: https://arxiv.org/abs/2611.18925
作者: Maryam Ghasemian,Lynette Hammond Gerido,Erman Ayday
类目: Cryptography and Security (cs.CR); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We propose a privacy-preserving selection layer for collaborative population stratification workflows, enabling an honest-but-curious server to recommend a strong clustering algorithm-hyperparameter configuration from privacy protected shares. Instead of using a fixed workflow such as PCA followed by K-Means with a preset number of clusters, the framework compares multiple clustering methods and parameter settings under privacy constraints. To avoid relying on a public reference basis, we construct a shared low-dimensional representation through a distributed-PCA (DPCA) procedure in which each party computes local top- r principal-component information and the server aggregates only low-rank summaries to recover a global basis. Building on this representation, the framework supports three LDP-protected pipelines: PCA then Noise, Noise then PCA, and Noise-Only. The server evaluates clustering methods and parameter settings using internal validation metrics and recommends a top-performing configuration from the perturbed data; this step is pure post-processing. Experiments on openSNP and 1000 Genomes show that the server-side selection layer reliably identifies strong settings from LDP-protected data. Across both datasets, the PCA-based pipelines outperform Noise-Only in clustering utility, communication cost, and runtime, and show lower membership-inference attack (MIA) power at small privacy budgets. Overall, they provide the best utility–scalability trade-off in our setting. Subjects: Cryptography and Security (cs.CR); Machine Learning (cs.LG) Cite as: arXiv:2611.18925 [cs.CR] (or arXiv:2611.18925v1 [cs.CR] for this version) https://doi.org/10.48550/arXiv.2611.18925 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-155] How Much Do Foundation Models Add to Gradient Boosting on Financial Event Sequences?

链接: https://arxiv.org/abs/2611.18914
作者: Carlos Eduardo Martins Relvas,Gustavo di Giovanni Bernardo
类目: Machine Learning (cs.LG)
*备注: 49 pages, 5 figures, 24 tables. Code, synthetic data generator and per-seed predictions available at this https URL

点击查看摘要

Abstract:Foundation models pretrained on raw transaction sequences are proposed for credit, fraud and churn modeling, with reported gains over gradient boosting, but the data and baselines behind these gains are rarely public. We measure how much the pretraining and adaptation recipes of these models add to a baseline of point-in-time window aggregates and LightGBM, and we propose TBS (Temporal Boosting-Sequence ensemble), which averages the ranks of a boosting model enriched with label-free event-transition features and of a pretrained sequence network fine-tuned jointly with the aggregates. On a synthetic credit-card dataset, the public MBD-mini banking dataset and two proprietary datasets of trade credit and B2B churn, we compare 27 methods at three label budgets with five seeds and paired bootstrap intervals, using sequence models of about 0.4 million parameters pretrained per dataset. Frozen self-supervised embeddings rarely improve the baseline, and a fine-tuned sequence network alone clears it in 9 of 21 target-budget cells. TBS clears it in 19 of 21 cells (18 after FDR adjustment), with a median gain of +1.8 AUC points, and the controls suggest that combining boosting with the pretrained network is the main source of the observed gain. Gains shrink as labels grow and with rich tabular data: with both, the gain on the proprietary datasets is +0.7 and +0.8 points, the latter with an interval including zero. We also re-examine two benchmarks. On IBM TabFormer, the test follows a change in the fraud pattern and the published baseline has no history; with the training period the vendor states, our history-aware LightGBM and our own foundation model stay below the published scores of a commercial model. On IBM AML, a simulation artifact lets a date rule reach F1 0.482, and a point-in-time LightGBM reaches the F1 of graph neural networks whose test graph includes later transactions.

[LG-156] Quantifying Membership Inference Risks in Genomic Data Sharing: The Role of Attackers Auxiliary Information

链接: https://arxiv.org/abs/2611.18909
作者: Maryam Ghasemian,Lynette Hammond Gerido,Erman Ayday
类目: Cryptography and Security (cs.CR); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Genomic data sharing is essential for biomedical research but introduces significant privacy risks, including membership inference attacks, in which an adversary determines whether a specific individual’s data was used in a study. Existing privacy protections are typically applied after a release format has already been chosen, leaving researchers without a principled method to assess dataset-specific risk beforehand. We propose a server-assisted, principal component analysis-based pre-release auditing framework that maps genomic data into a shared low-dimensional space, applies epsilon-local differential privacy via the Laplace mechanism with a formally derived sensitivity for principal component analysis scores, and estimates membership inference attack power at a controlled 5% false-positive rate using a distance-threshold attack. A membership-neutrality diagnostic based on Unbiased Reconstruction Robustness is reported alongside attack power, precision, and recall. Experiments using the openSNP and 1000 Genomes Project datasets across multiple dimensionalities and privacy budgets show that separation between the researcher’s dataset and public auxiliary data in principal component analysis space is a dominant determinant of attack success, exceeding the effect of perturbation level alone. Comparison against a genome-wide association study-style likelihood ratio test baseline further contextualizes the measured risk. This framework enables researchers, data repositories, and institutional review bodies to quantify dataset-specific membership inference risk before genomic data publication, supporting more informed and privacy-aware biomedical data-sharing decisions.

[LG-157] Factorized Self-Supervised Speech Tokenization NEURIPS2026

链接: https://arxiv.org/abs/2611.18898
作者: Benjamin Van Niekerk,Jean-Philippe Letendre,Nicol Visser,Hugo Seute,Herman Kamper,Mirco Ravanelli
类目: Machine Learning (cs.LG); Audio and Speech Processing (eess.AS)
*备注: 26 pages, 6 figures, 16 tables. Accepted at NeurIPS 2026

点击查看摘要

Abstract:Speech tokenizers convert audio into discrete units, enabling language models to process and generate speech. The goal is to learn compact, text-like representations while preserving enough detail to reconstruct a speaker’s voice and delivery. We propose Facto: a factorized tokenizer that disentangles time-varying information (content and prosody) from static components (speaker identity and recording conditions). We build on a recent linear factorization method for self-supervised features, extending it to handle unseen speakers. Then, we cluster the factorized space and train a lightweight decoder to reconstruct audio from the resulting tokens. By removing the static components before clustering, our tokens are more robust to small noise perturbations. We validate this by testing token stability under various noise conditions. Next, we theoretically analyze Facto’s convergence and scaling properties, quantifying how the factorization improves with the number of training speakers. Finally, we evaluate Facto on a range of downstream tasks. We show that greater token stability translates into improved text-to-speech performance and competitive results in voice conversion and spoken language modeling. Code and checkpoints at this https URL and demo page at this https URL.

[LG-158] SLICES: A Benchmark and Evaluation Framework for Self-Supervised Learning on Sparse ICU Time Series ALT ML4H ML4H-2026

链接: https://arxiv.org/abs/2611.18868
作者: Hannes Ill,Dimitrios Proios,Leo Anthony Celi,Peter J. Schüffler
类目: Machine Learning (cs.LG)
*备注: Accepted at the Machine Learning for Health (ML4H) Symposium 2026, Proceedings track (PMLR 346). 28 pages, 8 figures, 21 tables. Code: this https URL

点击查看摘要

Abstract:Self-supervised learning (SSL) results for intensive care unit (ICU) time series are hard to compare when data processing, encoders, and adaptation vary together. We present SLICES, a benchmark and evaluation framework for comparing and extending SSL methods under a shared clinical setup. It fixes 24-hour value and observation-mask inputs, a Transformer encoder, patient-level splits, four endpoints, two adaptation protocols, and five matched seeds across MIMIC-IV, eICU-CRD, and their pooled union. Masked autoencoding (MAE), joint-embedding prediction (JEPA), and contrastive learning provide reference implementations. Among these objectives, MAE leads mean AUPRC in all nine frozen classification comparisons, also under converged probes, and in seven of nine fine-tuning comparisons. Fine-tuned MAE leads hospital mortality at full labels and scores higher than a supervised Transformer of the same architecture at every label fraction, and SSL fine-tuning gives the lowest length-of-stay error. Supervised baselines remain strong elsewhere: GRU-D leads 24-hour mortality and XGBoost leads acute kidney injury in all three regimes. Rankings depend on the adaptation protocol, the comparator set, and whether the probe loss matches the metric. A two-seed MIMIC-IV rerun with the maintained code stays within two across-seed SDs of the reference in all 48 downstream runs. The code repository supports new objectives, shared downstream evaluation, and regeneration of the reference results, so new SSL methods can be compared with these references under the same conditions. Generalization to unseen hospitals is not tested here and is the natural next extension.

[LG-159] Exstill: The Most Efficient Single-Chip Processor for the Next Token

链接: https://arxiv.org/abs/2611.18859
作者: Valavan Manohararajah,Vithursan Thangarasa
类目: Hardware Architecture (cs.AR); Machine Learning (cs.LG)
*备注: 62 pages plus references (72 total), 16 figures, 22 tables. Technical report

点击查看摘要

Abstract:Generating a token with a language model requires a fixed amount of arithmetic, whose energy sets a floor beneath any hardware. For Qwen3-1.7B at a 16K-token context, gate-level measurements at 5 nm place this floor at 0.89 mJ per token, yet an NVIDIA B300 spends 391-1,120x that. The excess is the price of generality, software layers, programmable hardware and moving weights and key-value (KV) caches from external memory, and it widens with each process generation. We present a co-design approach that brings token generation close to this floor and then lowers the floor itself, optimizing silicon and model together in a gate-level search loop. Fixing one model architecture in the datapath, with weights still loadable, and holding weights and every conversation’s KV cache on-die, a pre-silicon 5 nm design brings an unchanged Qwen3-1.7B token to 3.4x its floor. Co-designing the model with the circuits at fixed quality lowers the floor to 0.29 mJ and the token to 0.52 mJ, 1.8x the new floor and less than the baseline model’s arithmetic alone, while on-die capacity grows from one user to twelve. Against a measured B300 serving the same twelve users, the design decodes each user’s tokens 110x faster at 690x lower energy. To show that one fixed architecture can serve, as a co-processor, the latency-critical workloads around a frontier model, we introduce Polyglot-DC (1.7B and 4B, each with its own design), specialized for each job by distillation and task-specific training into its own weights. As a drafter it speeds single-user speculative decoding of large open models by up to 6.6x at the maximum reasoning effort. Its guard returns multilingual verdicts in a millisecond or less, within a point of the strongest open guard. Its agent and financial-reasoning models lead comparable open models on established benchmarks, with a reasoning turn falling from seconds to tens of milliseconds.

[LG-160] Identification and external validation of the sets of 4- 6- 8- 12- and 24-hour CGM motifs

链接: https://arxiv.org/abs/2611.18844
作者: Benjamin Lobo
类目: Machine Learning (cs.LG)
*备注: 22 pages, 7 figures

点击查看摘要

Abstract:There is growing interest in using CGM for purposes other than those initially intended. Recent work of Lobo et al. identified \Omega : a set of motifs such that almost any daily CGM profile (the 24-hour CGM time series starting at midnight) can be classified - matched to a motif in \Omega . This work generalizes the prior work by identifying sets of motifs for 4-, 6-, 8-, 12-, and 24-hr CGM profiles where an x -hr CGM profile does not have to be 24 hours in length, and does not have to start at midnight. The prior work methods are modified to identify the five new sets of motifs using data from the same 6 studies in the original work. The newly identified sets were subjected to rigorous external validation using a data set comprised of data from 21 studies (3,455 subjects, 289,592 days of CGM data). The external validation data set has both temporal and geographic transportability and addresses limitations of the data set used to define the five new sets of motifs. The results of the external validation indicate that the prior work generalizes well: the performance of each x -hr set of motifs on the external validation set is almost identical to that obtained for \Omega . The generalization of the original work broadens the potential use cases of the framework to include short term predictions (e.g., impending hypoglycemia) and diagnostics which rely on detecting changes in CGM time series which are less than 24 hours in length.

[LG-161] Recovered or Memorized? Investigating Dynamical Consistency and Faithfulness in World Model Latents NEURIPS2026

链接: https://arxiv.org/abs/2611.18807
作者: Jerome Francisco,Kyle Deng,George Flint,Kevin Zhu
类目: Machine Learning (cs.LG)
*备注: NeurIPS 2026 Workshop on Interpretability for Discovery (Interp4Discovery)

点击查看摘要

Abstract:World models develop internal simulators that can learn the structure and dynamics of an environment. It has been demonstrated that their internal representations are both meaningful and causal, but are these representations faithful to the governing dynamics of the environment they are learning? We introduce Anchored Recovery Certification (ARC), a procedure that utilizes a discrete Fréchet distance calculation with ground truth to evaluate dynamical consistency on candidate world models up to an affine gauge. We use the procedure on TD-MPC2 (decoder-free world model) that has learned the pendulum swing on an observation lift under varying hyperparameters at k = 5, 10, 25 steps. The model is then steered within a PCA-derived intervention subspace and compared against a ground-truth simulator using relative error to investigate dynamical faithfulness. ARC sufficiency held for TD-MPC2 most consistently at k = 10, with six of seven constraints statistically significant after Holm-Bonferroni correction, while the relative error between the simulator and the causal latent remained substantial. Further research could be conducted with decoder world models such as DreamerV3 (reconstruction loss) to investigate whether the results differ.

[LG-162] Clustering-Based Artificial Tree Ensembles for Random Forest Approximation

链接: https://arxiv.org/abs/2611.18803
作者: Lea L. Kronziel,Maren S. Wieder,Silke Szymczak,Björn-Hergen Laabs
类目: Machine Learning (cs.LG)
*备注: 5 figures

点击查看摘要

Abstract:Random Forests (RFs) offer strong prediction performance but are difficult to interpret. Most representative trees (MRTs) and artificial representative trees (ARTs) provide surrogate explanations, but a single tree may miss patterns within a forest. We investigate benefits and limitations of clustering-based representative tree ensembles for RF surrogate modeling and assess the contribution of similarity-based clustering relative to random tree partitions. Using linear and nonlinear simulations and regression benchmarks, we evaluate fidelity, prediction accuracy, complexity, and prediction stability against equally sized ensembles of top-ranked MRTs and randomly selected trees. Increasing ensemble size generally improves fidelity, accuracy, and prediction stability. ART ensembles show higher fidelity in simulations and often greater prediction stability for small ensembles. Compared with random tree partitions, similarity-based clustering generally improves ART fidelity in simulations and often on benchmarks. On benchmark datasets, MRT-based ensembles generally achieve higher fidelity and prediction accuracy, whereas ART ensembles have fewer terminal nodes. ART ensembles provide compact surrogates when reduced complexity matters and their fidelity and prediction accuracy meet application requirements. These findings clarify both the trade-offs of representative tree ensembles as RF surrogates and the contribution of similarity-based clustering beyond random clustering.

[LG-163] Actuarial-Informed Neural Networks: Constrained Deep Learning for Sex-Specific Multi-Population Longevity Forecasting

链接: https://arxiv.org/abs/2611.18802
作者: Davide Rindori
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We introduce Actuarial-Informed Neural Networks (AINN), a constrained deep learning framework for sex-specific multi-population longevity forecasting that embeds actuarial knowledge into LSTM training through differentiable penalty terms. Applied to a high-longevity 6-country cluster (Switzerland, Sweden, Norway, West Germany, Netherlands, Japan; 1956-2020), the framework is a challenger to the Li-Lee benchmark and unconstrained neural approaches. Our central finding is that monotonicity constraint intensity lambda functions as a continuous governance dial with a quantifiable accuracy-stability trade-off. At lambda = 0.01, 30-year forecast confidence intervals narrow by 6.5% at zero RMSE cost. At lambda = 0.1, they narrow by 25% at only +0.84% RMSE cost. This trade-off is invisible to standard hyperparameter optimizers, which select lambda by minimizing one-step RMSE, the wrong metric for structural constraints in recursive forecasting. The framework adds three contributions: (i) residual-calibrated process noise that eliminates double-counting in the dual uncertainty framework, reducing confidence intervals by 53-60%; (ii) a 5-seed model ensemble with a coefficient of variation of 1.06%; and (iii) a time-varying credibility weight Z(t) = exp(-alpha t) connecting to Buhlmann credibility theory for blending neural and actuarial predictions. A 10-year out-of-sample backtest (trained on 1956-2010, validated against observed HMD life expectancy for 2011-2020) yields a mean absolute error of 0.38 years, with 11 of 12 country x sex combinations within the 90% confidence interval. The model produces regulatory-grade outputs including Solvency Capital Requirements under the Swiss Solvency Test (ES 99.0%: +1.91 years, CHE Male) and Solvency II, reverse stress testing, and model-based stability verification (amplification ratio 1.02x). Subjects: Machine Learning (cs.LG) Cite as: arXiv:2611.18802 [cs.LG] (or arXiv:2611.18802v1 [cs.LG] for this version) https://doi.org/10.48550/arXiv.2611.18802 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-164] Hybrid Compression for Memory-Efficient Matrix-Vector Multiplication

链接: https://arxiv.org/abs/2611.18792
作者: Felipe A. Louza,Giovanni Manzini,Guilherme P. Telles
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Iterative machine-learning algorithms repeatedly access large real-valued matrices through matrix-vector products, and their performance is limited by memory bandwidth rather than arithmetic throughput. Lossless compression schemes that support computation directly over the compressed data reduce both the data moved and the memory footprint of these workloads. The mm-RePair framework [11] follows this approach by applying the RePair grammar compressor to the Compressed Sparse Row/Value (CSRV) representation of a matrix and performing matrix-vector multiplications directly on the compressed data. Its effectiveness, however, degrades on matrices that are dense or contain many distinct values. In this paper we present a hybrid strategy for the mm-RePair framework that combines grammar and entropy compression. The CSRV sequence is partitioned into a grammar-compressible component, formed by the entries that mm-RePair can compress effectively, and an entropy-compressible component holding the remaining entries, which is encoded with the ans-fold coder [24]. Compression improves further by sorting and delta encoding the entries of each row in the CSRV representation, and by remapping the alphabet to a contiguous range. Experiments on real-valued machine-learning matrices, including word embeddings and transformer hidden states, show that our hybrid strategy is more compact than the original \mmRepair on every input and reduces the peak memory of multiplication, with the largest gains on dense matrices with many distinct values, at competitive multiplication times. It also outperforms the column-reordering variant of \mmRepair on most of the matrices, and combining the two is even more compact on the sparse ones, at the cost of a much longer compression time. Finally, our strategy scales with the number of threads as well as the original \mmRepair does.

[LG-165] Setting Optimal Reserve Prices using Fine-Tuned Language Models

链接: https://arxiv.org/abs/2611.18785
作者: Joshua Foster,Fredrik Odegaard
类目: Machine Learning (cs.LG); Econometrics (econ.EM); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Online auction houses must set reserve prices for highly differentiated products whose unstructured descriptions contain valuable demand information. Existing approaches generally require managers to specify product attributes in advance. We ask whether a language model can instead learn from listing descriptions and historical bids to set item-specific reserve prices before bidding begins. We fine-tune a pretrained text embedding language model using the structural likelihood of auction bids as its training objective. This directs representation learning toward the valuation distribution that determines the reserve decision. We show that consecutive upper bids identify the tail information needed to set an optimal reserve without observing or modeling the number of potential bidders. A split-sample estimator learns the representation on one half of the data and conducts structural estimation and inference conditional on that representation on the other. We apply the method to approximately 65,800 online used-vehicle auctions. With seller values calibrated to match the sale rate among listings, the model implies a counterfactual net increase of 165.8 million, representing a 14.6% increase above the incumbent reserve policy. The framework translates a listing description and a seller’s value into an item-specific reserve recommendation without constructing a product taxonomy or requiring a bid history for the item being priced. It provides an approach to adapting language models for operational decisions in marketplaces where products are heterogeneous and demand-relevant information is recorded in text.

[LG-166] Can a Vision-Language Model Tune Clustering In Context?

链接: https://arxiv.org/abs/2611.18768
作者: Javith Ranatunga,Paimon Goulart,Evangelos E. Papalexakis
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Clustering is a widely used unsupervised learning method, but selecting an appropriate algorithm and set of hyperparameters for a specific dataset can require many possible configurations and evaluations. In this work, we investigate whether a Vision-Language Model can instead visually evaluate clustering results and use those decisions to then tune an explicit clustering algorithm. We study this through three progressive experiments using Qwen3.5-27B and five controlled synthetic clustering scenarios. First, we test whether the VLM can distinguish the best clustering result from a pair of plots. Second, we test whether it can assign numerical scores to those plots where a higher score represents a better result. Finally, we use these judgments within an agentic loop in which the VLM scores a k-means clustering and proposes updated hyperparameters that may result in better clustering. Our exploratory results show that the VLM performs well when clear cluster structure is present, while struggling on more ambiguous and noisy settings. In the full agentic experiment, the VLM produces clustering configurations with silhouette scores relatively close to those of grid search and Bayesian optimization, while terminating in much fewer iterations. Overall, these results provide an initial proof of concept that VLM-based visual feedback can be used to autonomously tune an explicit clustering algorithm.

[LG-167] Stagewise Adaptive Boosting

链接: https://arxiv.org/abs/2611.18662
作者: Valerie R. Fu,Brayden Lu,Lu Tian
类目: Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Stagewise algorithms and sparsity-inducing regularization are two influential approaches to high-dimensional modeling. Existing stagewise theory, however, largely centers on convex regularization, whereas stronger sparsity and other structural preferences are often pursued through fixed-level nonconvex optimization rather than generating an entire adaptive solution path. We develop stagewise adaptive boosting, a general framework that builds a monotone path by selecting the update consistent with structural preference. The framework can encourage sparsity along the path itself and also admits extensions that promote other properties of the solution, including group selection, similarity across models for multiple outcomes, and known priorities or prerequisite relations among predictors. We study its theoretical property as well as finite sample performance using simulations and real-data analyses. Across these studies, the proposed pathwise approach often achieves predictive performance comparable to commonly used baselines while selecting substantially smaller models. These results suggest that adaptive stagewise learning provides a practical route to stronger sparsity and richer structural control without repeatedly solving separate weighted or nonconvex optimization problems.

[LG-168] Rethinking the Verification of Machine Unlearning

链接: https://arxiv.org/abs/2611.18658
作者: Lei Zhou,Youwen zhu,Rongke Liu
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Machine unlearning claims require evidence that a model has removed the influence of specified training data. Yet widely deployable post hoc verifiers. infer unlearning largely from predictive behavior, the final model response that remains under the provider’s control. We make this limitation concrete through Output-Level Deceptive Response (ODR), an attack that produces responses accepted by output based verifiers while preserving the original representation state. By separating output conformity from internal unlearning, ODR motivates a rethinking of machine unlearning verification, shifting the audit target from what the model returns to what changes inside it. We operationalize this shift through Representation Unlearning Verification (RUV), a gray box verifier that compares intermediate forward representations in the original and unlearned models. RUV traces a four-stage response through activation reorganization, representation displacement, neighborhood disruption, and loss of manifold support. Matched retained controls distinguish target specific responses from global model drift, while joint permutation calibration combines the four evidence views into an adaptive decision. Experiments spanning exact and approximate unlearning, class and sample level requests, multiple datasets, and diverse architectures show mean ODR deception success rates of 97.0% and 97.7% , respectively. RUV achieves mean verification accuracies of 98.9% and 94.9% , while rejecting 99.5% and 99.8% of the corresponding ODR submissions.

[LG-169] Beveridge: Open Weight System One Decision Models with Calibrated Probabilities

链接: https://arxiv.org/abs/2611.18655
作者: Sathiskumar Jothi
类目: Machine Learning (cs.LG)
*备注: 23 pages, 11 figures, 8 tables. Models: this https URL , this https URL , this https URL . Code: this https URL

点击查看摘要

Abstract:Many production systems need a model to make a narrow decision about a document (route a ticket, check a limit, refuse a request, pick the next step of a browser agent) rather than to write text about it. We present Beveridge, a family of open weight System One decision models that take a document and up to 15 typed questions (yes/no, choice among described options, or a level on an ordered scale) and return, in one forward pass without generating text, an answer with a calibrated probability for every option. Beveridge-Large-C (398M parameters) adds a 3.5M parameter decision network to ModernBERT-large: typed number and date features in the input, exact facts from a rule engine, question-to-document evidence attention, a three-step recursive core with learned halting and an exact yes/no path. Beveridge-Pro-C adapts Gemma 4 26B-A4B with LoRA to score option letters, and Beveridge-Pro-C-Web extends it to the next step of a browser agent. On held-out tests the 0.4B model reaches 93.0% on our operations-document benchmark against 91.1% for the best of five general LLMs (31B-675B) and 87.9% for a commercial System One API, and, because its confidences are calibrated (ECE 0.029 on that benchmark), it can hand over 76% of decisions automatically at 95% accuracy, where no general LLM we tested can hand over any. It answers in 4.9 ms per decision on one NVIDIA H200 GPU and 60 ms on the 14 core CPU of an Apple M4 Pro laptop. Beveridge-Pro-C-Web raises next-step accuracy on web pages from unseen site types from 49.4% to 98.2% and completes live browsing tasks without losing document accuracy. We describe the models, the training recipes and an evaluation on 19 held-out test sets against a commercial System One API and five general LLMs, including where the models are weak. Weights and inference code are released under Apache 2.0.

[LG-170] Earth World Model: Physics-constrained latent dynamics for forecasting and environmental scenario analysis

链接: https://arxiv.org/abs/2611.18654
作者: Huijun Qiao,Ziqiang Zhang,Jinnong Liao
类目: Machine Learning (cs.LG)
*备注: 20 pages, 5 figures, 11 tables; includes Supplementary Information

点击查看摘要

Abstract:World models have been widely applied to multi-agent planning and forecasting, but exogenous environmental changes are rarely incorporated as replaceable drivers in a unified state-evolution process. This limits assessment of whether environmental perturbations propagate through physical systems with faithful dynamics. We propose the Earth World Model (EWM), an Earth-centred forecasting framework that compresses heterogeneous multi-source observations into a reusable low-dimensional latent state representing the system’s history and current condition. Given explicit exogenous drivers, a shared transition operator recursively updates the latent state, and physics-constrained decoding produces multi-step forecasts. Replacing the drivers enables environmental scenario analysis and evaluation of dynamic-response fidelity. We evaluated EWM for power forecasting and environmental-perturbation response at six photovoltaic sites and one wind-power site. Across four photovoltaic sites and three forecast horizons, three independent training runs yielded mean R^2 values of 0.8578-0.9095. In a separate six-site decoder ablation, the mean net per-unit response \rho under the prescribed irradiance intervention increased from 0.180 with a multilayer perceptron decoder to 0.922 with an analytic physical decoder, close to site- and horizon-specific analytic references. The wind model achieved an injection elasticity of 0.9990 \pm 0.0006 , while pre-intervention outputs remained unchanged. These results show that forecast fit and intervention response should be evaluated separately and that the tested EWM configurations preserved response amplitude and temporal boundaries under the prescribed interventions.

[LG-171] MeltKV: Memory-Aligned KV Cache Compression via Variable-Bit Packed Vector Quantization for Long-Context LLM s

链接: https://arxiv.org/abs/2611.18652
作者: Jiyun Han,Seongjoon Cho,Seungkyu Choi
类目: Machine Learning (cs.LG)
*备注: 9 pages, 10 figures, 5 tables. To appear in the IEEE/ACM International Conference on Computer-Aided Design (ICCAD 2026)

点击查看摘要

Abstract:Large Language Models (LLMs) increasingly rely on long contexts, making Key-Value (KV) caches dominate memory usage and inference latency. Existing vector quantization (VQ) methods effectively compress KV caches to sub-4-bit levels but often overlook inefficiencies in memory alignment and decoding speed. This work introduces MeltKV, a memory-latency co-optimized KV cache quantization framework for long-context LLM serving. MeltKV employs a memory-aligned variable-bit index packing scheme that adaptively assigns bit-widths to grouped key channels and compactly stores indices within standard data-types, eliminating redundant bits. Additionally, head-wise codebook sharing reduces the total codebook size while maintaining accuracy. Experiments demonstrate that MeltKV effectively compresses KV representations to sub-4-bit levels down to 1 bit and, at comparable perplexity, achieves 48.1% lower memory usage and 52.9% lower latency than state-of-the-art VQ methods.

[LG-172] Predictive maintenance in manufacturing: The operational value of continual machine learning

链接: https://arxiv.org/abs/2611.18636
作者: Pascal Janetzky,Michael Klar,Manuel Wenzel,Tobias Schlagenhauf,Stefan Feuerriegel
类目: Machine Learning (cs.LG); Applications (stat.AP)
*备注:

点击查看摘要

Abstract:Predictive maintenance in high-volume manufacturing often faces a domain-incremental learning problem: as new product configurations enter production, prediction models must learn from newly available data while preserving performance on earlier configurations that remain in operation. In this paper, we empirically study the operational value of continual machine learning methods for addressing this challenge in predictive maintenance, using real-world maintenance data from Bosch. The setting at Bosch covers 25 product configurations over five years, which creates a practical challenge for predictive maintenance models, as such models must be updated with data from newly introduced configurations without reducing performance on configurations learned earlier, a phenomenon known as “forgetting”. At the same time, retaining all historical sensor data and repeatedly retraining models is often infeasible in high-volume manufacturing since industrial machines generate too large data volumes. To address this challenge, we combine continual machine learning with robust optimization to support predictive maintenance. We then compare continual machine learning with a broad set of baselines, including fine-tuning, domain-adaptation methods, online learning methods, and configuration-specific models. Across experiments, continual machine learning methods achieve stronger predictive performance and lower maintenance costs than the baselines, with a reduction of cost-related indicators of more than 40 % relative to the baselines. Our results demonstrate the operational value of continual machine learning for decision support in maintenance when product portfolios grow over time and historical manufacturing data cannot be fully retained.

[LG-173] Mechanistic Interpretability of Mambas Selective Gate in Financial Markets: A Leakage-Controlled Study

链接: https://arxiv.org/abs/2611.18599
作者: Dip Roy,Rajiv Misra,Sanjay Kumar Singh,Anisha Roy
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Selectively gated state-space models (SSMs), notably Mamba, are increasingly used for financial time series, yet how they internally represent market regimes (bull, bear, sideways) remains opaque. This paper applies Mechanistic Interpretability (MI) to a six-layer Mamba trained on approximately twenty-one years of daily data for a point-in-time set of SP 500 constituents plus macroeconomic and Fama-French factors, decomposing per-layer hidden states with Top-K Sparse Autoencoders (SAEs). Our central contribution is methodological and cautionary. A prior, uncontrolled configuration recovered market regimes from SAE features at F1 = 1.00. Under a controlled protocol that (i) removes the three input features that overlap the HMM label definition, (ii) disables the auxiliary regime-classification head, (iii) uses a point-in-time constituent universe, and (iv) evaluates against labels disjoint from the inputs, the same regime label is recovered at only F1 = 0.398, indistinguishable from a logistic baseline on raw inputs (0.435) and from an SAE trained on a randomly initialized, untrained Mamba (0.364). A 2x2 attribution grid shows the near-perfect result is an interaction of auxiliary supervision and label-defining inputs: neither factor alone exceeds F1 ~ 0.40. Effect sizes with block-bootstrap inference show the residual selective-gate (Delta t) regime effect is statistically detectable but economically negligible and uniform across layers rather than layer-specialized; the apparent polysemanticity gradient disappears under continuous, theory-grounded measures. We contribute a reproducible leakage-control protocol for interpretability evaluation of financial SSMs, a quantified demonstration that previously reported regime structure is a label-contamination artifact, and a modest economically grounded feature subset that genuinely survives the controls.

[LG-174] Not All Data Leakage Is Equal: A Controlled Benchmark of Evaluation Inflation in Tabular Classification

链接: https://arxiv.org/abs/2611.18598
作者: Ankit Sharma
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Data leakage is widely recognized as a threat to trustworthy machine-learning evaluation, yet the phrase covers mechanisms with very different statistical consequences. We present a controlled empirical benchmark that measures the inflation in held-out ROC-AUC caused by five common leakage mechanisms: global standardization, global median imputation, duplicate contamination, supervised feature selection before splitting, and target encoding before splitting. The benchmark spans six public datasets and two synthetic stress tests, three model families, and five repeated stratified train/test splits, yielding 600 paired clean-versus-leaked comparisons. To avoid pseudo-replication, inference is performed on dataset-level effects with cluster bootstrap confidence intervals and Holm-corrected one-sided Wilcoxon tests. Label-dependent or direct data-reuse mechanisms produced systematic optimism: global target encoding increased ROC-AUC by 2.42 percentage points on average (95% cluster-bootstrap CI 0.62–4.89), global supervised feature selection by 1.93 points (0.63–3.39), and duplicate contamination by 1.52 points (0.40–2.80). All three remained significant after multiplicity correction. In contrast, global scaling and global median imputation had near-zero average effects under the i.i.d. settings studied here. A controlled target-encoding stress test further shows that leakage becomes sharply worse as categories become sparse. The results support a severity-aware view of leakage: evaluation pipelines should prevent all held-out-data dependence, but label-dependent transformations and observation reuse deserve especially aggressive auditing.

[LG-175] Skill-Latent Rewards: LLM -Guided Reward Formulation for Hierarchical Reinforcement Learning

链接: https://arxiv.org/abs/2611.18597
作者: Maryam Shoaeinaeini,Brent Harrison
类目: Machine Learning (cs.LG)
*备注: 9 pages, 6 figures

点击查看摘要

Abstract:Long-horizon reinforcement learning under sparse and delayed feedback remains an open challenge. Existing return-decomposition methods redistribute episodic rewards to ease credit assignment, but they operate on raw state–action transitions and disregard the subgoal structure that governs meaningful progress in complex tasks. We propose Skill-LR, a hierarchical reinforcement learning framework in which a vision–language model identifies the agent’s active subgoal at each timestep and a large language model generates a matching latent reward formulation through a two-stage prompting process. By conditioning reward signals on the currently active subgoal, Skill-LR produces stage-specific credit that evolves with the hierarchical plan, reducing temporal variance and improving sample efficiency. On multi-stage MiniGrid and BabyAI benchmarks, including DoorKey and Unlock-to-Unlock, Skill-LR converges faster and more reliably than sparse environmental rewards and strong credit-assignment baselines such as RUDDER and LaRe, demonstrating that subgoal-aligned latent rewards offer a principled path toward stable, interpretable credit assignment in long-horizon tasks.

[LG-176] Check Token: Real-Time Self-Verification and Precise Truncation in LLM Reasoning

链接: https://arxiv.org/abs/2611.18564
作者: Fei Ding,Yongkang Zhang,Yuhao Liao,Zijian Zeng,Huiming Yang
类目: Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:ECC memory embeds 8 parity bits for every 64 data bits and automatically detects and corrects errors on each read. The parity bits carry no data and only safeguard integrity, at ~12.5% overhead. Yet the reasoning chains of large language models lack such built-in self-verification: once an error occurs it propagates along the chain, and existing methods can only verify externally after generation completes. We propose the check token, establishing built-in self-verification for language model generation streams for the first time: a functional marker CHECK is added to the vocabulary, and the model triggers self-checking (analysis, localization, truncation, rewriting) by outputting it at any position. The check token carries no reasoning content (discarded after triggering) and only safeguards reasoning correctness, directly corresponding to the role of ECC parity bits, also at only ~13% overhead. Speculative forking further eliminates false-positive latency, and segmented reward makes trigger timing end-to-end learnable. Experiments (Qwen3-32B / Qwen3-Next) show that the check token achieves +10.8 pp improvement on HMMT25 at 1.13x overhead, with token efficiency 88x that of Best-of-8, and the precise truncation advantage monotonically increases with chain length.

[LG-177] Surrogate-Assisted Compact-Model Extraction for SKY130 CMOS at 77 K

链接: https://arxiv.org/abs/2611.21594
作者: Aaron Wang,Olivia Gzamouranis,Vaibhav Lohia,Hershel Fei Xun Weiner,Olivia Seidel,Richard Cavanaugh,Giuseppe Di Guglielmo,Nhan Tran,Davide Braga
类目: Instrumentation and Detectors (physics.ins-det); Emerging Technologies (cs.ET); Machine Learning (cs.LG); Applied Physics (physics.app-ph)
*备注:

点击查看摘要

Abstract:Cryogenic circuit design requires compact transistor models that reproduce measured low-temperature behavior and run in standard circuit simulators. We fit BSIM4 parameters to measured I-V curves of SkyWater 130 nm (SKY130) transistors at 77 K using a neural surrogate of NGSpice. The surrogate maps compact-model parameters and device geometry to the complete set of output and transfer characteristics. Extraction is a bounded least-squares search: a bank of NGSpice simulations supplies initial candidates, finite differences of the surrogate supply the local Jacobian, and NGSpice re-simulates the resulting cards before selection. On 18 measured geometries, we compare a search over seven card parameters with a search over 43 parameters selected by NGSpice sensitivity. Relative to NGSpice re-simulations of the previously released cards, the two searches reduce the in-sample mean relative root-mean-square error (RRMS) by 17% and 26%, respectively. The 43-parameter search has the lower overall error, but the seven-parameter search has the lower mean on high-bias curves, and some individual curves have higher error than with the released cards under both searches.

[LG-178] Benchmarking memory and nonlinear processing capacity of quantum recurrent neural networks for time series prediction

链接: https://arxiv.org/abs/2611.21552
作者: José Daniel Viqueira,Andrés Gómez,David Mera
类目: Quantum Physics (quant-ph); Machine Learning (cs.LG)
*备注: 20 pages, 7 figures, 8 tables

点击查看摘要

Abstract:Quantum machine learning has emerged as a promising approach to time series prediction. While quantum reservoir computing has been widely explored for this setting, variational quantum models tailored to sequential data remain comparatively understudied, largely because they are difficult to train. In this work, we present a rigorous benchmark of variational quantum recurrent neural networks (QRNNs) on time series prediction tasks, focusing on two key capabilities: memory and nonlinear processing. We introduce a task-driven methodology that quantifies model capacity, enabling an intuitive and fair comparison across architectures, including classical recurrent baselines. We evaluate the proposed framework on two generic benchmark tasks and on a synthetic dataset describing a chaotic system. Overall, the quantum models achieve performance comparable to their classical counterparts, but classical models scale better with the memory size within the range studied. Notably, one QRNN variant shows higher average capacity on the task designed to jointly probe memory and nonlinear capacity. Moreover, our results show that quantum models offer better capacity when we are able to find a good parameter initialization. However, the increasing sensitivity to initialization leads to a lower probability of finding good solutions as the model size increases. These results provide a principled starting point for designing richer quantum recurrent architectures and for assessing where they may offer practical benefits.

[LG-179] ONAM: Fitting Orthogonal Neural Additive Models for Post-Hoc Interpretability of Black-Box Predictions

链接: https://arxiv.org/abs/2611.21550
作者: David Köhler,Matthias Schmid
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Methodology (stat.ME)
*备注:

点击查看摘要

Abstract:Modern machine learning models, such as artificial neural networks, achieve high predictive performance through increasingly complex black-box architectures, limiting their use in settings where transparency, interpretability, and fairness are required. Here we present the R package ONAM, which provides a novel and practical implementation of functional decomposition, a technique to facilitate the interpretability of black-box models through structured effect representations. ONAM implements an algorithmic framework for the post-hoc decomposition of predictions into components associated with different input features. The package is designed for usability and efficiency, including functionalities for visualization as well as quantitative measures to assess interpretability. As a model-agnostic method, it can be applied to a wide range of prediction models. We describe the mathematical foundations, algorithmic design, and software architecture of ONAM. A central result of the proposed methodology is that the resulting decomposition satisfies the condition of hierarchical orthogonality, thereby guaranteeing that all components are identifiable and uniquely defined.

[LG-180] he cost of enforcing long-run coverag e: a phase transition in online conformal prediction

链接: https://arxiv.org/abs/2611.21542
作者: Yu Gui,Rina Foygel Barber,Cong Ma
类目: atistics Theory (math.ST); Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Consider an online prediction setting, where at each time, we are given a new feature vector and asked to provide a prediction interval or prediction set that will likely contain the corresponding response. In this setting, if the data points are i.i.d. (or more generally, exchangeable), conformal prediction provides a mechanism for building valid prediction sets around the output of any trained model, offering coverage probability \geq 1-\alpha at each time t . In practice, however, issues such as distribution drift, changepoints, or other challenges may cause violations of the i.i.d. assumption. In this adversarial setting, online conformal prediction methods can still offer guarantees on long-run empirical coverage. However, there is a trade-off between these two types of goals: enforcing tight constraints on long-run empirical coverage to protect against the adversarial setting can lead to unstable oscillations in the instantaneous training-conditional coverage in the i.i.d. setting. In this work, we seek best-of-both-worlds algorithms that aim to minimize cumulative error in instantaneous training-conditional coverage when in the i.i.d. setting, while also offering long-run empirical coverage guarantees in the adversarial setting. Specifically, consider a condition that requires long-run empirical coverage to deviate from the target level 1-\alpha by at most \mathcalO(T^-\beta) , for some \beta\in[0,1] . We prove that this family of conditions can be described with a phase transition. For \beta\in[0,\frac12] , it is possible to construct algorithms satisfying such a constraint in the adversarial setting, while offering the same instantaneous coverage performance as split conformal prediction if the data is in fact i.i.d. On the other hand, for \beta\in(\frac12,1] , this type of condition comes at a cost, and does not achieve best-of-both-worlds behavior. Subjects: Statistics Theory (math.ST); Machine Learning (cs.LG); Machine Learning (stat.ML) Cite as: arXiv:2611.21542 [math.ST] (or arXiv:2611.21542v1 [math.ST] for this version) https://doi.org/10.48550/arXiv.2611.21542 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-181] Almost Controlled: The FDR of Benjamini-Hochberg on Weighted Conformal p-Values

链接: https://arxiv.org/abs/2611.21509
作者: Ryan Lim(Albert Einstein College of Medicine)
类目: Methodology (stat.ME); Machine Learning (cs.LG); Statistics Theory (math.ST); Machine Learning (stat.ML)
*备注: 55 pages, 3 figures, 1 table. Main text 10 pages; full proofs in the appendices

点击查看摘要

Abstract:Weighted conformal p-values generalize conformal inference to cases with covariate shift. One way to use them is to pass them to the Benjamini-Hochberg (BH) procedure, either to decide which test units to select or to detect outliers. But weighted conformal p-values may not be positively dependent, so we want to know: does BH still control the false discovery rate (FDR) in finite samples? Based on experiments, it appears to, but we prove it doesn’t. We give two counterexamples, certified in exact rational arithmetic, where the FDR exceeds its target level q . We also identify how large the difference between the FDR and q can be. With n calibration points, m test units and weight ratio B , the relative excess \mathrmFDR/q-1 is bounded by a function of n , m , B and q that is of order \sqrtBm/(nq) when Bm/(nq) is small. The bound holds for every configuration with B\le n and every n . This order can be reached when multiple non-null test units share a common threshold: in a Poisson-limit model, that configuration’s excess is c(q)\sqrtBm/(nq) to first order as B/n\to0 and then m\to\infty . For some fixed distributions with this structure, the excess decreases only like n^-1/2 as n grows. If every test unit is null, the excess is O(B/n) at fixed m , and in certain configurations it grows linearly with m . To assist practitioners in judging how large the FDR can be, we turn that bound, which depends only on n , m , q and B , into a lowered BH level with proved finite-sample control, and give a much less conservative rule of thumb that covers every worst case we were able to find but provides no guarantee.

[LG-182] Learning from Duality in Neural Network Field Theories

链接: https://arxiv.org/abs/2611.21503
作者: Pietro Capuozzo,Brandon Robinson,Benjamin Suzzoni
类目: High Energy Physics - Theory (hep-th); Machine Learning (cs.LG)
*备注: 1+76 pages, 7 figures

点击查看摘要

Abstract:We develop neural realizations of positive-level Wess-Zumino-Witten models using bosonic Wakimoto and fermionic Fourier-Berezin fields. Level-rank duality connects the charged sectors of the fermionic construction, where explicit neural observables recover conformal block data and enable the transfer of information about unknown correlators between paired theories. We also show that distinct neural presentations of the same continuum current algebra exhibit different finite-width behavior, with approximation accuracy depending on architecture, regulator, and parameter distribution. These results connect the exact mathematical structure of conformal field theory with the expressive and computational properties of neural field theories.

[LG-183] Generative Data Augmentation Without External Data: An Empirical and Theoretical Study on Deep Networks

链接: https://arxiv.org/abs/2611.21467
作者: Daniel Kua,Yan Song
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: 24 pages, 7 figures

点击查看摘要

Abstract:Generative data augmentation adds samples from generative models to a classifier’s training set. Many influential studies use generators pre-trained on large external datasets, and so mix the value of synthetic data with knowledge transferred through pretraining. We study the case without external data, where each generator learns only from the classifier’s own training set, and ask when and why adding generated samples to that training set raises or lowers the classifier’s test accuracy. We train classifiers with four backbones on six datasets, two of them medical, adding samples from four generators. We bound the excess risk of a classifier trained with synthetic data, that is, how far its population risk can lie above the minimum risk within its hypothesis class. The bound charges synthetic data for the total variation distance between the real and generated distributions and credits synthetic data for a larger training set. The gain from adding synthetic data, the fall in population risk, is at most the excess risk of the real-only classifier, which trains on the real images alone. Read through the bound, the experiments suggest that synthetic data help most when our classifiers detect only a small discrepancy between the generated and real data and the real-only classifier has room to improve, and that synthetic data help little or hurt when the detected discrepancy is large. Generators with a larger lower bound on the total variation distance, a bound that we estimate from our own classifiers, tend to help less on five of the six datasets, and on the CIFAR datasets the most accurate backbone gains the least. The bound gives these directions but not the size of the largest gains, most of which traditional augmentation also reaches at the same number of optimizer steps.

[LG-184] Comments on “Deep Fréchet Regression”

链接: https://arxiv.org/abs/2611.21454
作者: Jianbin Tan,Anru R. Zhang
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: Comment on arXiv:2407.21407 (“Deep Fréchet Regression”). This discussion comment will appear in the Journal of the American Statistical Association (JASA)

点击查看摘要

Abstract:We discuss the Journal of the American Statistical Association (JASA) discussion paper “Deep Fréchet Regression” and explore an alternative perspective based on conditional flow matching. We propose learning covariate-dependent transport directly within the response space, avoiding separate manifold learning and reverse mapping. This perspective offers a potentially more flexible framework for incorporating geometric structure and modeling full conditional distributions.

[LG-185] Self-attention is a Dyson equation for many-body machine learning

链接: https://arxiv.org/abs/2611.21398
作者: Jarvist Moore Frost
类目: rongly Correlated Electrons (cond-mat.str-el); Materials Science (cond-mat.mtrl-sci); Machine Learning (cs.LG); Chemical Physics (physics.chem-ph)
*备注: 7 pages, 3 figures. Julia code available as part of this http URL

点击查看摘要

Abstract:Why does symmetric sum-pooling (DeepSets) predict ground-state potentials to sub-meV accuracy, yet fail for correlated wavefunctions and excited states? Sum-pooling is a mean-field (Hartree) model: its O(1/N) off-diagonal Jacobian suffices for nearsighted ground-state energies, but erases nonlocal two-body correlations. We show that self-attention is a real-space Dyson iteration whose Boltzmann (softmax) kernel acts as a two-body self-energy. Because physical Hamiltonians act pairwise, Löwdin’s 2-RDM theorem and the Galitskii–Migdal relation guarantee that iterating a two-body propagator captures the full energetics of correlated matter. When fitting a fractal backflow determinant, DeepSets blurs the nodal surface onto a mean-field envelope, whereas self-attention recovers the fractal attractor. Seen through this lens, the unreasonable effectiveness of attention is not an accident; in a universe built from pairwise interactions, two-body attention is all you need.

[LG-186] Anchoring Motion Refining Trajectories: Generalizable Continuous Neural Decoding

链接: https://arxiv.org/abs/2611.21335
作者: Fanqi Shen,Junru Chen,Surui Cheng,Jiahe Li,ZhiZhang Yuan,Gaorui Zhang,ZhiYuan Cao,Xiaoquan Shen,Yang Yang
类目: Neurons and Cognition (q-bio.NC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Brain activity encodes motor information across multiple temporal scales, enabling continuous neural control beyond discrete action recognition. However, existing methods are typically tailored to specific subjects and tasks, with direct neural-to-kinematic mappings that can entangle subject-specific neural variability with task-dependent motion patterns, limiting generalization. To address these limitations, we introduce AMRT, a shared neural decoding framework for heterogeneous electrode configurations and movement paradigms. Motivated by cortical encoding of movement direction and speed, AMRT adopts a two-stage design that anchors local motion through velocity estimation and refines the resulting trajectories over longer horizons, encouraging reuse of local motion tendencies across subjects and tasks. Across three datasets, AMRT generalizes to unseen trials and subjects and demonstrates transfer potential across neural recordings and movement paradigms, highlighting the potential of shared neural representations and anchor-and-refine decoding for generalizable continuous neural control.

[LG-187] High-probability Uniform Mixing via Dobrushin Contraction and Accelerated Hamiltonian Monte Carlo

链接: https://arxiv.org/abs/2611.21305
作者: Stefan Oberdörster
类目: Probability (math.PR); Machine Learning (cs.LG); Computation (stat.CO); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Hamiltonian Monte Carlo, since its invention and subsequent rise to widespread use, has been motivated by its ability to accelerate convergence to equilibrium compared to diffusive methods, both in terms of dependence on condition number and dimension. In the present work, we establish this acceleration simultaneously in both aspects for critically damped Metropolis-adjusted kinetic Langevin with persistent uniforms. More precisely, for target measures \mu(dx)\propto e^-U(x)dx with m -strongly convex, L -gradient Lipschitz, and L_H -Frobenius Hessian Lipschitz potentials, we show that the method mixes to \varepsilon -accuracy in total variation using \mathcal O\bigl((\kappa^1/2+L_H^1/2m^-3/4),d^1/4\bigr) gradient evaluations up to logarithmic factors, both uniformly over starting points in domains of high equilibrium probability and from warm starts. The proof incorporates ideas from previous studies by Bou-Rabee and Chen et al. that show either aspect of acceleration separately, into novel coupling technology. This new coupling technology for general Markov chains extends earlier frameworks aimed at capturing fast local equilibration in the presence of slow global mixing. Specifically, via a weak- (1,1) maximal inequality that follows from Hopf’s maximal ergodic theorem, we uncover a stability mechanism in high-probability domains with respect to equilibrium. This enables straightforward localization to naturally arising domains of fast local mixing, and seamlessly fits broad local frameworks via Dobrushin contraction, Wasserstein contraction and local Harris’ theorems. We develop the Dobrushin formulation.

[LG-188] Compressed Simulation-Based Inference: Neural Posterior Estimation for Massive Data via Low-Dimensional Sketches

链接: https://arxiv.org/abs/2611.21286
作者: Michael Wieck-Sosa,Cosma Rohilla Shalizi
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Computation (stat.CO); Methodology (stat.ME)
*备注:

点击查看摘要

Abstract:Neural posterior estimation (NPE) is a popular Bayesian simulation-based inference (SBI) approach that uses neural networks trained on simulated data–parameter pairs to approximate the posterior given a user-chosen prior. For high-dimensional observations, NPE methods typically approximate the conditional density of the parameter given a user-chosen or learned summary statistic, rather than the entire dataset, due to the curse of dimensionality. However, both approaches face challenges: hand-crafting good summaries for complex models requires considerable effort from scientists, while learning them from massive datasets can be prohibitively computationally expensive. To address these challenges, we propose approximating the conditional density of a parameter given low-dimensional sketches of the data. We show that, for a model with a p -dimensional parameter, only 2p+1 random sketches are needed for posterior consistency. We compare with state-of-the-art embedding networks for NPE, and we find that our method \emphmatches the best embedding network’s performance while \emphincreasing computational efficiency.

[LG-189] A COMPASS to the Galaxy: Joint nucleosynthetic yield set comparison and parameter inference in Galactic chemical evolution

链接: https://arxiv.org/abs/2611.21175
作者: Berkay Günes,Giuseppe Viterbo,Sven Buder,Tobias Buck
类目: Astrophysics of Galaxies (astro-ph.GA); Instrumentation and Methods for Astrophysics (astro-ph.IM); Machine Learning (cs.LG); Data Analysis, Statistics and Probability (physics.data-an)
*备注:

点击查看摘要

Abstract:Galactic Chemical Evolution (GCE) models connect stellar nucleosynthesis to observed abundance patterns, but their interpretation requires simultaneous parameter inference and comparison between competing stellar yield prescriptions. This remains challenging due to intractable likelihoods and growing dataset complexity. Using compass, a novel simulation-based inference framework, we perform, for the first time, amortized joint Bayesian model comparison and parameter inference for GCE models with intractable likelihoods within a unified architecture. Compass employs a conditional score-based diffusion model with a transformer architecture to learn both the posterior and likelihood directly from forward simulations, enabling efficient inference and likelihood-based ranking of competing nucleosynthetic yield sets. In controlled tests, compass identifies the generating yield set and recovers ground-truth parameters with accuracy comparable to existing simulation-based inference (SBI) methods, while reducing computational cost by orders of magnitude. Applied to Milky Way data, the framework favors a hybrid NuGrid + IllustrisTNG yield prescription, confirmed by an independent classifier, and infers an initial mass function slope of \alpha_\rm IMF = -2.65 \pm 0.02 and a SN \rm Ia normalization of \log(N\rm Ia) = -2.81 \pm 0.02 , with a systematic spread of \approx 0.3 in \alpha_\rm IMF between the best-ranked yield sets, consistent with independent constraints. The attention mechanism confirms that the inferred parameters are driven by physically meaningful abundance signatures. By unifying amortized parameter inference, Bayesian model comparison, and interpretability in a single probabilistic framework, compass provides a scalable approach for precision chemical evolution studies with current and forthcoming spectroscopic surveys.

[LG-190] owards Microphone-Invariant Speech Representations for Transformer-Based Voice Disorder Detection

链接: https://arxiv.org/abs/2611.21140
作者: Iva Ewert,Felix Putze,Marvin Borsdorf,Tanja Schultz
类目: Audio and Speech Processing (eess.AS); Machine Learning (cs.LG)
*备注: 5 pages, 3 figures, 1 table, under review

点击查看摘要

Abstract:Automatic voice disorder detection from speech offers strong potential for early diagnosis and continuous disease monitoring. While models based on pre-trained transformers have improved detection performance, their robustness to microphone variation remains underexplored, particularly for consumer-grade devices. To the best of our knowledge, we present the first study of microphone mismatch in transformer-based binary voice disorder detection across 20 smartphone microphones. While most mismatched microphone conditions cause moderate performance degradation, the worst-case condition reduces unweighted average recall by up to 12.53 percentage points. To address this, we introduce a Barlow Twins-inspired module to learn microphone-invariant representations. Our approach reduces the performance gap between matched and mismatched conditions by up to 62%, while maintaining comparable matched-condition performance. Qualitative analysis of the embedding space demonstrates more compact, distinct speaker clusters across microphones. Importantly, results indicate generalisation to unseen microphones across manufacturers, smartphone generations, and release periods, without requiring any inference-time adaptation.

[LG-191] he Limits of Closed-Form Couplings for Few-Step Generation

链接: https://arxiv.org/abs/2611.21119
作者: Tobias Höppe,Stefan Bauer,Qiang Liu,Andrea Dittadi,Kirill Neklyudov
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Few-step generators via flow models require complicated technical solutions such as a pretrained flow model or a consistency loss. While rectified flows have gained popularity for their success in few-step sampling, they pay a cost of needing a pre-trained model and sampling a large number of couplings per dataset. Yet for a finite dataset and a Gaussian probability path, the probability-flow vector field of the empirical distribution is available in closed form, so one could use this field as the teacher and skip pretraining altogether. We demonstrate that, surprisingly, this approach is impossible for problems of practical relevance. In particular, the analytic solution of the diffusion process in high dimensions becomes almost a linear transformation. The integration returns the noise vector drawn at initialisation, plus a drift that is empirically a faint copy of the data point, almost independent of the noise. The coupling therefore differs from an independent pairing only through this drift. As a result, training a model on these couplings is almost equivalent to training the model on the independent draws of the noise and the data, which does not provide a useful teacher. We highlight that this is the direct consequence of the high-dimensionality of the data, and that for low enough dimensions or large enough datasets, analytic couplings become more efficient teachers. We quantify this effect by the fraction of endpoint energy carried by the drift, which scales as \frac\log Nd for a dataset of size N in a d -dimensional space. Finally, we provide an empirical study confirming all the theoretical results presented in the paper.

[LG-192] Out-of-training properties in frozen interatomic potential embeddings are decodable and ordered by physical scale NEURIPS2026

链接: https://arxiv.org/abs/2611.21116
作者: Natalija Stepurko,Joonatan E.M. Laulainen
类目: Chemical Physics (physics.chem-ph); Materials Science (cond-mat.mtrl-sci); Machine Learning (cs.LG)
*备注: 61 pages (9 main text), 18 figures, 23 tables. Accepted at AI4Mat and AI4Science workshops, NeurIPS 2026. Code: this https URL ; ORB-v3 embeddings: this https URL ; UMA-S embeddings (gated): this https URL

点击查看摘要

Abstract:Machine-learned interatomic potentials (MLIPs) are trained on energies, forces and stresses, yet their frozen representations can predict other materials properties. We probe ORB-v3 and UMA-S across layers, read-outs and equivariant channels for nine properties of 154,875 Materials Project crystals. Both models outperform composition-only descriptors on all targets, including five outside their training objectives; errors in physical units help assess potential screening applications. Optimal depth varies by property, consistent with a physical-scale ordering. Random-weight controls reveal strongly property-dependent benefits of pretraining. UMA-S’s rotation-invariant angular channels improve crystal-system classification, while higher scores from its full tensor disappear under rotation; matched-model controls do not isolate equivariance as the explanation. Accuracy varies across different probes, but nearest-neighbour prediction often approaches XGBoost performance, even though materials do not form distinct clusters according to their properties. We also test whether a linear predictor trained on one model’s embeddings can predict properties using the other model’s embeddings after a learned linear transformation. Predictive accuracy is largely preserved, despite differences in which atoms the two models place close together. These results guide the reuse of frozen MLIP representations for materials prediction.

[LG-193] CONGA: Compiling Gaussian Fiber Geometry for Structured Bayesian Inference NEURIPS2026

链接: https://arxiv.org/abs/2611.20995
作者: Víctor Gallego
类目: Computation (stat.CO); Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注: Accepted to the Workshop on Geometric Distributional Deep Learning (GDDL), NeurIPS 2026

点击查看摘要

Abstract:In structured Bayesian models, a large latent block is often Gaussian conditional on a smaller set of parameters, yet generic inference methods relearn this geometry during warmup. We study how this structure can be used to construct transports for posterior inference. The conditional mean and precision determine exact affine maps, including symmetric Brenier and triangular Knothe-Rosenblatt maps, that separate a standard normal block from the remaining non-Gaussian core. KL and relative Fisher objectives then reduce to objectives on this core. We introduce CONGA, a compiler that identifies candidate Gaussian blocks and selects a temporal, spectral, or sparse implementation according to their structure and computational cost. Marginalization requires an identity established from the model; numerical evidence alone permits only a target-preserving change of variables, and a failed check leaves the model unchanged. On the Stock-Watson inflation model, CONGA improves minimum ESS per second by 12.5x for NUTS. It also restores convergence where the raw target fails: for WALNUTS on the same model, and for NUTS on a 6760-dimensional electricity load model reduced to 34 core coordinates. Experiments on graph-indexed fields show how coloring and fill affect the choice of map, including cases where retaining the original target is preferable. Code at this https URL Comments: Accepted to the Workshop on Geometric Distributional Deep Learning (GDDL), NeurIPS 2026 Subjects: Computation (stat.CO); Machine Learning (cs.LG); Machine Learning (stat.ML) Cite as: arXiv:2611.20995 [stat.CO] (or arXiv:2611.20995v1 [stat.CO] for this version) https://doi.org/10.48550/arXiv.2611.20995 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-194] ransferable neural-network preconditioners for Wilson–Clover propagator inversions

链接: https://arxiv.org/abs/2611.20990
作者: Shiyi Zhong,Yuanyi Zhang,Jian-Hui Zhang,Dian-Jun Zhao
类目: High Energy Physics - Lattice (hep-lat); Machine Learning (cs.LG); Computational Physics (physics.comp-ph)
*备注: 15 pages, 6 figures

点击查看摘要

Abstract:We investigate two single–layer neural-network architectures for accelerating four-dimensional Wilson–Clover propagator inversions on lattice ensembles in practical use; both transfer directly without retraining. The first is a gauge-covariant parallel transport convolution (PTC) network that directly reduces the number of generalized conjugate residual (GCR) solver iterations. The second is a CNN-like network used together with a multigrid-inspired algorithm. The two approaches can be combined to further reduce GCR iterations. PTC transfers without retraining across lattice volumes, lattice spacings, and pion masses while maintaining comparable acceleration, whereas the CNN-like network transfers directly from small- to large-volume ensembles with strong acceleration. A cost comparison accounting for both basis construction and solves shows that the proposed scheme is more advantageous for large-volume ensembles than for small-volume ones.

[LG-195] FARCH: A Functional ARCH Model for Calibrated Prediction Bands

链接: https://arxiv.org/abs/2611.20972
作者: Mohamed Chaouch
类目: Methodology (stat.ME); Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Forecasting a curve, rather than a handful of scalar summaries of it, requires a notion of uncertainty that is itself functional. We propose FARCH, a functional nonlinear autoregressive model with conditional heteroscedasticity, in which the conditional mean and the conditional volatility of the next curve are forecasted jointly and nonparametrically, by reusing a single kernel weighting scheme — built on a wavelet semi-metric between curves — first on the curves themselves and then on their squared residuals. From this pair of estimators we derive two constructions of prediction bands: a pointwise band, obtained from marginal scenario quantiles, and a uniform, sup-norm band, calibrated for decisions that concern the entire forecasted path rather than any single instant. Applied to day-ahead forecasting of half-hourly French national electricity demand (1996–2009), and compared, under an identical protocol, against four existing prediction-interval constructions that do not model the conditional variance explicitly, the uniform FARCH band attains pointwise coverage statistically indistinguishable from its nominal target at the 80%, 90% and 95% levels, and a curve-wise coverage of 90.1%, significantly above the 70.8% reached by the best competing method, at the cost of a wider interval. Pointwise bands, in contrast, undercover regardless of whether the variance is modelled explicitly, indicating that state-dependent volatility and simultaneous calibration are two distinct ingredients that a trustworthy prediction band must combine.

[LG-196] Parallelizable Interior-Point ADMM for Block-Structured Nonconvex Programs

链接: https://arxiv.org/abs/2611.20483
作者: Zherong Pan
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Physics simulation is an increasingly important computational primitive for AI, and local–global/ADMM splitting exposes its parallel structure but lacks matching finite-time convergence guarantees. We close this gap with two ADMM variants, Interior-Point ADMM (IP-ADMM), for block-structured nonconvex composite programs with smooth nonlinear inequality and affine constraints. A one-sided finite-activation wall keeps each constraint split state strictly feasible while the original-space infeasibility vanishes, giving closed-form constraint updates, one proximal call per objective block, and one equality-constrained quadratic consensus solve per iteration. \revTo our knowledge, this makes IP-ADMM the first single-loop method that updates nonconvex inequality constraints locally and in parallel with finite-time convergence guarantees. Ordinary IP-ADMM attains a CQ-free best-iterate weak-KKT rate of O(K^-1/3) or a banded Fritz–John rate of O(K^-1/4) , which converts to KKT at O(K^-1/4) under a uniform constraint qualification; these are nearby-point certificates that reduce to conventional KKT for smooth blocks. A checkpoint variant, IP-ADMM-Ckpt, improves the asymptotic KKT exponent to O(K^-1/3) . We also show how the bounds scale with the objective- and constraint-block counts and the coupling norm.

[LG-197] On the Convergence of Continuous-Time AdamW

链接: https://arxiv.org/abs/2611.20437
作者: Bangjie Wang,Ning Ning
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注: 55 pages, 9 figures

点击查看摘要

Abstract:Continuous-time AdamW provides a framework for understanding whether an optimizer widely used to train large-scale neural networks settles to an equilibrium or sustains persistent motion. We study this long-time behavior while retaining bias correction and fixed positive weight decay. Under suitable assumptions, we establish that the dynamics are well defined for all time and remain bounded when the decay is sufficiently small. Our main result connects their limiting behavior to the geometry of the objective: trajectories approach a locally invariant manifold whose dimension equals the Hessian nullity along the associated critical component. When critical points are isolated and have at most one such direction, the trajectory converges to a single equilibrium. With two degenerate directions, however, bounded persistent motion can occur even for a smooth, strictly convex objective with a unique minimum. We also quantify how close the long-time trajectory remains to the critical set as weight decay decreases, without assuming convergence to a point. These results identify objective geometry as a key to understanding when continuous-time AdamW converges and when persistent motion is possible.

[LG-198] High-dimensional transfer learning with unaligned ordinal responses

链接: https://arxiv.org/abs/2611.20411
作者: Ruopei Fu,Ying Yang
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Statistics Theory (math.ST)
*备注:

点击查看摘要

Abstract:In this work, we study high-dimensional transfer learning with unaligned ordinal responses, allowing different domain-specific regression coefficients, cumulative links and thresholds. We develop the sample-split and full-target algorithms that pool source and target likelihoods and then estimate a target correction. Under suitable conditions, our algorithms achieve minimax-optimal rates for parameter estimation and excess risk. Simulations favor the full-target algorithm over the sample-split algorithm for finite-sample estimation and prediction. An application to a medical dataset further demonstrates the predictive gains of our full-target algorithm over the target-only and pooled baselines when transferring information from the five-category source data to binary hospital mortality.

[LG-199] Generalized Prediction Set with Bandit Feedback

链接: https://arxiv.org/abs/2611.20387
作者: Zhou Wang,Xingye Qiao
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:In high-stakes environments where uncertainties abound, set-valued prediction offers a cautious and robust mechanism by presenting multiple potential labels as the prediction for each test instance to mitigate the potential risk associated with prediction errors. Yet, integrating this paradigm with out-of-distribution (OOD) detection remains scarcely explored in such settings as online learning with bandit feedback. The bandit feedback mechanism informs the learner about the correctness of the pulled arm/action instead of the explicit ground truth label, leaving the true class label unknown when an incorrect action is taken. To address this challenge, we introduce BanditGPS which conducts set-valued prediction with OOD detection in the bandit feedback setting, using an estimation to the ground truth of class labels. BanditGPS achieves three objectives: render small/informative prediction sets, enhance the OOD detection performance, and control the recall for all normal classes to meet prescribed requirements. Our approach is characterized by the loss function, which trades off between high OOD detection and small prediction sets. Theoretically, we prove that the convergence rate of the regret is \tilde\mathcalO(T^-1/2) . The empirical results further show that BanditGPS effectively controls the recalls with promising performances on OOD detection and informative prediction.

[LG-200] Multi-Library Feature Engineering and Statistically Validated Classification for Power Swing Detection and Stability Assessment

链接: https://arxiv.org/abs/2611.20364
作者: Pallav Kumar Bera,Samita Rani Pani,Ranjeeta Patel
类目: ignal Processing (eess.SP); Machine Learning (cs.LG)
*备注: Accepted at Annual Conference of the IEEE Industrial Electronics Society (IECON) 2026; 52nd Edition; Doha, Qatar

点击查看摘要

Abstract:Faults occurring during power swings, and the evolution of a swing into an unstable condition, cause distance-relay maloperation and cascaded outages. This paper identifies faults, faults during power swings, and power swings, and distinguishes stable from unstable swings in a 9-bus WSCC system modeled in PSCAD/EMTDC. Three time-series feature libraries (tsfresh, TSFEL, and tsfeatures) are applied to the 3-phase relay voltages and currents, giving 3286 statistical, temporal, spectral, and fractal descriptors per event window. A two-step feature selection (Boruta all-relevant filtering followed by RFECV- and SHAP-guided minimal-optimal pruning) reduces the feature set to 18 compact, non-redundant features. 14 classifiers spanning linear, kernel/instance-based, probabilistic, neural, and tree-ensemble/boosting families are then benchmarked under repeated stratified cross-validation, with model differences assessed by a Friedman/Nemenyi omnibus-and-post-hoc test and a Nadeau–Bengio corrected paired t-test. Rotation Forest (ROF) identifies the three event types with 99.98% and SVM-RBF distinguishes stable from unstable swings with 97.97% balanced accuracy. The results show that automated, library-driven feature extraction combined with rigorous statistical model comparison achieves near-perfect discrimination while retaining feature-level interpretability through SHAP importance.

[LG-201] Canonical activation patterns and ReLU pattern varieties

链接: https://arxiv.org/abs/2611.20346
作者: Bella Finkel,Yaoying Fu,Chenxi Wu
类目: Combinatorics (math.CO); Machine Learning (cs.LG); Algebraic Geometry (math.AG)
*备注:

点击查看摘要

Abstract:A ReLU pattern variety records the polynomial relations among the coefficients of the affine maps associated with a prescribed set of activation patterns of a ReLU network. We study how the combinatorics of the canonical set of activation patterns of an unbiased ReLU network is reflected in the geometry of its pattern variety. In particular, we introduce the notion of pattern stability and characterize pattern stability and combinatorial stability for shallow networks. When the associated wall graph is connected and contains an edge of every neuron label, we resolve a problem of Alexandr and Montúfar by determining the defining ideal of the associated pattern variety explicitly. We establish a hierarchy of progressively weaker equivalence relations relating labeled combinatorial equivalence, equality and permutation equivalence of canonical pattern sets, and equivalences of the associated varieties. Finally, we relate functional dimension to the differential of the pattern map and show that it is bounded above by the dimension of the canonical pattern variety at every depth. For nondegenerate shallow networks, we give an exact formula for the functional dimension; for every shallow architecture, it equals the dimension of the canonical pattern variety on an open dense set of full Lebesgue measure.

[LG-202] Measuring Coherence in Predictive Models NEURIPS2026

链接: https://arxiv.org/abs/2611.20325
作者: Jamie Reason,Vik Shirvaikar,Mark van der Wilk,Chris Holmes
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: Accepted at NeurIPS 2026

点击查看摘要

Abstract:Many sequential decision-making procedures, such as active learning, implicitly assume that a model’s current predictions agree with those it expects to make after updating on new data, a property we call coherence. Most models in use do not update by exact conditioning and are not guaranteed to be coherent, which can lead to worse decisions. Yet there is currently no general way to measure incoherence or to tell when it matters. We develop tools to measure incoherence and characterise when it affects decisions. We introduce the incoherence ratio, an MMD-based measure of how far any updating procedure departs from coherence, computed only from the procedure’s own predictions and updates, and accompany it with an anytime-valid test of coherence. In Bayesian experimental design, acquisition objectives that are equivalent for coherent models choose different data under every approximate model we study, and the incoherence ratio tracks how far incoherence distorts them. We also introduce action coherence, which captures only the incoherence that changes a decision. What a resulting disagreement costs, however, is determined primarily by predictive accuracy. Sequential decisions depend on how a model updates as well as on what it predicts, so coherence, like accuracy, should be checked rather than assumed.

[LG-203] Near-Optimal Methods for Generalized Lipschitz Convex Optimization

链接: https://arxiv.org/abs/2611.20318
作者: Dmitry Kovalev,Grigory Malinovsky
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:In this paper, we consider a convex optimization problem, where the upper bound on the subgradient norm depends on the objective function. We establish a lower bound on the number of iterations required by any first-order method to find an approximate solution to the problem. Furthermore, we present two near-optimal algorithms that can match the lower bound up to logarithmic factors. The first algorithm is a projected GD with a carefully designed stepsize schedule which incorporates warm-up and decay phases. The second algorithm is scalar-stepsize clipped AdamW which extends the near-optimal guarantees to the stochastic setting, where only a noisy unbiased estimator of the subgradient is available.

[LG-204] he Manifold-Index Model: provably and efficiently avoiding the statistical and computational curse of dimensionality

链接: https://arxiv.org/abs/2611.20275
作者: Yantao Wu,Mauro Maggioni
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Statistics Theory (math.ST)
*备注: 89 pages, 17 figures, 8 tables

点击查看摘要

Abstract:We study vector-valued regression with predictors filling a tube around an unknown m -dimensional manifold \mathcalM in \mathbbR^D . In the Manifold-Index Model, the mean response is F=f\circ\Pi_\mathcalM , where \Pi_\mathcalM is the nearest-point projection onto \mathcalM and f:\mathcalM\to\mathbbR^p is a possibly non-injective link function. Multiscale ANisotropic lOcal Regression (MANOR) estimates tangent planes by multiscale PCA of predictors with similar responses, then refines them by pooling outer products of local slopes. The neighborhoods narrow tangentially while retaining complete normal fibers. We assume a Lipschitz link differential and bounded response noise. For p\ge m , under our geometric and sampling conditions, n independent observations give mean tangent-space error of order n^-2/(m+4) and integrated squared prediction error at the m -dimensional nonparametric rate n^-4/(m+4) , up to logarithmic factors. In a scaling regime with fixed intrinsic dimension, model and scale parameters, and ratio of manifold volume to the m -th power of its reach, the required sample size grows polynomially in D+p , and computation is polynomial in n , D and p . Numerical experiments on curves with non-monotone links and surfaces with vector-valued responses compare observed convergence rates with the theoretical bounds; a Langevin dynamics example tests prediction of hitting probabilities under an approximate manifold-index model.

[LG-205] Phase Transitions in Heavy-Tailed Mean Estimation under ell_p Norms NEURIPS2026

链接: https://arxiv.org/abs/2611.20176
作者: Ishaq Aden-Ali,Yeshwanth Cherapanamjeri,Mikael Møller Høgsgaard,Kasper Green Larsen,Nikita Zhivotovskiy
类目: atistics Theory (math.ST); Machine Learning (cs.LG)
*备注: Accepted at NeurIPS 2026

点击查看摘要

Abstract:We study the problem of estimating the mean of a random vector in \mathbb R^d under \ell_p norms, assuming only the existence of a covariance matrix. In the Euclidean norm, the sample mean is already optimal in expectation, and robust estimators are usually needed only to obtain high-probability tails. Our results show that this picture changes for \ell_p norms with p2 : robustification is already necessary even for the expected risk. We prove that the minimax risk exhibits a sharp phase transition, controlled by the interaction between the sample size n , the dimension d , and the norm parameter p . In the large-sample regime, the classical Gaussian rate remains achievable. In the complementary regime, heavy-tailed distributions create a genuinely non-Gaussian obstruction. Furthermore we show that the simple estimator: coordinate-wise median of means whose number of buckets chosen according to the geometry of the loss, optimally tracks this phase transition in its expected risk minimax. We also show that natural projection-based median-of-means estimators for general norms, despite achieving the optimal confidence trade-off relative to the sample-mean benchmark of Lugosi and Mendelson (Probab. Theory Relat. Fields, 2019), can be polynomially suboptimal for \ell_p norms even in the constant-probability regime.

[LG-206] Cross-Subject Priors for Maximum a Posteriori Estimation of Latent Structure Influence Models in Motor-Imagery EEG With Few Trials

链接: https://arxiv.org/abs/2611.20164
作者: Sagnik De,Anil Kumar Gupta
类目: ignal Processing (eess.SP); Machine Learning (cs.LG)
*备注: 9 pages, 4 figures

点击查看摘要

Abstract:Coupled hidden Markov models such as the latent structure influence model (LSIM) describe how the sensorimotor rhythms of the two hemispheres evolve and interact during motor imagery, but subject-specific maximum likelihood (ML) estimates are unstable when few trials are available. This paper derives a maximum a posteriori (MAP) estimator for the LSIM whose conjugate prior equals the expected statistics of \lambda pseudo-trials from a population model fitted to other subjects. The M-step remains in closed form, expectation maximization keeps its monotone ascent, and ML ( \lambda=0 ) and the population model ( \lambda\to\infty ) are limiting cases. The prior strength was selected on 35 subjects of the PhysioNet EEG Motor Movement/Imagery Dataset and applied unchanged to 70 held-out subjects and to BCI Competition IV-2a. With \lambda=1 the held-out log-likelihood exceeded that of ML in 61 of 70 subjects (Holm-adjusted p=9.1\times10^-8 ) and in all 70 subjects with five training trials per class. A prior learned only from the PhysioNet data improved the held-out likelihood of all nine BCI IV-2a subjects at five trials per class and reduced the error of the joint transition matrix by 61%. Decoding accuracy did not improve ( p=0.17 ). Cross-subject priors therefore stabilize few-trial dynamic models of motor-imagery EEG, but they do not improve decoding, which in these data relies mainly on structure shared across subjects.

[LG-207] A plug-in energy-balance freezing module for transformer frost nowcasting without leaf-sensor inputs

链接: https://arxiv.org/abs/2611.20114
作者: Juyeon Kim,Chaeyeon Yi,Yun Am Seo
类目: Applications (stat.AP); Machine Learning (cs.LG); Atmospheric and Oceanic Physics (physics.ao-ph)
*备注: 45 pages main text + 19 pages supplementary material

点击查看摘要

Abstract:Frost protection requires local warnings where leaf-sensor inputs are unavailable. In the frost literature we surveyed, physics entered machine-learning models through inputs, numerical-model coupling, or losses, rather than model structure. We develop a transformer-compatible module that embeds leaf energy-balance physics in freezing-probability forecasts: an energy-balance temperature reference and a learned residual predict leaf temperature, which drives a multiplicative gate imposing a temperature-dependent probability cap. We mathematically prove conditional monotonicity and an alarm boundary, and test whether placing the same information in the output pathway or in attention yields different benefits. We evaluated 12 hourly leads at 19 Korean stations over five frost seasons, with and without leaf-sensor inputs. In the 2025-2026 hold-out season without these inputs, our time-token transformer (PET-E) achieved a lead-mean critical success index (CSI) of 0.516 versus 0.478 for fixed-configuration, joint-lead LightGBM, with significant gains at all 12 leads, and the fraction of alarms at observed leaf temperatures of at least 2 degrees C fell from 0.087 to 0.051 relative to the unmodified backbone. Across the hold-out and a retrospective 2024-2025 evaluation, the module reduced Brier scores and warm-surface alarm fractions in both the time-token transformer and iTransformer, with no significant CSI decrease at any lead. At held-out stations without static or leaf-sensor inputs, both module-equipped backbones exceeded LightGBM at all 12 leads (CSI 0.468 and 0.467 versus 0.437). In the time-token backbone, the temperature pathway reduced this alarm fraction in all four settings, whereas the attention prior increased CSI in three. These findings link an explicit temperature-dependent probability cap to improved forecasts and distinguish output-side from attention-based embedding.

[LG-208] Washout-free synchronization parameter estimation and uncertainty quantification of echo state networks

链接: https://arxiv.org/abs/2611.20112
作者: Andrea Nóvoa
类目: Chaotic Dynamics (nlin.CD); Machine Learning (cs.LG)
*备注: 15 pages, 12 figures

点击查看摘要

Abstract:Echo state networks (ESNs) forecast chaotic systems for several Lyapunov times. Like all recurrent networks, ESNs require a washout before forecasting, i.e., a phase in which the full state drives the network until the reservoir state synchronizes with the system. At deployment, however, sensors measure only part of the state, with noise and at a lower sampling rate than the training data. We propose a washout-free deployment of ESNs via ensemble data assimilation. In closed loop, the ESN is a Markovian dynamical system; hence, a sequential filter can correct its reservoir state whenever data are available. The filter synchronizes an ensemble of reservoir states, drawn from the training data, with the system, and quantifies the forecast uncertainty. We extend the method to parametric ESNs, to infer physical parameters, and to parallel ESNs with localization, to forecast high-dimensional systems. We test the method on the Lorenz-63 and Lorenz-96 systems. First, the filter synchronizes the ESN with a fraction of the data of a washout. Second, the filter deploys one full-state ESN from a subset of its components, which a washout cannot do. Third, the filter estimates a forcing that is not in the training data, and the estimated forcing lengthens the forecast. Fourth, localization synchronizes a parallel ESN with eight times fewer members than the global filter. Fifth, the ensemble spread signals, without the true state, when the ESN synchronizes and when the filter fails. This work enables the deployment of ESNs from partial, sparse and noisy observations.

[LG-209] Kolmogorov-Arnold Lookup Table Demappers for Amplitude Phase-Shift Keying over Nonlinear Satellite Links

链接: https://arxiv.org/abs/2611.20068
作者: Kürşat Tekbıyık,Güneş Karabulut Kurt,Antoine Lesage-Landry
类目: ignal Processing (eess.SP); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Next-generation high-throughput satellites increasingly demodulate and decode on board, so the per-symbol cost of the receiver constrains payload power and throughput. Soft demappers for nonlinear satellite links must therefore produce accurate bit log-likelihood ratios (LLRs) at low cost. We introduce a KAN-LUT demapper, a two-layer Kolmogorov–Arnold network (KAN) whose learned univariate edge functions are stored as 128-entry lookup tables (LUTs), so that inference needs only table reads, linear interpolation, and additions. We bound the LLR error introduced by the tables and evaluate the demapper on amplitude phase-shift keying (APSK) links with low-density parity-check (LDPC) coding and a nonlinear traveling-wave-tube amplifier. For 16-, 32-, and 64-APSK, KAN-LUT stays within 0.2 dB of a per-cluster Gaussian log-MAP (LogMAP-CG) demapper, while its cost relative to LogMAP-CG falls from 87% to 27% as the order grows. It matches or slightly exceeds multilayer perceptrons (MLPs) at about one fifth of their cost, remains competitive under operating-point mismatch, and needs only tens of kilobytes of tables. KAN-LUT is therefore a low-cost candidate demapper for regenerative payloads.

[LG-210] SWAN-Rx: Efficient and Bandwidth-Scalable Neural Reception for MIMO-OFDM

链接: https://arxiv.org/abs/2611.20058
作者: Zhoubin Kou,Cong Shen
类目: ignal Processing (eess.SP); Information Theory (cs.IT); Machine Learning (cs.LG)
*备注: This work has been submitted to IEEE for possible publication

点击查看摘要

Abstract:We present a Scalable Windowed Axial-attention Neural Receiver (SWAN-Rx) for multiple-input multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) systems. The receiver improves reception accuracy and reduces forward-pass computation while supporting frozen-weight reuse across bandwidth allocations and channel conditions. It combines a common resource-element representation of received observations and pilot context with alternating time-axis and local frequency-axis attention. Shared parameters and relative positional information permit soft detection across allocation lengths without inference-time parameter updates. At fixed subcarrier spacing, model configuration, and symbol count, bounded frequency neighborhoods give linear growth of forward-pass computation with bandwidth. Evaluations compatible with the 5G New Radio physical layer reuse weights trained on four physical resource blocks (PRBs) at up to 273 PRBs and on unseen channels. At the training bandwidth, SWAN-Rx achieves lower pre- and post-decoding bit error rates than the implemented three-step neural receiver (NRX) and DeepRx-MIMO baselines, with 16.9% and 39.7% fewer parameters and 56.3% and 35.7% fewer forward-pass operations, respectively. The same weights achieve lower pre-decoding error rates than these baselines throughout the evaluated bandwidth-extrapolation and 16-PRB channel-transfer tests, with an estimated 0.4-0.9 dB signal-to-noise ratio reduction relative to NRX at a pre-decoding error rate of 10^-2 in the 16-PRB tests.

[LG-211] U-HamGNN: a universal model for magnetic Hamiltonians conditioned on spin and Hubbard U

链接: https://arxiv.org/abs/2611.20035
作者: Hongyu Yu,Yinze Hu,Yizhou Chen,Yang Cheng,Yingwei Chen,Xingao Gong,Yang Zhong,Hongjun Xiang
类目: Computational Physics (physics.comp-ph); Materials Science (cond-mat.mtrl-sci); Machine Learning (cs.LG)
*备注: 25 pages, 5 figures

点击查看摘要

Abstract:The electronic structure of a magnetic crystal depends jointly on its spin configuration and the on-site Hubbard correction U, requiring a separate DFT+U calculation for each combination. We introduce U-HamGNN, an equivariant network that takes these two inputs directly and predicts spin-resolved Hamiltonians across diverse crystal chemistries. A spectral-hybrid optimizer that weights each update direction by how consistently the gradient points along it reaches 8% lower Hamiltonian error than Muon and 26% lower than AdamW in a ten-epoch benchmark. The predicted bands reproduce U-dependent gap opening and altermagnetic spin splitting within the trained window and beyond it; exchange couplings and magnetic-order calls derived from the Hamiltonian agree with DFT+U on a calibration set. This universal model provides a new tool for magnetic research.

[LG-212] owards GPROF V08: A Machine-Learning Framework for Simulating Passive Microwave Observations of Precipitation

链接: https://arxiv.org/abs/2611.19966
作者: Simon Pfreundschuh,Christian D. Kummerow,Paula J. Brown
类目: Atmospheric and Oceanic Physics (physics.ao-ph); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Spatially resolved precipitation retrievals based on convolutional neural networks (CNNs) can improve precipitation estimates from passive microwave (PMW) observations and will be used to implement the next operational PMW retrieval algorithm for the Global Precipitation Measurement (GPM) mission. However, the conventional radiative transfer simulations used to extend the retrieval algorithm to the various sensors of the GPM constellation do not provide the spatial coverage required to train CNN-based retrievals. To address this, we introduce a machine-learning-based simulation framework that enables the training of the CNN-based GPROF-NN 3D retrieval algorithm for the full GPM constellation. At the core of this framework is Satformer, a new neural-network architecture that can translate PMW observations between any pair of GPM sensors. The resulting synthetic observations exhibit lower errors than conventional radiative transfer simulations and lead to more accurate precipitation retrievals when used to train GPROF-NN precipitation retrievals. We further demonstrate the applicability of the framework to new instruments by developing GPROF-NN retrievals for the Microwave Imager aboard the Weather System Follow-on Microwave satellite. While GPROF-NN retrievals based on the new simulation framework achieve lower instantaneous retrieval errors and higher correlations compared to independent reference data than conventional retrievals, residual simulation errors can introduce systematic biases in the resulting retrievals. The proposed simulation framework provides a key component for extending GPROF-NN to the full GPM constellation as part of the upcoming operational GPROF V08 retrieval algorithm. The results demonstrate the potential of machine-learning-based simulations to support spatially resolved precipitation retrievals across heterogeneous satellite constellations. Subjects: Atmospheric and Oceanic Physics (physics.ao-ph); Machine Learning (cs.LG) Cite as: arXiv:2611.19966 [physics.ao-ph] (or arXiv:2611.19966v1 [physics.ao-ph] for this version) https://doi.org/10.48550/arXiv.2611.19966 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-213] Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization for Multi-Asset Portfolios

链接: https://arxiv.org/abs/2611.19903
作者: Supriadi
类目: Portfolio Management (q-fin.PM); Machine Learning (cs.LG)
*备注: 10 pages, 7 figures, 2 tables. Replication package and dataset available at this https URL

点击查看摘要

Abstract:Nested Clustered Optimization (NCO) insulates asset allocation from Markowitz covariance ill-conditioning. However, classical tree clustering relies on static pairwise Pearson correlation, failing to capture time-varying, asymmetric downside risk contagion across multi-asset markets. We propose a unified deep graph machine learning architecture: Dynamic Spatio-Temporal Graph Attention Networks with Nested Clustered Optimization (ST-GAT-NCO) for multi-asset tactical allocation. The framework extracts temporal momentum and downside risk dynamics via causal dilated 1D-convolutions, then deploys a multi-head Graph Attention Network (GAT) over dynamic market topologies to learn latent spatial node representations. These learned embeddings define an angular metric distance space that guides two-stage downside NCO optimization under institutional constraints and turnover smoothing. Evaluated over a 73-month non-anticipative walk-forward out-of-sample window (January 2020 to January 2026) across 13 global asset classes net of 10 bps transaction costs, ST-GAT-NCO achieves a 5.32% CAGR and 37.06% cumulative net return, delivering an excess Sharpe ratio of 0.30 and a Sortino ratio of 0.48. This substantially outperforms classical Minimum Variance (3.68% CAGR, 0.08 Sharpe), Hierarchical Risk Parity (2.65% CAGR), and Plain Symmetric NCO (4.11% CAGR, 0.15 Sharpe), while restricting maximum drawdown to -13.07% (versus -19.44% for Equal-Weight and -16.36% for Downside NCO) under a controlled one-way turnover of 15.31%/month. Regime analysis confirms active capital preservation across the 2020 COVID-19 crash, 2022 monetary tightening, and 2023-2026 disinflation rallies.

[LG-214] First-order smoothed exact-penalty methods for equality-constrained regularized nonconvex optimization

链接: https://arxiv.org/abs/2611.19887
作者: Shiji Zuo,Qihang Lin,Wei Liu,Yangyang Xu
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:We consider regularized nonconvex optimization over a closed convex set \mathcalX \subseteq \mathbbR^d , which involves a convex nonsmooth regularizer r in the objective and n smooth equality constraints. To find a near-KKT point of the problem, we propose a prox-linear first-order method, named as SEPi-Prox, based on a smoothed exact-penalty formulation. At each outer iteration of SEPi-Prox, a strongly convex prox-linear subproblem is formed and solved inexactly to a sufficient accuracy. Under a local non-singularity condition that is weaker than some other commonly used regularity conditions or constraint qualifications, SEPi-Prox finds an \epsilon -KKT point within O(\epsilon^-2) first-order oracle calls, matching the lower bound for unconstrained smooth nonconvex optimization. We employ Nesterov’s accelerated proximal-gradient method (APG) and a dual cutting-plane method (DCP) for the prox-linear subproblems. Let \mathcalC_\mathrmFO and \mathcalC_r denote the numbers of arithmetic operations of evaluating one first-order oracle and one proximal mapping of r over \mathcalX , respectively. Then SEPi-Prox requires \widetildeO(\mathcalC_\mathrmFO\epsilon^-2 + (dn + \mathcalC_r)\epsilon^-5/2) and \widetildeO((\mathcalC_\mathrmFO + dn^2 + n\mathcalC_r + n^3.373)\epsilon^-2) total arithmetic operations to produce an \epsilon -KKT point, respectively by using APG and DCP as the subsolver. The latter result is preferable when n is small and nearly matches the optimal dependence on \epsilon . Numerical experiments on instances of the multi-class Neyman-Pearson classification demonstrate the efficiency of SEPi-Prox with both subsolvers.

[LG-215] Sharp Gaussian Approximation for Q-Learning

链接: https://arxiv.org/abs/2611.19850
作者: Xingtu Liu
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Statistics Theory (math.ST)
*备注:

点击查看摘要

Abstract:Gaussian approximation provides a foundation for online statistical inference in reinforcement learning. We establish sharp non-asymptotic Gaussian approximation rates in Wasserstein distance for asynchronous Q-learning with Polyak-Ruppert averaging. We prove an upper bound of \widetildeO(n^-1/4) , where n is the number of iterations, together with a matching lower bound up to logarithmic factors. Q-learning is more challenging to analyze than linear stochastic approximation due to the nonlinearity of the Bellman operator. Asynchronous updates introduce further challenges due to Markovian noise. Our analysis separates the errors from Bellman nonlinearity, Markovian noise, and the finite-sample covariance discrepancy. In particular, we show that Markovian noise induces an unavoidable bias even when the Bellman operator is linear. The trade-off between this bias and the covariance discrepancy determines the convergence rate. We further extend our analysis to the synchronous setting and establish a sharp rate of \widetildeO(n^-1/2) .

[LG-216] ProTTS: End-to-End Prominence-aware Natural and Controllable Text-to-Speech via Automatic Word Stress Discovery

链接: https://arxiv.org/abs/2611.19832
作者: Sai Harshitha Aluru,Chiranjeevi Yarra
类目: Audio and Speech Processing (eess.AS); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Natural speech conveys meaning not only through words, but also through the prominence on a few words. Yet most text-to-speech (TTS) systems remain prominence-unaware, existing prominence-aware TTS require external stress annotations, while unsupervised stress detectors operate in isolation, disconnected from synthesis. This creates a fundamental gap no existing system can automatically discover where stress should be placed, realize it as natural-sounding prosody, and offer user control over emphasis placement, all within a single model. To circumvent this, we propose ProTTS, a unified framework where stress detection emerges from the TTS model itself and, in turn, guides it to synthesize more natural and controllable speech. We incorporate a clustering head into the variance adaptor of FastSpeech_2 that discovers word stress directly from aggregated prosodic features. The discovered stress then drives emphasis variance predictors that modify pitch, energy, and duration for stressed words. The framework is trained end-to-end: TTS reconstruction and clustering losses jointly shape a stress-discriminative feature space without requiring stress labels at any stage. At inference, the same architecture also supports user-controllable emphasis by accepting a custom stress mask. We evaluate ProTTS on ISLE (non-native), LibriTTS, VCTK (native training), and three evaluation benchmarks (Tatoeba, TinyStress-15k, EmphAssess). Listeners prefer ProTTS over the baseline with a MOS of 3.96 vs. 3.12 and 83.5% preference rate. For stress detection, ProTTS achieves 0.839/0.861 F1 on ISLE (German/Italian L2), outperforming unsupervised (0.70) and supervised (0.82) baselines, and 0.933 F1 in zero-shot transfer to Tatoeba.

[LG-217] Hermitian Positive Definite Neural Networks for Matrix-CFAR Geometric Signal Detection in Nonhomogeneous Clutter

链接: https://arxiv.org/abs/2611.19820
作者: Xiyao Chen,Takumi Sato,Linyu Peng
类目: ignal Processing (eess.SP); Machine Learning (cs.LG); Applications (stat.AP)
*备注: 11 pages, 9 figures

点击查看摘要

Abstract:Matrix information geometry (MIG) detectors exploit the Riemannian geometry of Hermitian positive definite (HPD) matrices to achieve robust object detection in nonhomogeneous clutter. The matrix constant false alarm rate (matrix-CFAR) method is a representative MIG detector that relies on the geometric structures with respect to specific metrics. In most cases, computing representative HPD covariance matrices requires numerical methods, typically involving computationally intensive iterative optimization algorithms. Although explicit formulas are available for certain cases, such as the log-Euclidean (LE) mean, they are generally unavailable for many other matrix means and medians. To address this challenge, we propose integrating the Hermitian positive definite neural network (HPDNet) into the matrix-CFAR framework, replacing conventional numerical optimization procedures with a learned, geometry-preserving mapping that directly estimates representative HPD covariance matrices from data. In the simulations, the framework is evaluated for MIG detectors using both means and medians associated with the affine-invariant Riemannian metric, the LE metric, and the Bures–Wasserstein metric, on both simulated compound-Gaussian clutter and real IPIX sea clutter. In both cases, the HPDNet is trained solely on simulated data to assess its applicability to the study of real data. The results show a significant reduction of computational cost in testing, while maintaining detection performance and robustness.

[LG-218] IANN: A Unified Framework for Equivariant InterAtomic Neural Networks with User-Friendly Implementation

链接: https://arxiv.org/abs/2611.19793
作者: Changzhi Ai,Xueping Qin,Dimosthenis Sokaras,Frank Abild-Pedersen,Johannes Voss
类目: Materials Science (cond-mat.mtrl-sci); Machine Learning (cs.LG); Computational Physics (physics.comp-ph)
*备注: 31 pages, 7 figures, 4 tables; Supplementary Information included (49 figures, 11 tables). IANN is MIT-licensed. Code: this https URL - Documentation: this https URL - Curated DFT databases: this https URL - Foundation models: this https URL

点击查看摘要

Abstract:Machine-learning interatomic potentials (MLIPs) now approach density functional theory (DFT) accuracy at a fraction of its cost, but the practical benefit is limited by software fragmentation: each state-of-the-art architecture is released as a separate codebase, with its own data format, configuration system, training loop and molecular-dynamics interface. Comparing two architectures on the same dataset, or moving a trained model into a production simulation, therefore requires substantial re-engineering. We introduce IANN (InterAtomic Neural Network), an open-source PyTorch framework that reimplements equivariant graph neural networks - PaiNN, NequIP, MACE, Allegro, EquiformerV2, EquiformerV3 and UMA - behind a unified data object, a unified trainer and a single deployment route, so that the architecture is a one-word change in a configuration dictionary. Forces, stresses and virials are obtained for every architecture from one analytic differentiation of the energy with respect to the edge vectors of the atomic graph, exact under periodic boundary conditions. Distributed training is configured automatically from the SLURM, PBS or MPI environment, so the same script runs unchanged from a workstation to a multi-node GPU allocation. Self-describing checkpoints serve directly as ASE calculators, as ensembles reporting uncertainties, or as TorchScript models run inside LAMMPS through bundled C++ pair styles with single- and multi-GPU inference. Trained on a common dataset of millions of structures, the architectures can be compared under identical conditions. We further release PaiNN foundation models trained on 41.9 million curated DFT structures from MPtrj, Alexandria, OMat24, MatPES, OC20, OC22, OC25, MPr2SCAN and MatPESr2SCAN, spanning the PBE, RPBE and r2SCAN functionals. Additionally, IANN provides an agent interface to facilitate integration with autonomous agents.

[LG-219] PIRanNN: Random-Feature Pricing and Calibration under Jump-Diffusion and Stochastic-Volatility Models

链接: https://arxiv.org/abs/2611.19792
作者: Saurabh Bansal,Srikanth K. Iyer,Shashi Jain
类目: Computational Finance (q-fin.CP); Machine Learning (cs.LG); Numerical Analysis (math.NA)
*备注: 29 pages, 4 figures, 14 tables. Code: this https URL

点击查看摘要

Abstract:We study a random-feature method for pricing and calibrating European options under jump-diffusion (Merton, Kou) and stochastic-volatility (Heston, SABR) models. The network has one hidden layer with a probit activation and fixed random inner weights, so fitting it to the pricing equation, the payoff and the boundary conditions is a single regularised least-squares problem, with no iterative training. For Merton and Kou the jump integral of each feature is explicit, so no quadrature is needed. One factorisation of the least-squares matrix gives the price at every state and time, its Greeks, and the exact derivatives of the prices with respect to the model parameters for gradient-based calibration. For the jump models, subtracting the Black-Scholes price removes the payoff kink and reduces the pricing error by about two orders of magnitude. The method is then about as accurate as a standard Crank-Nicolson finite-difference scheme at the same run time, although a finite-difference scheme that also subtracts the Black-Scholes price is about two orders of magnitude more accurate still. On the test problems of a recent physics-informed neural network study, the method matches that network’s accuracy in seconds instead of about half an hour. The calibration experiments show that a solver can fit the quotes and still return wrong parameters when its pricing error is not small compared with the effect of the parameters on the prices: without the subtraction the solver misses the jump parameters by up to 50%, and with it they are recovered to within 3%. For Heston and SABR, used as an applicability study with boundary conditions that need no reference prices, pricing errors are about one percent and the calibrated parameters are within about 15% of the true values.

[LG-220] Conditional Asset Pricing Transformer

链接: https://arxiv.org/abs/2611.19783
作者: Hao Fu,Xinbing Kong,Zeyu Li,Junfan Mao
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:In this paper, we propose a conditional asset pricing transformer (CAP-T) that captures nonlinear interactions among firms through attention to forecast excess returns. In the first stage, a transformer estimates conditional expected returns, and kernel-weighted principal component analysis of its residuals estimates the local factor loading space. Projecting the conditional mean onto this space separates factor risk compensation from conditional mispricing. Then, in the second stage, separate transformers learn these two components, whose predictions are combined to forecast returns. As a result, the proposed framework decomposes returns into conditional mispricing, factor risk compensation, a centered factor component, and an unpredictable innovation. We establish approximation and generalization bounds and characterize how first-stage prediction errors propagate through local factor estimation and component learning, providing statistical guaranties for conditional mispricing, factor loadings, and factor risk premia. Finally, simulations and empirical studies demonstrate the predictive and interpretive benefits of the proposed method.

[LG-221] ght Sequential Confidence Intervals for Online Kernel Regression

链接: https://arxiv.org/abs/2611.19772
作者: Haichen Hu,David Simchi-Levi
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Probability (math.PR)
*备注:

点击查看摘要

Abstract:We characterize, in the worst case over kernels, the sharp minimax width of online confidence intervals for elements of a reproducing kernel Hilbert space (RKHS). This resolves the open problem posed by Vakili et al. (2021) in its original discrete-time form. A learner observes y_i=f(x_i)+\epsilon_i at adaptively chosen points, with |f|_\mathcal H_k\le B and i.i.d. R -sub-Gaussian noise, and reports the interval |f(x)-\mu_n(x)|\le\rho_n(\delta)\sigma_n(x) , where \mu_n and \sigma_n are the posterior mean and standard deviation of a surrogate Gaussian process model with regularization parameter \lambda . Writing \rho_n^(\delta) for the minimax confidence width for a fixed kernel, we show that, for the sum of a linear kernel and a periodic Sobolev kernel, all sufficiently large n , and all \delta\in(0,1/4] , \rho_n^(\delta)\asymp B+(R/\lambda)\sqrt\gamma_n+\log(1/\delta) , where \gamma_n is the maximal information gain and grows polynomially in n . Building on Lattimore (2023), we construct an adaptive design on a fixed two-dimensional domain for which, under Gaussian noise, the normalized error at a prespecified point is \Omega(B+(R/\lambda)\sqrt\gamma_n) with constant probability. This rules out the conjectured polylogarithmic width in the worst case over kernels and online designs.

[LG-222] Where the Error Lives: Separating Lid-Edge and Body-Induced Error in Neural Pressure Prediction for 3D Cavity Flow

链接: https://arxiv.org/abs/2611.19741
作者: Zirun Zhou,Zhi-Qian Zhang,Yixuan Hong,Guang Liu,Zhigang Liu,Lili Zhang
类目: Fluid Dynamics (physics.flu-dyn); Machine Learning (cs.LG)
*备注: 37 pages, 10 figures

点击查看摘要

Abstract:Can a pressure surrogate score much better globally without better resolving the pressure around immersed bodies? We test this in a three-dimensional lid-driven cavity with an unregularized lid and centered pressure at native 128^3 resolution. On 60 cases from 12 unseen geometries, with predictions sealed before the CFD truth existed, adding a body-free, Reynolds-conditioned reference correction to the same frozen 64^3 surrogate lifted to 128^3 lowers the mean full-volume relative L_2 error from 20.6% to 7.5%. In the post hoc regional analysis, the lid-edge junction occupies 1.5% of the fluid but holds 69% of the truth pressure energy and 95% of the uncorrected surrogate’s squared error. Interior error changes by just +0.16% relative, and object-local peak error and recall change little on average. Retrospective ablations locate the learned term’s contribution in the interior and near the bodies; rebasing on the reference lowers global error further with little change in cohort-mean near-body metrics. Reference-relative body-response error remains about 25% in the interior near-body band. Grid-refinement, lid-regularization and convergence checks delimit the numerical scope. In this benchmark, a large global gain chiefly reflects correction of a shared boundary field, not better resolution of the body response. We recommend reporting regional truth-energy and model-error shares alongside near-body and reference-relative body-response errors.

[LG-223] Relax-and-Sparsify Approximation Algorithms for Sparse Ridge Regression

链接: https://arxiv.org/abs/2611.19738
作者: Shixuan Yi,Yongchun Li
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:This paper studies sparse ridge regression in the low-rank, high-dimensional setting that often arises when data collection is expensive. By exploiting the low-rank structure, we develop approximation algorithms within a relax-and-sparsify framework: solve a convex relaxation and then sparsify its solution to satisfy the original sparsity constraint. Specifically, we establish a tight rank-dependent sparsity bound for the convex relaxation by decomposing its support into two parts. This decomposition in turn motivates two approximation strategies, parameter reduction and binary redistribution, each exploiting a different part. Experiments on synthetic and real datasets show that the binary redistribution algorithms outperform the benchmarks in most instances.

[LG-224] Set-based State Estimation for Unknown Dynamical Systems from Noisy Output Data with Probabilistic Guarantees

链接: https://arxiv.org/abs/2611.19716
作者: Georgios Pantazis,Michelle S. Chong
类目: Optimization and Control (math.OC); Machine Learning (cs.LG); Systems and Control (eess.SY)
*备注:

点击查看摘要

Abstract:We develop a methodology for set-based state estimation from noisy input-output trajectories, without a predefined model structure for the unknown state dynamics or access to training state measurements. We derive a bound on the state set diameter that depends on the sensor map and leverage it as a surrogate to formulate a tractable linear program balancing state set tightness against output trajectory coverage. Combining this certificate with deterministic or probabilistically certified measurement noise sets and leveraging the results from scenario optimization with relaxation, we provide joint a posteriori probabilistic guarantees for state inclusion and state set diameter. Numerical simulations corroborate our developed theoretical guarantees.

[LG-225] Minimax Estimation of Discrete Distributions under Hockey-stick Divergence

链接: https://arxiv.org/abs/2611.19702
作者: Manaka Ueno,Geoffrey Wolfer
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Statistics Theory (math.ST)
*备注:

点击查看摘要

Abstract:We initiate the study of estimating a discrete distribution with respect to the hockey-stick divergence from independent samples. We provide non-asymptotic upper and lower bounds on the minimax expected loss in terms of the alphabet size, the number of samples, and the divergence parameter. The upper bounds are achieved by add-constant estimators, while the lower bound is derived via a Bayes risk argument. Both upper bounds and lower bounds reveal a transition between polynomial and exponential regimes as the parameters of the problem vary.

[LG-226] A Geometric Understanding of Denoising Autoencoder with Applications to Image Denoising

链接: https://arxiv.org/abs/2611.19653
作者: Haoran Yan,Rongjie Lai,Wenjing Liao
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Denoising autoencoders have shown strong empirical performance in recovering clean data from their noisy observations, yet their underlying statistical behavior remains underexplored. In this paper, we study denoising autoencoders under the manifold hypothesis, assuming that clean data are supported on low-dimensional smooth manifolds embedded in a high-dimensional ambient space, while noisy observations are generated by isotropic Gaussian noise. We establish generalization guarantees in three geometric scenarios: a single manifold, a pairwise separated finite union of manifolds, and a finite union of intersecting manifolds. Using local chart representations, partitions of unity, and neural network approximation, we derive upper bounds for the expected squared reconstruction error for denoising autoencoder. In the single-manifold and separated-union scenarios, the error consists of a sample-dependent statistical term of order n^-2/(d+2)\log^4 n , a residual term of order d\sigma^2/D caused by tangential noise, and an exponentially small tail term associated with failure of the manifold projection. Here D is the data ambient dimension, d is the intrinsic dimension of the clean-data manifold, and \sigma^2 is the expected squared noise magnitude. In the scenario of intersecting manifolds, we further identify an additional ambiguity term determined by the probability mass near the intersection neighborhood. Numerical experiments on synthetic manifolds and real image datasets support the predicted sample-size scaling law and illustrate the additional reconstruction ambiguity near manifold intersections.

[LG-227] NAP-Wave: A Neural Adaptive Plane-Wave Architecture for High-Frequency Helmholtz Problems

链接: https://arxiv.org/abs/2611.19603
作者: Hamza Kamil,Mohammad Mahdi Abedi,David Pardo
类目: Geophysics (physics.geo-ph); Machine Learning (cs.LG)
*备注: 35 pages, 15 figures, 7 tables

点击查看摘要

Abstract:Training neural networks to simulate highly oscillatory wavefields presents optimization challenges, often hindered by spectral bias and poorly conditioned physics-based loss landscapes. To address these challenges and enhance convergence, we propose NAP-Wave, a joint neural architecture and training strategy that expands the target scattered wavefield in an adaptive neural plane-wave basis. The basis functions are extracted from all hidden layers and modulated by trainable plane-wave carriers, embedding the high-frequency oscillatory physics directly into the network’s trial space. NAP-Wave is trained on a physics-informed loss built from the Lippmann-Schwinger integral equation, so no differentiation of the network with respect to its input is required. We develop a decoupled optimization strategy in which Adam updates the parameters of the neural plane-wave basis, while the linear combination coefficients are computed exactly by a least-squares (LS) solver. We couple this with an adaptive-depth training strategy in which the network starts shallow and grows progressively layer-by-layer. At each growth step, we avoid the loss spikes typical of dynamic network growth using the LS-computed coefficients, the warm-started parameters, and the preservation of the first- and second-order optimizer moments of the previous stage. Ablation studies confirm that the plane-wave basis, the exact coefficient solve, and the progressive schedule independently accelerate convergence and improve accuracy. Validated on exact-solution benchmarks, NAP-Wave achieves relative L2 errors of order 1%. On highly heterogeneous geophysical benchmarks, it achieves relative errors of 5%, 9%, and 10%, significantly outperforming published PINN methods in both accuracy and convergence time. To ensure reproducibility, the code is available at this https URL.

[LG-228] Steering Bayesian Inference and Decisions through Data Poisoning

链接: https://arxiv.org/abs/2611.19557
作者: Roi Naveiro,William N. Caballero,Juan Maroñas
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: 64 pages, 13 figures

点击查看摘要

Abstract:Adversarial machine learning has revealed that statistical models can be manipulated through poisoned data, and Bayesian inference is no exception. Earlier work has shown that subtle poisoning attacks can drive the posterior toward a distribution of the attacker’s choosing. Such posterior attraction attacks require a fully specified target posterior; however, attackers typically care about only a few posterior quantities or, more consequentially, about a downstream decision. Herein, we study how an attacker can alter Bayesian inference and decisions by deleting or replicating a few observations, chosen to minimize the maximum mean discrepancy between the resulting posterior and a target distribution. A suitable kernel lets the attacker steer selected posterior moments without specifying the rest of the target. Because Bayes actions depend on the posterior only through expected utilities, which are themselves posterior moments, the same approach can target a decision directly. We construct the attacks with a gradient-based heuristic that requires only posterior samples, and analyze its convergence. Experiments with regression and hierarchical models compare moment-based and full-posterior attacks, and illustrate how deleting a handful of observations can flip a Bayesian decision.

[LG-229] Geometric Stability and Sample Complexity for Compositional Function Priors NEURIPS2026

链接: https://arxiv.org/abs/2611.19523
作者: Christopher L. Hammond
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: 25 pages. An extended abstract was accepted at the NeurIPS 2026 Workshop on Symmetry and Geometry in Neural Representations (NeurReps), non-archival. Ancillary file: companion document with extended proofs and examples

点击查看摘要

Abstract:We study when compositional structure yields dimension reduction in supervised learning. We frame learning as empirical risk minimization with a function-space prior: a regularization functional selects a subset of functions with favorable approximation and metric entropy. The models are compositions f(x)=g(Ax+\Delta h(x)) , where A\in\mathbbR^k\times d is a linear base map, \Delta h is a nonlinear correction, and g is an outer function; that is, f is an arbitrary smooth function of a k -dimensional representation of the input that is linear up to a correction. We prove covering-number bounds showing that the statistical complexity of this class interpolates between an intrinsic regime, parametric in the kd entries of the linear map and nonparametric only in the representation dimension k , and an ambient d -dimensional nonparametric rate; the interpolation is controlled by the correction size \varepsilon=|\Delta h| . We then connect \varepsilon to data geometry. If the data lie in a small tube around a k -dimensional manifold and the regression target depends on the input only through its nearest point on that manifold, then locally the target factorizes exactly as g(Ax+\Delta h(x)) with A a tangent chart and a correction whose Sobolev norm scales linearly with curvature at the localization scale. Patching such local models over an atlas gives a finite-sample excess-risk bound for a least-squares estimator over a finite net of local models, whose ambient term vanishes with curvature. For a one-dimensional model problem we also prove a converse: on curved data, every affine compositional predictor faces an error floor unless its correction budget scales with curvature, while the matching curved model class attains the intrinsic rate. The correction size \varepsilon is thus the effective-dimension parameter linking curvature to statistical complexity.

[LG-230] Equilibrium measures for anisotropic exponential interaction energies with applications to transformers

链接: https://arxiv.org/abs/2611.19518
作者: Oleksii Kachaiev,Martin Burger,Samira Kabri,Yury Korolev,Tim Roith
类目: Analysis of PDEs (math.AP); Machine Learning (cs.LG); Probability (math.PR)
*备注: 23 pages, 3 figures

点击查看摘要

Abstract:We study the minimization of anisotropic exponential interaction energies over probability measures on the sphere, motivated by mean-field models of self-attention in transformer architectures. For every symmetric positive definite interaction matrix, we relate the minimizer to a corresponding anisotropic Brownian motion and use the Feynman–Kac formula and elliptic regularity theory to show that it admits a smooth, strictly positive density on the sphere. We further show that it is the only stationary measure with full support for the dynamics considered. We use the stochastic viewpoint to derive an algorithm for sampling from the minimizer. In dimension two, we also propose a numerical scheme for approximating its density based on a Mathieu series expansion.

[LG-231] Over-the-Air Semantic Alignment via Dual Stacked Intelligent Metasurfaces

链接: https://arxiv.org/abs/2611.19515
作者: Mario Edoardo Pandolfo,Kyriakos Stylianopoulos,George C. Alexandropoulos,Paolo Di Lorenzo
类目: ignal Processing (eess.SP); Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注:

点击查看摘要

Abstract:Semantic communications enable task-oriented information exchange, but suffer from latent-space mismatch when transmitter and receiver employ heterogeneous machine learning models. Existing semantic alignment methods rely on digital processing, introducing computational and latency overhead. This paper proposes a dual stacked intelligent metasurfaces (dual-SIM) architecture that performs semantic alignment entirely in the wave domain. By jointly exploiting programmable metasurfaces at both ends of a link and the wireless channel, the proposed architecture realizes an end-to-end linear transformation between heterogeneous latent spaces. We show phase-only multilayer SIM can approximate both supervised linear aligners and zero-shot Parseval-frame equalizers. We investigate single-SIM, disjoint asymmetric and symmetric dual-SIM, and jointly optimized dual-SIM configurations, considering channel equalization and constructive channel exploitation. For joint design, we develop an alternating projected gradient algorithm with convergence guarantees to first-order stationary points. We further introduce a deep-unrolled network that unrolls gradient-based iterations to reduce computational cost of SIM reconfiguration under channel variations. Numerical results with heterogeneous vision transformer encoders show accurate semantic alignment from moderate to high signal-to-noise ratios. The joint and symmetric disjoint dual-SIM designs achieve the highest accuracy over compact channels, the joint design leads at shallow stack depths, and the advantage over single-SIM designs diminishes at equal depth.

[LG-232] Interpreting Random Forests: Comparing Surrogate Models and Rule Extraction Methods

链接: https://arxiv.org/abs/2611.19471
作者: Lea L. Mairhöfer,Jonathan B. Krause,Björn-Hergen Laabs,Silke Szymczak
类目: Methodology (stat.ME); Machine Learning (cs.LG); Machine Learning (stat.ML)
*备注: 4 figures

点击查看摘要

Abstract:Random forests (RFs) often achieve high prediction performance, but their ensemble structure can be difficult to interpret. Surrogate models and rule extraction methods offer interpretable approximations, yet guidance on selecting an appropriate method for a specific research question remains limited. We compared 12 methods for interpreting RFs, with a decision tree as a prediction-oriented reference, across 11 simulation settings and 67 benchmark datasets for regression and classification. Evaluation addressed prediction and structural similarity to the RF, model complexity, prediction performance, and runtime. Simulations additionally assessed explainability using known effect and noise variables. No method performed best across all objectives. For regression, SLIM provided the strongest general performance for prediction similarity or performance on test data, while a linear model offered a simpler global interpretation of linear effects when performance was comparable. For classification, inTrees achieved the strongest overall benchmark performance, with CTree and OC-CART providing useful alternatives. ART consistently produced the most compact representations. NH and SIRUS offered strong structural similarity for regression, while MRT most closely reproduced RF variable usage in the classification simulations, illustrating that prediction-oriented rankings do not necessarily identify the best structural explanation. Outcome-type restrictions, runtime limits, and method-specific failures reduced result availability for several approaches and should be considered alongside performance when selecting a method. Method selection should therefore be guided by the outcome type, anticipated data structure, and intended interpretability objective. Prediction similarity alone is insufficient for explaining an RF and should be considered jointly with structural similarity and model complexity.

[LG-233] Generalization Bounds for Neural Networks with Sparse Connectivity NEURIPS2026

链接: https://arxiv.org/abs/2611.19382
作者: Nong Minh Hieu,Antoine Ledent
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注: Accepted at NeurIPS 2026, 61 pages, 2 figures

点击查看摘要

Abstract:Empirical evidence suggests that large neural networks rarely make effective use of all available parameters, with learned solutions exhibiting pronounced sparsity. Despite this ubiquity, existing generalization theory only partially explains how such sparsity influences statistical performance. In this paper, we derive non-asymptotic generalization gap bounds for deep neural networks whose layer-wise weight matrices are restricted by \ell_p quasi-norm ( 0p\leq 1 ) constraints. Our analysis gives rise to a novel result that interpolates between classical norm-based generalization guarantees and parameter-counting bounds. Remarkably, in the extreme sparsity limit as the quasi-norm order p approaches zero, our analysis reveals a transition to a square-root dependence on the number of \emphactive (non-zero) parameters, resulting in bounds that are strictly tighter than existing parameter-counting approaches. Experiments demonstrate that our bounds exhibit a milder dependence on initial architectural choice at the over-parametrization regime, effectively capturing the generalization benefit of the sparsity arising from training on natural data.

[LG-234] High-dimensional limit theorems for adaptive variants of SGD I: Ballistic limits

链接: https://arxiv.org/abs/2611.19059
作者: Aukosh Jagannath,Varnan Sarangian
类目: Machine Learning (stat.ML); Machine Learning (cs.LG); Probability (math.PR); Statistics Theory (math.ST)
*备注:

点击查看摘要

Abstract:We develop the high-dimensional effective dynamics for adaptive, momentum-based variants of Stochastic Gradient Descent (SGD). These effective dynamics are given by a Markov process. They key observation is that the non-Markovianity in the algorithm due to the exponential moving average and preconditioning can be treated using the method of averaging. Our approach applies to a broad class of algorithms including to Adaptive momentum estimation (ADAM), Momentum orthogonalized by Newton-Schultz (MUON), and its variants. It also applies to a broad class of learning tasks, such at tensor PCA, Gaussian projective models, teacher-student networks, and a multi-layer neural network trained on the XOR problem using a hybrid of MUON and ADAM as commonly done in practice. We demonstrate how to this technique can give precise, quantitative comparisons between algorithms. In particular, we find interesting phenomena, such as the cross-over phenomenon where online SGD has better generalization than ADAM early in training until the learning curves cross and ADAM dominates. Furthermore, we prove a version of the Edge of Stability phenomenon. In particular, we introduce an analogue of the infinity Laplacian that sharply characterizes the Edge of Stability.

[LG-235] Averag ed-Cost Control with Shared Feedback under Model Uncertainty

链接: https://arxiv.org/abs/2611.19019
作者: Piernicola Bettiol,Daisuke Inoue,Matías Gómez-Aedo,Dante Kalise,Richard B. Vinter
类目: Optimization and Control (math.OC); Machine Learning (cs.LG)
*备注: 26 pages, 6 figures

点击查看摘要

Abstract:This paper studies the shared-feedback design problem for finite-horizon averaged-cost control under model uncertainty represented by a joint law over the model parameter and initial state. One feedback map is shared across all realizations and uses only time and the current state as online inputs. We develop two methods: the nonlinear adjoint method combines forward closed-loop solves, backward adjoints, and joint-law policy gradients, while the linear moment method gives exact cost and gradient formulas for a homogeneous linear-quadratic design model by propagating conditional second moments. Numerical studies compare averaged and nominal designs for time-dependent neural feedback and static linear feedback in nonlinear Van der Pol and Stuart-Landau systems; the static gains are designed from oscillator linearizations and evaluated on the original nonlinear systems. Averaged design has lower cost in most comparisons.

[LG-236] aking Band-Limited Maximum Likelihood to High Dimensions: Linear Memory the Parametric Rate and a Self-Tuning Cut-off

链接: https://arxiv.org/abs/2611.19011
作者: Rahul Agarwal,Ankit Agarwal
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Band-limited maximum likelihood (BLML) estimates a density as the square of a band-limited function. The estimate is non-negative, and when the square root of the true density is band-limited its error falls at the parametric rate n^-1 . Its algorithm bins the data, which is fast only in one or two dimensions: from three on every point has its own bin, the kernel system is dense, and memory grows as n^2 , 40 GB at 10^5 points. Whether the rate survives in higher dimension, where kernel estimators slow down, had not been tested. We study dense data, densities that fill their space, not data near a low-dimensional manifold. We show that the binned equations are the stationarity condition of a strictly convex objective and minimise it by block coordinate descent without forming the kernel matrix: memory is linear in n , a sweep costs O(n^2 d) and a fit takes about ten, and one A100 GPU fits 10^6 points in eight dimensions in about ten minutes. On band-limited truths the error then decays as n^-1 in every dimension we tested, up to eight, while the slopes of kernel estimators flatten from -0.67 to -0.28, and it is 2 to 18 times lower than a neural spline flow’s. A streaming fit learns the cut-off, BLML’s only tuning parameter, from analytic gradients of the held-out likelihood, for about the cost of one more fit. On real data BLML is the strongest nonparametric estimator we tested: it leads kernel and nearest-neighbour estimators on electricity prices by 0.27 to 0.47 nats per hour, and a tuned kernel estimator on the full POWER and GAS benchmarks by 0.39 and 1.58 nats per point, placing it between early and recent normalizing flows. It loses on densities with a kink.

[LG-237] Accelerating Atomistic-Continuum Simulations of Multilayer 2D Materials with Functional Principal Component Analysis-Based Machine Learning

链接: https://arxiv.org/abs/2611.18961
作者: Jonathan Oleson,Abhishek Keripale,Shashank Pathrudkar,Susanta Ghosh
类目: Computational Physics (physics.comp-ph); Machine Learning (cs.LG); Mathematical Physics (math-ph)
*备注:

点击查看摘要

Abstract:Atomistic-continuum simulations of complex wrinkling deformations in large-area multilayer two-dimensional (2D) materials are computationally expensive. To accelerate these simulations, this work presents a machine-learning (ML) framework that first constructs a low-dimensional representation of the deformation field, then learns the deformation in this reduced space. Dimensionality reduction of deformed 2D membranes is challenging because of the extremely high dimensionality of discretized deformation fields, the need to preserve surface continuity and prescribed geometric constraints, the stochastic nature of wrinkling patterns, and the limited availability of training data. We address these challenges using two-dimensional functional principal component analysis (FPCA). The proposed FPCA representation reduces deformation fields containing millions of degrees of freedom to latent spaces with only tens of dimensions while preserving the continuity of the deformed surface. In addition, prescribed homogeneous boundary constraints are satisfied exactly through an appropriate choice of basis functions in the FPCA representation. A neural network is then trained to predict the principal components from the system and loading parameters. The resulting ML-predicted deformation fields are used as high-quality initial configurations for the atomistic-continuum solver, which subsequently performs energy minimization to obtain equilibrium solutions. The framework is demonstrated for multilayer graphene subjected to uniaxial compression, biaxial compression, and combined biaxial compression and simple shear. For the most computationally intensive large-strain cases, the proposed approach achieves speedups of at least 6.8 times while converging to final energies comparable to, and in some cases lower than, those obtained using conventional atomistic-continuum initialization.

[LG-238] Multi-Assay Antibody Developability Prediction with Shared Structure-Language Representations

链接: https://arxiv.org/abs/2611.18846
作者: Daria B. Kokh
类目: Biomolecules (q-bio.BM); Machine Learning (cs.LG)
*备注: 27 pages, 11 figures, 4 tables, Supplementary Material

点击查看摘要

Abstract:Early assessment of antibody developability is critical in drug discovery, but computational approaches remain limited by the scarcity of experimental data. We explore two strategies for working within that limit. First, we designed per-residue physicochemical descriptors encoding the local structural environment, including new terms such as directional charge asymmetry and screened electrostatic potential. These feed a convolutional network, combined with protein language model (PLM) embeddings in a two-stream architecture: the two are encoded and predicted independently and merged only at the prediction stage. Second, a shared trunk with per-assay heads passes information between related assays. Both are implemented in the Multi-Assay Developability Model (MADM) and evaluated on eight assays: HIC, heparin affinity (HAC), AC-SINS, and polyspecificity ELISA against BVP, dsDNA, insulin, ovalbumin and soluble membrane proteins (SMP), with labels from 29 open sources. Structural descriptors match or exceed PLM embeddings (AbLang2, ESM2) on several assays; combining the streams gives a small but consistent gain in accuracy and reduces seed-to-seed variance. Multi-assay training raises Spearman rho for ovalbumin and SMP by about 0.1, where assays correlate and share sequence space. Under cluster-split cross-validation the strongest results are for HAC, ovalbumin, insulin, SMP and HIC (rho = 0.70-0.82); AC-SINS, BVP and dsDNA are weaker (0.50-0.55). In these three, sequence clusters largely coincide with data source, so a held-out fold measures transfer to differently calibrated measurements as well as to new sequences. On the 80-antibody GDPa3 hold-out, performance falls by rho around 0.1 for all but SMP (0.22). Transfer to the VHH panel of Ginkgo GDPa5 gives comparable results, and scFv validation against binary SMP labels yields F1 = 0.89, AUC = 0.95.

[LG-239] rajectory-Based Bayesian Discovery of a Host-Pathogen Model from Quasi-Steady-State Observations

链接: https://arxiv.org/abs/2611.18815
作者: Reymart Salcedo Lagunero,Andreas Langer,Cordula Reisch
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:System identification of nonlinear dynamical systems becomes challenging when the measured data come from a nearly settled state. We study this situation in the case of a bistable host-pathogen interaction model with a cubic term, whose seven true mechanisms are hidden among twenty candidate terms. The data are from one noise-free trajectory on a quasi-steady-state observed window and the initial state. In this situation, classical sparse identification approaches fail to identify the correct mechanisms as the candidate library becomes ill-conditioned along the quasi-steady-state trajectory. To address this, we propose an alternative approach in which sparse Bayesian learning and \ell_0 iterative hard thresholding with pursuit provide a good start, and a two-swap climb on the Bayesian evidence of the candidate ODE solutions recovers the true model. Our results show successful model recovery, i.e., correct identification of its support and accurate estimation of its coefficients, as validated by exhaustive scoring.

[LG-240] Hybrid quantum learning for imbalanced blood-cell classification: domain adaptation and classifier design

链接: https://arxiv.org/abs/2611.18757
作者: Gabriele D’Angeli,Cyrill Aebischer,Eryk Schiller,Gianluca Nicchiotti,Su Fong Chien,Jérôme Extermann,Enrico Pomarico
类目: Image and Video Processing (eess.IV); Machine Learning (cs.LG)
*备注: Submitted to Quantum Machine Intelligence

点击查看摘要

Abstract:Limited annotated data, class imbalance, and morphological similarity make white blood cell classification challenging. We benchmark hybrid variational quantum classifier heads against an affine baseline on 821 microscopy images across seven morphological classes. Using ImageNet-initialised ResNet34 backbones and repeated stratified five-fold cross-validation, we first compare direct transfer with intermediate adaptation on MLL23, a larger peripheral blood cell microscopy dataset. To guide the head comparison, we explore different encoding strategies, variational depth, and the number of independent circuit blocks. We then compare the classifier performance using the affine head and two hybrid heads across multiple random seeds, under a shared training and evaluation protocol. For this comparison, all backbones share the same MLL23-adapted initialisation. Intermediate adaptation increases mean macro-F1 across all four affine-head fine-tuning configurations. The largest gain occurs with a frozen backbone, where macro-F1 rises from 0.685 to 0.786 , a relative improvement of approximately 14.7% . The shallow two-block hybrid head achieves a mean pooled out-of-fold macro-F1 of 0.865 \pm 0.010 , compared with 0.879 \pm 0.008 for the affine head. The corresponding balanced accuracies are 0.886 \pm 0.008 and 0.889 \pm 0.010 , respectively. In the architectural exploration, the two-block hybrid head achieves the highest macro-F1 and balanced accuracy among the evaluated hybrid configurations, while the effect of encoding varies across the considered metrics. Our results help guide the design and evaluation of hybrid classifier heads for imbalanced microscopy by linking architectural choices to overall performance and class-specific precision-recall trade-offs.

[LG-241] Random matrix discrepancy – CLuP algorithm and parametric RDT approach

链接: https://arxiv.org/abs/2611.18650
作者: Mihailo Stojnic
类目: Disordered Systems and Neural Networks (cond-mat.dis-nn); Information Theory (cs.IT); Machine Learning (cs.LG); Probability (math.PR)
*备注:

点击查看摘要

Abstract:We study information-theoretic and algorithmic aspects of random matrix discrepancy in proportional regime. The problem amounts to finding \kappa=\min_\x\in\mB^n|\sum_i=1^n\barA_i\x_i|_2 , where \barA_i\in\mR^d\times d is a collection of n iid GOE matrices, \mB^n are vertices of the unit cube in \mR^n , and \alpha=\lim_n\rightarrow \infty\fracnd^2 is a proportionality constant. Strong progress is made in several directions: (i) We develop a Random duality theory (RDT) framework and utilize its a lifted variant to establish lower bounds on \kappa , which precisely match the bounds obtained via the first moment in [81]. (ii) To address analytical challenges, we propose a negative Hopfield model as a proxy for implementation of a fully lifted RDT. On the 2nd lifting level, this model fits remarkably well between the information-theoretic lower bounds from [81] and the algorithmic upper bounds from [38]. (iii) On the 3rd and 4th lifting levels, parametric RDT concepts predict a sizeable computational gap for lower \kappa values in the proxy model. Extending proxy information-theoretic fitting to algorithmic suggests a similar computational gap for matrix discrepancy in low \kappa regime, aligning with [73]. (iv) We design a Controlled Loosening Up (CLuP) algorithm for matrix discrepancy. Its performance strongly matches our analytical results – closely approaching information-theoretic predictions in lower \alpha ranges (indicating a small or nonexistent computational gap) and following the proxy model’s algorithmic predictions in higher \alpha ranges (indicating a sizeable computational gap). Both the information-theoretic and algorithmic predictions match the corresponding phenomenologies of the symmetric binary perceptron (SBP) vector analogue in the entire \alpha\in(0,\infty) range [1,2,14,31,51,92,111]. Subjects: Disordered Systems and Neural Networks (cond-mat.dis-nn); Information Theory (cs.IT); Machine Learning (cs.LG); Probability (math.PR) Cite as: arXiv:2611.18650 [cond-mat.dis-nn] (or arXiv:2611.18650v1 [cond-mat.dis-nn] for this version) https://doi.org/10.48550/arXiv.2611.18650 Focus to learn more arXiv-issued DOI via DataCite (pending registration)

[LG-242] Domino: Joint Unmasking via First-order Markov Chain for Few-Step Masked Diffusion

链接: https://arxiv.org/abs/2611.18645
作者: Mansoor Ahmed,Lilia Chebbah,Jenny Dinh,Murray Patterson,Hemanth Venkateswara
类目: Machine Learning (stat.ML); Machine Learning (cs.LG)
*备注:

点击查看摘要

Abstract:Masked diffusion models generate a sequence by revealing masked positions over many denoising steps. A few-step sampler has to reveal many positions at once, but the standard reverse kernel draws them independently and primarily ignores the conditional dependence between tokens revealed in the same step. We propose a novel framework called Domino, which models this dependence with a first-order Markov chain over the revealed positions, given the current state. The first revealed position keeps the posterior of a pre-trained masked diffusion model (MDM), and each subsequent position follows a transition given the token of its predecessor. Notably, the resulting step is normalized by construction, contains the factorized step in MDMs as a special case, and has a likelihood in closed form. We implement the Domino transitions via a lightweight module on the hidden states of a pre-trained MDM denoiser with only 2.6% parameter overhead and train by keeping the MDM denoiser frozen. Our extensive evaluations across text, molecules, and images show that Domino lowers the few-step generative perplexity of MDMs by a factor of 1.7 to 3.8 on One Billion Word (LM1B), WikiText-103, and OpenWebText at nearly unchanged sample entropy. It also raises the validity of few-step molecule samples by 1.5x. The source code is available at: this https URL

[LG-243] Ensemble Bounds for Pairwise Dissimilar Decomposition of Deterministic Datasets

链接: https://arxiv.org/abs/2611.18625
作者: Ghurumuruhan Ganesan
类目: Machine Learning (stat.ML); Information Theory (cs.IT); Machine Learning (cs.LG)
*备注: Accepted for publication in Springer Proceedings in Mathematics and Statistics, QEDS conference, Ranchi, India

点击查看摘要

Abstract:In this paper, we study reduced similarity decompositions of \emphdeterministic datasets using redundancy graphs and obtain \emphensemble bounds for the minimum size of such constrained decomposition. We consider batch decompositions that reduce similarity \emphacross batches and use the Lovász local lemma to obtain bounds for the minimum size of a decomposition with a given pairwise dissimilarity. We then consider unique batch decomposition with a given relaxation parameter and obtain upper and lower ensemble bounds for the minimum size. Throughout, we use Binomial random graphs to obtain the desired ensemble bounds, by conditioning on the number of similar \emphpairs of data points.

[LG-244] Physics-informed and event-aware forecasting of ammonia emissions in aerobic chicken manure composting bioreactors: A PINN-PatchTST-GPR framework with symbolic dynamic extraction

链接: https://arxiv.org/abs/2611.18614
作者: Lanbin Wei,Haiyong Weng
类目: Quantitative Methods (q-bio.QM); Computational Engineering, Finance, and Science (cs.CE); Machine Learning (cs.LG)
*备注: 46 pages, 6 figures

点击查看摘要

Abstract:Predicting ammonia (NH3) concentrations during aerobic chicken manure composting provides critical feedforward signals for intelligent aeration. However, the concurrence of long-term biochemical evolution and intermittent forced-aeration pulses poses significant modeling hurdles across steady baselines, transient surges, and risk quantification. Existing data-driven models lack physical consistency, while mechanistic kinetic models struggle with environmental disturbances. To bridge this gap, we propose PINN-PatchTST-GPR, a physics-informed neuro-symbolic forecasting framework. Validated on a 28-day composting campaign across six vertical bioreactors, the framework couples a channel-independent PatchTST backbone with macroscopic kinetic constraints, dynamic physical bounds, and a Gaussian process regression (GPR) event specialist. Multi-channel sensor streams drive discrete differential equation residuals and non-negativity bounds, strictly ensuring physical plausibility. A block-split conformal prediction module yields nominal 90% distribution-free prediction intervals, while PySINDy and PySR extract low-complexity symbolic equations for quiescent and surge regimes. The model achieves an overall RMSE of 0.8256 +/- 0.0132 ppm, R2 of 0.9817 +/- 0.0006, Epeak of 1.6793 +/- 0.0725 ppm, and Etime of 0.7 +/- 0.2 min. For uncertainty quantification, it attains PICP90 of 99.792 +/- 0.047%, MPIW of 6.951 +/- 0.128 ppm, and IS0.1 of 6.968 +/- 0.128 ppm. Symbolically, PySINDy yields an RMSE of 0.4624 +/- 0.0758 ppm (R2 = 0.9896), while PySR yields 2.4084 +/- 0.0864 ppm (R2 = 0.9523). Biologically, the 20-min aeration strategy minimizes cumulative NH3 emissions. Overall, PINN-PatchTST-GPR harmonizes global forecasting, localized transient responsiveness, and risk quantification for precision composting.

[LG-245] No Convergence in Learned Eavesdropping-Detection Dynamics for BB84 Quantum Key Distribution

链接: https://arxiv.org/abs/2611.16372
作者: Ramon Jose C. Bagunu
类目: Quantum Physics (quant-ph); Cryptography and Security (cs.CR); Machine Learning (cs.LG)
*备注: 5 figures

点击查看摘要

Abstract:In prior work we showed that a reinforcement-learning eavesdropper can intercept BB84 quantum key distribution while holding the quantum bit error rate below the abort threshold and keeping a usable distillable key. In this paper we ask whether an equally adaptive learning detector can close this stealth window. We cast the interaction as an alternating best-response game between a tabular eavesdropper and a tabular threshold-adapting detector. We measure exploitability, which is the detection rate and distillable key fraction achieved by a fresh attacker best-responding to the final detector. Across three learning rules (Q-Learning, SARSA, Double Q-Learning), three channel-noise levels, and a uniform false-alarm discipline, the dynamics do not settle out to twenty rounds. The exploitability does not approach a fixed point, and its round-to-round swings do not decay. Two controls, one retraining the attacker against a frozen defender and one retraining the defender against a frozen attacker, show that the swings on each side match the variance of the best-response procedure itself. The dynamics fluctuate at the scale of best-response variance rather than converging, with neither player settling to a fixed response. The behavior is noise-dependent, and the stealth window stays open at the low noise where practical QKD operates. The defender contains the attacker on average at intermediate noise without the contest settling, and at high noise the channel rather than the learned policy suppresses the attack. As a candidate mechanism we examine a last-mover information asymmetry, supported by significant anti-persistence at intermediate noise and consistent with best-response noise at low noise. The results suggest that a learned detector cannot be certified from any single round, and that closing the window calls for breaking the information structure rather than out-adapting the attacker.

附件下载

点击下载今日全部论文列表