ArXiv Domain 2026-09-10

数据来源：ArXiv Domain

LLM Domain Papers

1. When Agent Governance Helps

Abstract:No specification says how a governed autotelic AI agent organization, where agents pursue self-generated goals inside guardrails, should be designed and evaluated. We answer in two parts. First, we synthesize the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework from a document-based qualitative evidence synthesis of 321 sources, integrating agency, agile, platform, and governance theory into a runnable specification. Second, we probe a prompt-layer instantiation of GAMPO on CHI-Bench, a long-horizon healthcare benchmark, across open and frontier models. The result is a boundary condition: governance benefit is gated by a model’s spare capacity and is domain- and model-specific. On capacity-constrained open models the full procedure yields no reliable benefit, whereas a single “verify your writes” sentence doubles task success (pass@1 2/20 to 4/20). At the frontier the same scaffold lifts prior-authorization 24% to 40% but nets zero on another model, a gap traced to a stable recommendation-override disposition. A second result refines the first: replacing the generic procedure with an answer-blind, per-task definition-of-done, keyed only to the case’s own policy and published standards, never the hidden key, raises prior-authorization to 84% under best-of-five self-consistency (68% single-attempt, confirmed by a held-out board) and utilization-management to 44%, while care-management meets a subjective content-quality wall. The contribution is a named, auditable framework and capability-gated evidence that governance should be sized to spare capacity, and that at the frontier a case-grounded specification beats a uniform procedure. Findings are exploratory: partial instantiation, small per-cell samples (n = 5-25), and single trials.

中文摘要

摘要：目前没有规范说明如何设计和评估一个受治理的自目标人工智能代理组织，在该组织中，代理在护栏内追求自主生成的目标。我们从两个方面进行回答。首先，我们从321个文献来源的文档基础定性证据综合中综合出受治理的自目标多代理产品组织（GAMPO）框架，将代理理论、敏捷、平台和治理理论整合为可运行规范。其次，我们在CHI-Bench（一项长期医疗保健基准测试）上探讨GAMPO的提示层实例化，在开放模型和前沿模型中进行评估。结果显示一个边界条件：治理效益受模型剩余能力的限制，并且具有领域和模型特异性。在能力受限的开放模型上，全流程没有带来可靠效益，而仅添加一句“验证你的书写”即可使任务成功率翻倍（pass@1从2/20升至4/20）。在前沿模型上，相同的框架将事前授权从24%提升至40%，但在另一模型上则为零，差异源于稳定的推荐覆盖倾向。第二个结果完善了第一个：将通用流程替换为答案盲、每任务定义完成标准，仅依据病例自身政策和已发布标准，而不使用隐藏钥匙，可在最佳五次自一致性下将事前授权提升至84%（单次尝试68%，由预留委员会确认），将利用管理提升至44%，而护理管理达到主观内容质量壁垒。贡献在于提供一个命名且可审计的框架，以及按能力限制的证据表明治理应根据剩余能力进行规模化，同时在前沿模型中，基于个案的规范胜过统一流程。研究结果属于探索性：部分实例化、每单元样本量小（n=5-25）、单次试验。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05531 (HTTP 429)

Authors: Michael Ray Johnson, Linda Naimi

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05531.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05531

Published: 2026-09-10T01:26:51.917Z

2. TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

Abstract:A voice agent has to decide, at every pause, whether the user has finished speaking. Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait. Open semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language. We release TamilEOT: 18,485 labelled turn boundaries cut from 116 real Tamil telephone conversations, and two audio-only detectors fine-tuned from Smart Turn v3. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both models run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Rule-derived labels, checked against a blind human listening pass, were right 95.9% of the time on the positive class and 44.4% on the negative class, which is below chance, because the rule answered a different question than the model is asked. Replacing them with an audio-LLM labeller measured at 97.5% human agreement cost US$5.69. Of every training lever we measured, only encoder capacity moved the result; three runs at identical config and seed span 0.87 accuracy points, which is the floor below which none of our other deltas mean anything. Replaying the same labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries are never surfaced to the model at all. Data, weights, code and every negative result are public.

中文摘要

摘要：语音代理必须在每次停顿时决定用户是否已经说完。如果没有语言模型，这个决定就会退回到固定的静音超时：设置得短，代理会打断；设置得长，每一次轮次都要等待完整的时间。已有开放语义的轮次结束检测器，但据我们所知，没有覆盖南印度语言。我们发布了 TamilEOT：从 116 个真实泰米尔语电话对话中切出的 18,485 个标注轮次边界，以及两个从 Smart Turn v3 微调的纯音频检测器。在 30 个未见过的通话中划分出的 4,168 个剪辑上，准确率从零样本的 70.30% 提升到 83.71%（8.7 MB）和 86.13%（21 MB）；ROC-AUC 从 0.751 提升到 0.921。两个模型在笔记本 CPU 单线程运行均低于 150 毫秒。我们还报告了构建该系统的成本。基于规则生成的标签，经盲检人工听审验证，正类正确率为 95.9%，负类为 44.4%，低于随机水平，因为规则回答的问题与模型所问不同。用音频-LLM 标注器替换它们，在测得 97.5% 人类一致率的情况下花费为 5.69 美元。在我们测量的每个训练杠杆中，只有编码器容量影响了结果；在相同配置和随机种子下的三次运行准确率跨度为 0.87 个百分点，这是我们其他变化无意义的下限。通过生产环境 VAD 和流适配器重新播放相同的标注边界会额外消耗 2.60 个百分点，并且有 7.8% 的边界根本不会传递给模型。数据、权重、代码以及所有负面结果均已公开。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05631 (timeout of 15000ms exceeded)

Authors: Santhoshkumar V

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05631.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05631

Published: 2026-09-10T01:26:51.917Z

3. Better Together: Complementary Query Rewriting Under a Strong RAG Baseline

Abstract:A popular way to improve Retrieval-Augmented Generation (RAG) is to rewrite the user’s question into several variants and search with all of them. We test whether this actually helps once the underlying search is already strong. Under one fixed, competitive pipeline (BGE dense retrieval, cross-encoder reranking, and MMR diversification), we compare four query-rewriting strategies (S1-S4) against two strong LLM baselines (HyDE, Query2Doc) on three datasets (HotpotQA, AmbigNQ, and the 512K-document EnterpriseRAG-Bench) over three seeds with paired-bootstrap significance tests. Our headline result is that rewriting alone is at best competitive with a strong baseline, but combining methods yields outsized gains because different strategies fail on different questions. A post-hoc union of four methods (S1+S3+S4+HyDE) improves HIT@10 over the baseline by +12.5 points on enterprise data (51.70 vs 39.22), and a five-method union reaches 52.98 (+13.8). Budget-matched controls capture only ~40% of this gain, confirming that complementarity, not retrieval budget, is the primary driver. On HotpotQA the union adds +1.6 to +1.8 points (p<0.001), saturating the all-method oracle; on AmbigNQ the same fusion hurts (-2.4 below the best solo, p<0.001), and we analyze when and why. Because rewriting is expensive, we evaluate in simulation a confidence-gated router that runs rewriting only when the baseline’s own top-1 score is low. It captures about half of the enterprise full-merge gain (+4.3 HIT@10) while paying rewriting cost on <40% of queries, and automatically declines to rewrite on AmbigNQ. A downstream answer-quality evaluation confirms the router improves F1 by +1.92 (p<0.01) at roughly 40% of the expansion cost. In short: treat query rewriting as a complementary coverage source, applied through cost-aware routing, not as a standalone replacement for a strong baseline.

中文摘要

摘要：一种改进检索增强生成（RAG）的流行方法是将用户的问题重写为多个变体，并使用所有变体进行搜索。我们测试了在底层搜索已经很强的情况下，这种方法是否真的有帮助。在一个固定且具有竞争力的流水线（BGE 稠密检索、交叉编码重排序和 MMR 多样化）下，我们将四种查询重写策略（S1-S4）与两种强大的大型语言模型基线（HyDE，Query2Doc）在三个数据集（HotpotQA、AmbigNQ 和 512K 文档的 EnterpriseRAG-Bench）上进行比较，使用三组随机种子并通过配对自助法显著性检验进行评估。我们的主要结果是，单独的重写最多只能与强基线竞争，但方法组合会带来显著增益，因为不同策略在不同问题上失败。四种方法的事后联合（S1+S3+S4+HyDE）在企业数据上的 HIT@10 比基线提高了 +12.5 个点（51.70 对 39.22），五种方法的联合则达到 52.98（+13.8）。预算匹配的控制仅捕获了大约 40% 的增益，确认互补性，而非检索预算，是主要驱动力。在 HotpotQA 上，联合增加了 +1.6 到 +1.8 个点（p<0.001），达到了全方法预知的饱和值；在 AmbigNQ 上，相同的融合反而有害（低于最佳单一方法 -2.4，p<0.001），我们分析了其原因和发生的情况。由于重写代价高昂，我们在模拟中评估了一个置信度门控的路由器，仅当基线自身的 top-1 分数较低时才进行重写。它捕获了企业全合并增益的大约一半（+4.3 HIT@10），而重写成本仅发生在小于 40% 的查询上，并且自动在 AmbigNQ 上拒绝重写。下游答案质量评估确认该路由器将 F1 提升了 +1.92（p<0.01），成本扩展大约为 40%。简而言之：将查询重写视为互补的覆盖来源，通过成本感知的路由应用，而不是作为单独替代强基线的方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05637 (HTTP 429)

Authors: Sara Shanian, Xiaoqin Yi, Pavlo Ruban, Kurt MacDonald

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05637.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05637

Published: 2026-09-10T01:26:51.917Z

4. Who Maintains Agent Skills? A Longitudinal Study of Human-Governed, AI-Assisted Skill Maintenance

Abstract:Lifelong LLM agents increasingly rely on external skill artifacts as one element for preserving and reusing capabilities over time. These skills (usually portable Markdown files such as this http URL) describe when and how to apply a capability and must be corrected, expanded, and consolidated as tools and usage patterns shift over deployment. Recent work seeks to automate skill curation, but it largely evaluates against automated baselines and treats human maintenance as an unmeasured bottleneck. We study that missing process directly. We mine the full commit histories of five public AI-skill repositories, a purposive sample of AI-tooling organizations, covering 873 commits, 143 skill files, and 254 substantive post-creation edits from October 2025 to June 2026. We code each edit with pre-registered governance, operation, and trigger-evidence codebooks. Three findings emerge. First, every substantive edit is authored or merged through a named human account, while 62% carry an AI co-author trailer, with large repository-level variation. Second, these edits are genuine curation: an audited sample shows that most change skill content, and the coded operations are dominated by additions and corrections. Third, a pre-registered rule-likeness axis fails its reliability gate; reliably coding rule-likeness from commit artifacts remains an open measurement problem. We release the corpus, codebooks, mining scripts, and a replay protocol for automated skill curators. For self-evolving agents, public skill maintenance currently looks less like an autonomous pipeline than a human-governed, AI-assisted loop that future curators must measure against and operate within.

中文摘要

摘要：终身学习的LLM（大语言模型）代理越来越依赖外部技能工件，作为维持和重复使用能力的一个元素。这些技能（通常是可移植的Markdown文件，例如此HTTP URL所示）描述了何时以及如何应用某项能力，并且必须随着工具和使用模式在部署过程中发生变化而进行纠正、扩展和整合。近期的研究试图自动化技能策划，但主要是以自动化基线进行评估，并将人工维护视为未被测量的瓶颈。我们直接研究了这一缺失的过程。我们挖掘了五个公共AI技能仓库的完整提交历史，这些仓库是人工挑选的AI工具组织样本，涵盖873次提交、143个技能文件以及2025年10月至2026年6月期间254次实质性创建后编辑。我们使用预注册的治理、操作和触发证据代码表对每次编辑进行编码。三个发现浮现出来。首先，每一条实质性编辑都是由一个命名的人工账户撰写或合并，同时62%的编辑带有AI共同作者标记，但在不同仓库层面存在较大差异。其次，这些编辑是真正的技能策划：经过审计的样本显示，大多数编辑会改变技能内容，并且编码操作主要以添加和修正为主。第三，预注册的规则相似性轴未通过其可靠性测试；从提交工件中可靠地编码规则相似性仍是一个开放的测量问题。我们发布了语料库、代码表、挖掘脚本和自动化技能策划的重放协议。对于自我进化的代理，目前公共技能维护看起来更像是一个人工治理、AI辅助的循环，而非自主管道，未来的策划者必须针对这一点进行测量并在其内部操作。

LLM Analysis

Q: 这篇论文试图解决什么问题？这篇论文试图解决\*\*自动化技能策展研究中对人工维护过程缺乏系统实证测量\*\*的问题。具体而言，现有关于终身LLM智能体技能库的研究主要关注如何自动化地生成、编辑和演化技能文件（如SKILL.md），并在自动化基线上评估这些方法。然而，这些研究将人工维护视为一个未被测量的可扩展性瓶颈——即认为其"需要大量专业知识且无法扩展"——却未曾直接考察公共仓库中真实的人类维护行为究竟包含什么、由谁执行、以及遵循何种模式。论文提出一个前提性的实证问题：\*\*在公共可观测的AI技能仓库中，是谁在维护这些智能体技能，以及这种维护具体由哪些操作构成？\*\* 为回答该问题，作者对五个公共AI技能仓库的完整Git提交历史进行了纵向挖掘，系统编码了254次实质性维护编辑的治理结构（人类主导 vs. AI协助）、策展操作类型（增强性 vs. 纠正性）以及触发证据，从而首次对"人类治理、AI协助"的技能维护循环进行了定量描述。 Q2: 有哪些相关研究？该论文的相关研究可分为四个主要类别，其核心差异在于：既有工作要么自动化地处理技能生成却未测量人工维护过程，要么挖掘演化的文本产物但未聚焦可执行的SKILL.md技能文件，也未同时考察治理结构与策展操作。 ### 1. 自动化技能策展（Automated Skill Curation） - \*\*早期终身智能体技能库\*\*：Wang et al. (2024) 提出可不断扩展的可执行技能库；Shinn et al. (2023) 与 Zhao et al. (2024) 通过反思机制将经验提炼为可复用的洞察。 - \*\*近期自动编辑与演化系统\*\*：Yang et al. (2026) 利用文本空间优化器从经验中编辑技能文本；Ni et al. (2026) 通过轨迹蒸馏生成技能；Ouyang et al. (2026) 使用强化学习训练外部仓库的策展者；Lin et al. (2026)、Shen et al. (2026)、Yu et al. (2026b) 则构建了完全自主的自演化技能库系统。 - \*\*基准测试\*\*：Zhong et al. (2026)、Li et al. (2026a)、Han et al. (2026) 评估生成的技能对下游任务的帮助程度，其中 Zhong et al. (2026) 报告人工撰写技能的任务成功率显著高于自动化方法。 \*\*与该研究的关系\*\*：上述工作均将人工维护视为可扩展性瓶颈而未直接测量其过程，且 Ouyang et al. (2026) 未设置人工策展对照组。 ### 2. 演化文本产物的经验挖掘（Empirical Mining of Evolving Artifacts）该方法论领域与本文最接近，即对开源仓库中持续演化的文本产物进行类型化挖掘： - \*\*智能体上下文文件\*\*：Chatlatanagulchai et al. (2025) 发现智能体 README/AGENTS.md 类文件"像配置代码一样通过频繁、小规模的增补演化"；Mohsenimofidi et al. (2026) 研究了面向 AI 智能体的上下文工程。 - \*\*提示与规则文件\*\*：Tafreshipour et al. (2025) 挖掘仓库中的提示演化，发现仅 21.9% 的提示变化能在提交信息中找到痕迹；Cai et al. (2026) 研究了 AI IDE 规则文件的演化；Galster et al. (2026) 构建了智能体 AI 编码工具配置的数据集。 - \*\*工作流演化\*\*：Rostami Mazrae et al. (2026) 发现 CI 工作流编辑以修改为主，而非增补。 - \*\*最接近的并行注册报告\*\*：Voria et al. (2026) 计划挖掘智能体上下文文件（CLAUDE.md/AGENTS.md）并推导基于维护的变化分类法，但\*\*排除了技能文件\*\*，且\*\*不归属每次编辑的作者身份\*\*。 \*\*与该研究的关系\*\*：这些研究在方法论上相似，但均未以可执行的 SKILL.md 技能文件为对象，也未同时测量提交级别的作者治理与编码的策展操作。 ### 3. 结构化技能与经验记忆（Structured Skill/Experience Memory） - \*\*图结构记忆\*\*：Li et al. (2026b) 的 SkillGraph 将键入技能图与策略共同训练；Feng et al. (2026) 的 ExpGraph 使用图结构经验记忆。 - \*\*反思与运行时纠正\*\*：Hassell et al. (2026) 利用反思性语义/情景记忆；Zhou et al. (2026) 的 TRACE 将用户纠正编译为运行时执行钩子。 \*\*与该研究的关系\*\*：这些系统操作于内部存储或运行时产物，均未挖掘纵向技能文件维护历史，也未将通用规则写回外部 Markdown 技能文件。 ### 4. 其他相关研究 - \*\*横截面市场分析\*\*：Ling et al. (2026) 对 Claude 技能市场进行了数据驱动的横截面分析，缺乏纵向视角。 - \*\*合成历史构造\*\*：Li & Hu (2026) 的 SkillHone 利用持久决策历史进行持续技能演化，但其历史为\*\*合成生成\*\*，而非真实的公开提交记录。 ### 总结定位如下表逻辑所示，既有工作要么是自动化系统（缺乏人工维护测量），要么是端点基准（缺乏过程分析），要么是其他产物（非 SKILL.md 文件）。该研究首次对公开 Markdown 技能文件的纵向维护历史进行了挖掘，同时编码了策展操作与作者治理结构。 | 研究方向 | 纵向 | 研究对象 | 人工策展分析 | |---|---|---|---| | SkillOpt / SkillOS / Trace2Skill | ✓ | 技能文件 | 仅自动化 | | SkillLearnBench / SWE-Skills-Bench | – | 技能文件 | 仅端点 | | Agent READMEs (Chatlatanagulchai et al., 2025) | ✓ | 上下文文件 | 粗略、未编码 | | Claude-skills 市场 (Ling et al., 2026) | – | 技能文件 | – | | SkillHone (Li & Hu, 2026) | ✓ | 技能文件 | 合成历史 | | \*\*该研究\*\* | \*\*✓\*\* | \*\*SKILL.md 编辑历史\*\* | \*\*编码 + 测量\*\* | Q3: 论文如何解决这个问题？论文通过\*\*纵向挖掘公共 Git 仓库的完整提交历史并建立系统化的三维编码框架\*\*，直接测量此前未被刻画的人工维护过程。具体方法路径如下： ### 1. 构建 purposive 语料库从五个公共 AI 技能仓库（getsentry/skills、trailofbits/skills、obra/superpowers、anthropics/skills、cloudflare/skills）的固定提交快照中，追踪 143 个 \`SKILL.md\` 文件的完整历史（含重名解析），覆盖 873 个提交。通过预筛选定义\*\*实质性创建后编辑\*\*：文件创建后、至少修改 5 行、非机器人作者且非琐碎变更，最终得到 254 个分析单元。 ### 2. 建立预注册的三轴编码手册每个实质性编辑沿三个正交轴进行编码： - \*\*治理轴（Governance）\*\*：基于提交元数据\*\*确定性\*\*分类——是否携带 AI 共同作者标记（\`Co-Authored-By\` trailer）、人类作者/提交者身份、或网页合并（web-merge）身份。 - \*\*策展操作轴（Curation Operation）\*\*：预注册的八类维护操作（内容扩展、事实纠正、失败修复、重构、合并、弃用等），可折叠为经典的\*\*纠正性\*\*（corrective）与\*\*增强性\*\*（enhancement / perfective）二分。 - \*\*触发证据轴（Trigger Evidence）\*\*：0–3 级阶梯，量化提交记录中可观测到的失败触发证据强度（L3 显式问题链接、L2 命名具体失败、L1 通用、L0 无）。 ### 3. LLM 辅助编码与可靠性控制 - 使用 Opus 4.8 对全部 254 个编辑进行双通道编码，并通过 50 编辑的重叠分层样本估计稳定性。 - \*\*关键控制\*\*：八类操作的模型间一致性仅为 kappa = 0.62 ，但将其折叠为\*\*纠正 / 增强\*\*二分后，一致性提升至 kappa = 0.64 （原始一致性 0.82），进一步经盲法人类重编码验证（ kappa = 0.72 ）。论文据此将 headline 发现严格建立在该二分分类上，而将八类叶节点仅作描述性报告。 - 对未能通过可靠性门控的预注册轴（“规则相似性”， kappa = -0.02 ），按预注册承诺\*\*不报告其分布\*\*，仅将其作为开放测量问题披露。 ### 4. 多维度分析维护过程在编码基础上展开四项互补分析： - \*\*维护内容\*\*：统计操作类型分布，计算纠正性与增强性编辑的比例。 - \*\*演化动态\*\*：追踪 120 个技能文件的规模轨迹（resident token 长度净变化）与编辑节奏（中位间隔 5 天），并通过确定性解析器将每次编辑的变更行归因至六个组件（指令、代码、路由器、前言、示例、引用）。 - \*\*治理结构\*\*：量化 AI 共同作者标记的占比（62%）、人类账户的绝对控制权（100% 实质性编辑均经人类作者或合并），并揭示组织层面的双峰差异。 - \*\*维护效用边界\*\*：通过预注册的迁移任务实验（13 个技能 × 11 个任务，共 143 个任务三元组）检验维护版本是否显著优于早期版本。该设计控制了求解器族属（gpt-5.4-mini）、多轮交互协议、盲法跨族评判、以及统计功效（预注册功效 0.77–0.90）。 ### 5. 发布可复现资源与重放协议论文发布完整的提交级记录（631 条文件历史记录）、三份编码手册、挖掘与编码脚本，以及一个\*\*重放协议\*\*（replay protocol）。该协议允许未来的自动化策展系统以历史编辑为靶标，在相同上下文中回溯评估其提案与真实人类编辑在操作类型、组件位置及内容重叠上的匹配程度。 Q4: 论文做了哪些实验？论文围绕公共 \`SKILL.md\` 文件的维护历史开展了多组相互补充的实证分析与受控实验，可归纳为以下五个层面： --- ### 1. 语料库挖掘与单元定义实验 - \*\*数据来源\*\*： purposive 选取 5 个公开 AI 技能仓库（getsentry/skills、trailofbits/skills、obra/superpowers、anthropics/skills、cloudflare/skills），固定提交快照后解析完整 \`git log --follow\` 历史。 - \*\*产出规模\*\*：共覆盖 873 个提交、143 个 \`SKILL.md\` 文件、631 条 commit-file 历史记录。 - \*\*实质性编辑筛选\*\*：定义“创建后、修改不少于 5 行、非机器人作者且非琐碎”的编辑为标准分析单元，得到 \*\*254 个实质性编辑\*\*。 - \*\*审计校验\*\*：对 50 编辑的审计样本进行人工分类，其中 70% 确认为真正的技能内容变更（其余为迁移、依赖版本或纯风格变更），排除了机械性噪音主导的可能。 --- ### 2. 三轴编码与可靠性验证实验对全部 254 个实质性编辑实施预注册的三轴编码，并系统检验信度： | 编码轴 | 方法 | 信度结果 | |---|---|---| | \*\*治理轴\*\* | 基于提交元数据确定性分类：AI co-author trailer、人类 author/committer、网页合并身份 | 模型双通道一致率 99.2%（252/254），无需推断性编码 | | \*\*策展操作轴\*\* | 预注册 8 类操作（内容扩展、事实纠正、失败修复、重构等），Opus 4.8 双通道编码 | 8 类 kappa = 0.62 ；折叠为 corrective/enhancement 二分后 kappa = 0.64 （原始一致率 0.82） | | \*\*触发证据轴\*\* | 0–3 级阶梯（L3 显式问题链接、L2 命名失败、L1 通用、L0 无） | 模型双通道 kappa = 0.59 （边界水平） | | \*\*规则相似性轴\*\* | 预注册 3 类：rule-like / instance-bound / mixed | \*\*可靠性门控失败\*\*： kappa = -0.02 ，按预注册承诺不报告分布 | - \*\*人类验证\*\*：单编码者对 50 编辑重叠样本进行盲法重编码。8 类操作人类-LLM 一致率为 kappa = 0.46 ；corrective/enhancement 二分达 kappa = 0.72 （95% CI $

0.51, 0.88
），支持 headline 发现建立在二分折叠上。 - 跨模型族验证：使用 gpt-4o 与 gpt-5.5 对 50 编辑重新编码。8 类 kappa 降至 0.36–0.40；但二分折叠仍保持 kappa = 0.51 –0.61，进一步证实该折叠的稳健性。 —- ### 3. 维护内容与演化动态分析实验在编码基础上展开描述性分析： - 操作分布量化：254 个编辑中，60% 为 enhancement（perfective）、38% 为 corrective、2% 为 other。内容扩展（72）与事实纠正（56）合计占半数；排除多文件大规模重构后，内容扩展、失败修复与重构仍稳健，事实纠正从第 2 大类别降至第 4。 - 规模轨迹：对 120 个至少两次观测的技能文件，追踪 resident token 长度净变化。32 个增长超 10%，7 个缩减超 10%，81 个稳定，整体呈稳定至累加态势。 - 编辑节奏：全部技能文件提交间隔的中位数为 5 天；各仓库差异显著（obra 中位 1.5 天，trailofbits 中位 14.5 天）。 - 组件归因：将 SKILL.md 解析为 6 个组件（frontmatter、router、instructions、examples、code、references），85% 的编辑触及 instructions，56% 触及 code，38% 触及 router，其余组件变更比例更低。 —- ### 4. 治理结构分析实验基于提交元数据对作者身份与仓库控制模式进行统计： - AI 共同作者标记 prevalence：62%（158/254）的实质性编辑携带 AI co-author trailer，38% 无可见 AI 标记；按 commit 计为 61%（114/187）。所有编辑均通过具名人类账户 author 或 merge，未观察到 agent-only 实质性提交。 - 组织异质性：5 个仓库呈双峰分布——obra 几乎全为人类作者；anthropics 几乎全为人类网页合并且无 AI trailer；getsentry 与 trailofbits 的 AI trailer 比例超过 90%。 - 规模与范围检验：Mann-Whitney 检验显示 AI-trailered 与 AI-trailer-absent 编辑在变更行数（中位 15 vs. 21， p = 0.20 ）与触及文件数（中位 4 vs. 6， p = 0.31 ）上无显著差异。 - 组件-治理关联检验：对 6 个组件分别执行 Fisher 精确检验（Holm 校正）、排除大规模重构后的 Fisher 检验、以及仓库分层 Cochran-Mantel-Haenszel 检验。无一组件能在全部三项控制下存活；frontmatter 的 90% AI-trailered 表象完全由多文件大规模重构的 fan-out 效应导致。 —- ### 5. 维护效用迁移任务实验（§7 与 Appendix D）为检验“维护是否提升技能效用”，设计预注册的受控迁移任务实验： - 样本：13 个具有 ≥ 6 次实质性编辑的公开技能，跨 4 个仓库；每个技能配 11 个迁移任务，共 143 个任务。 - 实验条件： no-skill, v(old), v(new) ，其中 v(old) 为窗口内最早版本， v(new) 为最新版本；每单元 3 次独立重复。 - 求解器：gpt-5.4-mini（推理模型，与评判器不同模型族），通过按需文件读取暴露完整技能包，并采用均等化多轮交互协议（最多 3 轮，由约束用户模拟器回答澄清问题）。 - 评判：Claude 模型盲评去标识化、指纹清除后的输出，按 1–5 分评价 correctness、coverage、actionability、non-redundancy；另设 gpt-5.5 参考面板覆盖 100 个 triad 以估计跨族一致性。 - 统计设计：推断单元为技能（ n = 13 ），预注册功效 0.77–0.90 以检测 pilot 规模的效应；采用符号检验与 Wilcoxon 符号秩检验。 主要结果： - 技能层面的 v(new) - v(old) 平均差异为 -0.09 （95% CI

-0.28, +0.10
）；符号检验 p = 0.58 ，Wilcoxon p = 0.55 ；13 个技能中仅 5 个倾向 v(new) 。 - 稳健性检验：长度控制残差分析（ -0.06 ， p = 0.58 ）、deep-five 子集（ -0.14 ）、弱求解器 gpt-4o-mini 重跑（ ≈ 0 ）均保持零效应。 - 评判器可靠性：100 triad 重叠上的 Pearson r = 0.57 ，Spearman rho = 0.51 ，ICC(2,1) = 0.52，低于预注册的 0.6 门控，提示测量噪声可能衰减了小效应。 —- ### 6. 策展重放协议设计（Appendix E）论文未运行自动化策展基线，但发布了一项回溯性重放协议，用于未来系统的离线评估： - 输入：历史时刻的 SKILL.md 版本及可用上下文（提交信息、关联 issue/PR、仓库）。 - 目标：预测下一次人类实质性编辑，匹配操作标签、触及组件、与真实 diff 的文本相似度。 - 评估指标：操作与组件匹配率、token 重叠度、ROUGE-L 摘要相似度，以及是否触发 consolidation/deprecation（观察到的基准率约 4%）。 Q5: 有什么可以进一步探索的点？基于论文的发现与局限性，以下方向值得进一步探索： —- ### 一、测量方法与编码可靠性 - 规则相似性（rule-likeness）的可操作化：论文预注册的“规则-实例”轴因信度不足（ kappa = -0.02 ）而被迫弃用。未来可探索基于句法模式（如条件语句、禁止性模态词）或语义约束的替代方案，或引入维护者访谈作为效标，以解决“某次编辑是否编码了可泛化规则”这一核心测量难题。 - 触发证据（trigger evidence）的稳定标注：同一批编辑在不同模型族下的具体失败证据比例从 24% 跳变至 63%，说明基于提交元数据的失败溯源极不稳定。可尝试结合 issue/PR 文本的显式链接解析、或执行日志的异常匹配，构建更客观的触发因素识别管道。 - AI 参与度的精确估计：Co-Authored-By trailer 既可能遗漏未披露的 AI 使用，也可能包含自动注入的噪声。可探索基于编辑时间模式、代码风格计量或维护者访谈的三角测量方法，以校准“真实 AI 贡献”与“可见治理信号”之间的差距。 —- ### 二、语料库扩展与外部效度 - 代表性样本与组织类型：现有语料为 purposive 选取的五个 AI 工具组织，维护深度可能为上限估计。未来可向非 AI 原生组织、个人开发者、企业私有仓库（在合规前提下）扩展，检验治理模式与操作分布是否随组织规模、领域和开源文化系统性变化。 - 更长观察窗口与技能生命周期：当前数据右删失于 5–8 个月，难以观测技能的长期衰减、退休或被替代模式。延长追踪周期可揭示技能库是否存在“过度增长—性能阴影—大规模清理”的周期性动态。 - 跨文件依赖与技能网络：论文以单文件为分析单元，未涉及技能间的引用、冲突或冗余关系。可构建技能依赖图，分析维护事件如何在网络中传播（如某工具 API 变更触发的级联修正）。 —- ### 三、自动化策展器的设计与评估 - 建立重放协议基线：论文已发布 replay 协议，但尚未运行自动化策展系统的基线实验。可直接在该协议上评测现有 SOTA 策展器（如 SkillOS、Trace2Skill 的变体），量化其操作类型匹配率、组件定位准确率及与真实人类 diff 的 token 重叠度。 - 显式修剪与合并预算：观察到 consolidation 与 deprecation 仅占 4.3%，技能库呈累积式增长。可探索带显式“修剪预算”的策展算法，如定期触发技能合并、冗余检测或基于使用频率的退休机制，以对抗无界增长。 - 从维护记录中蒸馏规则：鉴于 rule-likeness 的编码困难，可尝试不直接对单次编辑分类，而是将同一技能的多次编辑聚合为时序模式，通过聚类或概念学习从中抽取稳定约束，再经人类审查写回技能文件。 —- ### 四、维护效果的因果机制 - 真实工作流中的效用检验：迁移任务实验（ S 7）在受控 harness 下得到零效应，但任务为作者构造，且评判器信度未达门控。未来可在真实 IDE 环境或 agent 执行轨迹中，对比“维护后技能”与“冻结版本”的任务成功率，以验证维护在原生场景中的价值。 - 维护异质性的来源解析：论文发现不同技能对维护的反应方向相反（如 brainstorming 显著下降而 pr-writer 提升）。可通过细粒度 diff 分析，识别导致正向迁移的编辑特征（如输出约束、边界条件）与导致负向迁移的特征（如过度流程化），建立编辑特征与下游效用的映射。 - 技能阴影（skill shadowing）与维护的关系：结合 Song & Wei (2026) 的发现，可研究维护行为（尤其是内容扩展）是否加剧了技能库中的阴影效应，以及何种维护操作（如重构、路由器调优）能缓解该问题。 —- ### 五、治理与组织因素 - 治理双峰的制度解释：论文发现组织间 AI 共著比例呈双峰（0% 与 90%+），但未能区分这是真实 AI 使用差异、还是 squash 合并工作流或披露文化差异所致。可通过仓库策略文件、维护者问卷或提交钩子日志，解释治理模式的组织决定因素。 - 人类-AI 协作的微观过程：当前仅在提交级别记录“人类合并/AI 共著”，未捕捉单次编辑内部的协作迭代（如人类提示工程、AI 生成草稿、人类再编辑）。对 Pull Request 评论、内联 diff 对话的细粒度挖掘，可揭示具体的分工模式。 - 治理信号的预测价值：当前发现治理身份（AI-trailered vs. human-only）既不预测编辑大小也不预测触及组件。若能在更大规模上确认此零效应，则意味着自动化策展器无需按作者身份差异化路由；反之，若发现特定治理模式与特定操作类型关联，则可设计针对性的审查工作流。 Q6: 总结一下论文的主要内容这篇论文对公共 AI 技能仓库中的人类治理、AI 辅助的技能维护过程进行了首次纵向实证研究，核心内容可概括如下： —- ### 1. 研究背景与动机终身 LLM 智能体越来越多地依赖外部 Markdown 技能文件（如 SKILL.md）来保存和复用能力。这些技能需要随工具变化和用法演进而持续修正、扩展与合并。近期研究致力于自动化技能策展，但主要将人工维护视为“无法扩展的瓶颈”，且缺乏对这一过程本身的系统测量。论文指出，自动化策展器若要真正替代或辅助人类，必须首先理解公共仓库中真实的维护行为由谁执行、包含何种操作。 —- ### 2. 研究问题论文围绕一个前提性实证问题展开： - 谁在维护公共仓库中的智能体技能？ - 这种维护具体由哪些操作构成，呈现何种演化动态？ —- ### 3. 研究方法与语料库研究采用纵向仓库挖掘与预注册编码相结合的方法： - 语料库： purposive 选取 5 个公共 AI 技能仓库（如 getsentry/skills、anthropics/skills 等），固定提交快照，覆盖 873 个提交、143 个 SKILL.md 文件。 - 分析单元：从中筛选出 254 个实质性创建后编辑（修改不少于 5 行、非机器人、非琐碎变更）。 - 三轴编码手册： - 治理轴：基于提交元数据记录作者身份（AI co-author trailer、人类 author/committer、网页合并）。 - 策展操作轴：预注册的 8 类维护操作（内容扩展、事实纠正、失败修复、重构、合并、弃用等），可折叠为纠正性（corrective）与增强性（enhancement）二分。 - 触发证据轴：0–3 级阶梯，量化提交记录中可观测的失败触发证据。 - 可靠性控制：通过 LLM 双通道编码、盲法人类重编码（ kappa = 0.72 ）及跨模型族验证，确保 headline 发现建立在可信的纠正/增强二分之上；对未通过信度门控的“规则相似性”轴，按预注册承诺不予报告。 —- ### 4. 主要发现 #### （1）治理结构：人类治理、AI 辅助 - 100% 的实质性编辑均通过具名人类账户撰写或合并，未观察到智能体独立完成的实质性提交。 - 62% 的编辑携带 AI 共同作者标记（Co-Authored-By trailer），38% 无可见 AI 标记。 - 该比例呈组织依赖的双峰分布：有的仓库超过 90% 的编辑为 AI 共著，有的则几乎完全由人类完成或仅通过人类网页合并，反映披露文化与工作流差异，而非单纯的 AI 使用率差异。 #### （2）维护内容：以增长和修正为主的真实策展 - 折叠为纠正/增强二分后，60% 为增强性编辑，38% 为纠正性编辑，2% 为其他。 - 具体操作以内容扩展（72 次）和事实纠正（56 次）为主；排除大规模重构后，内容扩展、失败修复与重构仍稳健，事实纠正占比下降。 - 整合与弃用极少（合计仅 4.3%），技能库呈累加式增长，鲜见主动修剪。 #### （3）演化动态：频繁、稳定至累加 - 编辑节奏中位数为 5 天，组织间差异显著（从 1.5 天到 14.5 天不等）。 - 规模轨迹以稳定为主：120 个可追踪技能中，81 个规模稳定，32 个增长超过 10%，仅 7 个缩减超过 10%。 - 变更高度集中于技能主体（instructions 占 85%，code 占 56%），路由器、前言、示例等边缘组件变更较少。 #### （4）维护效用的迁移任务检验：零效应 - 对 13 个技能的预注册迁移任务实验（11 任务/技能，3 条件，跨族评判器）显示： - 维护版本（ v(new) ）与早期版本（ v_(old) ）在评判质量上无显著平均差异（均值差异 -0.09 ，95% CI

-0.28, +0.10
）。 - 单个技能效应异质（5 个提升，8 个下降），但聚合层面未复现早期 pilot 的正向趋势。 #### （5）规则相似性的测量失败 - 预注册的“规则相似性”（rule-like vs. instance-bound）轴在双通道编码中未能通过可靠性门控（ kappa = -0.02$）。论文据此指出，从提交产物中可靠区分“泛化规则”与“单例修复”仍是开放的测量问题。 —- ### 5. 贡献与意义 - 实证贡献：首次系统测量了公共 Markdown 技能文件的纵向维护历史，揭示了当前技能维护的真实形态——一个人类主导、AI 辅助的循环，而非全自动管道。 - 资源贡献：发布了完整的提交级语料库、三份编码手册、挖掘与编码脚本，以及面向自动化策展系统的回溯性重放协议（replay protocol）。 - 设计启示：为自动化策展器提出五项 grounded 假设，包括需显式预算用于修剪、保持人类账户的合并控制、优先按操作类型而非大小编辑分类、以及按仓库建模治理文化等。 - 负面结果的价值：明确报告了 rule-likeness 轴的可靠性失败，为未来研究厘清了测量障碍。 Q7: 想要进一步了解论文以上只是了解一篇论文的几个基本FAQ。如果你还想与Kimi进一步讨论该论文，请点击这里为你跳转Kimi AI网页版，并启动一个与该论文相关的新会话。

Authors: Chen Shen, Estevam Hruschka

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05677.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05677

Published: 2026-09-10T01:26:51.917Z

5. A Rubric-Guided Large Language Model Solution for Opioid Use Disorder Computable Phenotyping

Abstract:Opioid use disorder (OUD) remains a public health crisis in the United States, yet it is difficult to identify from electronic health records (EHRs) because missing diagnosis codes and supporting evidence are buried in clinical narratives. Accurate OUD identification is critical to support interventions and improve health outcomes. This study developed a rubric-guided large language model (LLM) that incorporated Optimization by PROmpting (OPRO) for OUD computable phenotyping (CP). The framework used an 18-item, expert-identified rubric to instruct LLMs to automatically extract critical text with supporting evidence to determine OUD flags. Two UF Health physicians (GMR and WMG) chart-reviewed 253 patients, including 68 OUD-positive cases. Our LLM-based computable phenotype (CP) achieved the best F1 score of 0.774 and an AUROC of 0.934, outperforming the machine learning-based CP using EHR and natural language processing-extracted variables, and zero-shot LLMs by relative F1 improvements of 12.8% and 44.4%, respectively. The proposed LLM-based CP could link LLM-extracted evidence to OUD phenotyping for better explainability.

中文摘要

摘要：阿片类药物使用障碍（OUD）仍然是美国的公共卫生危机，但由于电子健康记录（EHR）中的诊断编码缺失且支持证据埋藏在临床叙述中，因此很难识别。准确的OUD识别对于支持干预措施和改善健康结果至关重要。本研究开发了一种由评分标准指导的大型语言模型（LLM），该模型结合了通过提示优化（OPRO）用于OUD可计算表型（CP）。该框架使用了18项由专家确定的评分标准来指导LLM自动提取关键文本及其支持证据，以确定OUD标志。两位UF Health的医生（GMR和WMG）对253名患者进行了病历审查，其中包括68例OUD阳性病例。基于LLM的可计算表型（CP）达到了最佳F1分数0.774和AUROC 0.934，优于使用EHR和自然语言处理提取变量的机器学习CP以及零样本LLM，F1相对提高幅度分别为12.8%和44.4%。所提出的基于LLM的CP可以将LLM提取的证据与OUD表型联系起来，从而具有更好的可解释性。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05682 (HTTP 429)

Authors: Mengxian Lyu, Paredes Pardo, Cheng Peng, Ziyi Chen, Mengyuan Zhang, Jieting Li Lu, Gary M Reisfield, William M Greene, Jenny Lo-Ciganic, Yonghui Wu

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05682.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05682

Published: 2026-09-10T01:26:51.917Z

6. Intra-Prompt Parallel Decoding for Common-Context Question Answering

Abstract:In common-context question answering (CCQA) tasks, multiple input questions share a common context to base their answers from. However, Large Language Models typically generate each answer using an independent prompt. While existing batching and caching techniques help improve parallelism and reduce repeated computations, the separation of questions across prompts limits the achievable speedup, as modern GPUs are underutilized due to a memory bottleneck during attention. We present Intra-Prompt Parallel Decoding (IPPD), a novel inference method that answers multiple common-context questions in parallel within a single prompt. IPPD directly addresses the bottleneck by efficiently sharing both memory and computation during the attention process, as the next token for every question is decoded in a single inference step. IPPD uses virtual position IDs and attention mask manipulation to generate the same output as standard prompting without requiring fine-tuning or any changes to the LLM architecture. Since all parallelism occurs within a prompt, IPPD is fully compatible with batched inference, even when each prompt features a different context. Our experiments show that IPPD delivers up to 7X the effective throughput as standard decoding without quality degradation, and outperforms prefix caching with PagedAttention in most settings.

中文摘要

摘要：在常见上下文问答（CCQA）任务中，多个输入问题共享一个用于回答的公共上下文。然而，大型语言模型通常使用独立的提示生成每个答案。尽管现有的批处理和缓存技术有助于提高并行性并减少重复计算，但由于问题被分散到不同的提示中，限制了可实现的加速效果，因为在注意力机制过程中由于内存瓶颈，现代 GPU 的利用率不足。我们提出了提示内并行解码（IPPD），这是一种新颖的推理方法，可在单个提示内并行回答多个公共上下文问题。IPPD 通过在注意力过程中高效共享内存和计算，直接解决了瓶颈问题，每个问题的下一个 token 都在单次推理步骤中解码。IPPD 利用虚拟位置 ID 和注意力掩码操作来生成与标准提示相同的输出，无需微调或修改 LLM 架构。由于所有并行操作都发生在单个提示内，IPPD 与批量推理完全兼容，即使每个提示包含不同的上下文。我们的实验表明，IPPD 在不降低质量的情况下，实现的有效吞吐量比标准解码高出最多 7 倍，并且在大多数设置中优于采用 PagedAttention 的前缀缓存。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05707 (HTTP 429)

Authors: Theodore Glavas, Nikhita Vedula, Dushyanta Dhyani, Antonios Valkanas, Yilun Zhu, Shervin Malmasi

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05707.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05707

Published: 2026-09-10T01:26:51.917Z

7. MedWER: A Reproducible, Model-Free Evaluation Protocol for Medical Speech Recognition

Abstract:Overall word error rate hides clinically critical errors: a transcript can be 95% correct and still swap one drug for another. The usual fix weights errors on medical entities, and almost always depends on an evaluation-time named-entity recognition (NER) model or cloud API, which makes the metric’s denominator a versioned black box. We present MedWER, an evaluation protocol and open-source tool for medical ASR whose denominator is a fixed, license-clean term list: 19,373 drug, diagnosis, symptom, and injury-mechanism entries projected from public sources. The protocol couples a pinned text normalizer with a phrase-aware term-restricted WER, the MedWER, so the only versioned component is a normalizer dependency held at an exact release and checked against committed golden fixtures. Coverage is validated against an independent provincial drug-benefit file the list was not built from; the matching heuristic is calibrated against ground-truth entity spans. Baselines for Moonshine~base, Whisper~this http URL, and MedASR on two open benchmarks are scored with the released tool and reported with 95% confidence intervals from resampled per-utterance scores.

中文摘要

摘要：总体词错误率掩盖了临床上关键的错误：一个转录文本可以 95% 正确，但仍然可能将一种药物替换为另一种。通常的解决方案是对医疗实体的错误加权，并且几乎总是依赖于评估时的命名实体识别（NER）模型或云 API，这使得指标的分母成为一个版本化的黑箱。我们提出了 MedWER，一种用于医疗 ASR 的评估协议和开源工具，其分母是固定的、许可清晰的术语列表：从公共来源投射出的 19,373 个药物、诊断、症状和损伤机制条目。该协议将固定的文本标准化器与短语感知的、受术语限制的 WER（MedWER）结合，使得唯一的版本化组件是以确切版本发布并与提交的黄金样本核对的标准化器依赖项。覆盖范围通过与构建列表时未使用的独立省级药物补助文件进行验证；匹配启发式方法基于真实实体范围进行校准。在两个开放基准上，使用发布的工具对 Moonshine~base、Whisper~这个 http URL 和 MedASR 的基线进行评分，并报告每次语音重采样得分的 95% 置信区间。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05728 (HTTP 429)

Authors: Justin Behling

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05728.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05728

Published: 2026-09-10T01:26:51.917Z

8. Some Tokens Behave like Magnets: Revealing Linguistic Organization in the Layers of Language Models

Abstract:We identify a special group of token vectors inside large language models (LLMs), which we term magnetic vectors, that organize the surrounding tokens by either attracting or repelling them. Particularly, tokens pointing the same way as an attracting magnet are elongated; tokens pointing the same way as a repelling magnet are compressed. Just as physical magnets pull or push away the iron filings around them, these vectors organize their surroundings through two opposing polarities. Moreover, we identify a statistically significant pattern in linguistic category where function words consistently act as repelling magnets in early layers, and we also find magnets consistently reorganize their polarities in unique ways deeper in the model. In a further case study we find this observation may unveil a deliberate, layer-wise organization in how LLMs process language. This pattern is consistent across different LLM architectures, sizes, and layer configurations. It is also causally relevant. When the LLM is fine-tuned for a downstream task, the task-functional tokens emerge as magnets. E.g., in question answering, the answer-span tokens become uniquely repelling magnets in the final layer, geometrically carving the answer out of the surrounding context. Furthermore, removing early-layer repelling magnets devastates syntactic tasks (POS tagging accuracy drops from 91% to below 10%) while sparing semantic ones, and removing late-layer attracting magnets does the reverse. We believe this phenomenon warrants further investigation, as it opens the first probe-free path to understanding how language models geometrically organize linguistic computation across their layers.

中文摘要

摘要：我们在大型语言模型（LLMs）中识别出一组特殊的词向量，我们称之为磁性向量，它们通过吸引或排斥周围的词来组织这些词。具体来说，指向像吸引磁铁方向的词会被拉长；指向像排斥磁铁方向的词会被压缩。就像物理磁铁拉动或排斥周围的铁粉一样，这些向量通过两种相反的极性组织它们的周围环境。此外，我们在语言类别中识别出一个具有统计学显著性的模式：功能词在早期层始终表现为排斥磁铁；同时我们发现磁性向量在模型更深层中总是以独特的方式重新组织它们的极性。在进一步的案例研究中，我们发现这一观察可能揭示了LLMs处理语言时逐层安排的刻意结构。这个模式在不同的LLM架构、规模和层配置中是一致的，并且具有因果相关性。当LLM为下游任务进行微调时，任务功能词会作为磁性向量显现出来。例如，在问答任务中，答案跨度（answer-span）词在最后一层会成为独特的排斥磁铁，从几何上将答案从周围的上下文中分离出来。此外，移除早期层的排斥磁铁会严重破坏句法任务（词性标注准确率从91%下降到低于10%），而语义任务则不受影响；而移除晚期层的吸引磁铁则情况相反。我们认为这一现象值得进一步研究，因为它为理解语言模型如何在各层几何组织语言计算开辟了首条无需探针的方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05743 (HTTP 429)

Authors: Andrew Liu, Devan Srinivasan, Gerald Penn

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05743.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05743

Published: 2026-09-10T01:26:51.917Z

9. CrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse

Abstract:Identifying the target of emotional words or phrases in crisis situations, especially health-related ones, is important for understanding public concerns across cultural and linguistic contexts. We propose CrisisKD, a five-stage teacher—student knowledge distillation framework for aspect-level sentiment and emotion analysis on unannotated social media data. A teacher LLM generates aspect-level labels and reasoning traces that supervise a smaller student model across aspect extraction, syntactic parsing, opinion extraction, sentiment classification, and emotion classification. Using this framework, we construct and release a dataset containing 50,615 aspect-level labels, together with the annotation and fine-tuning scripts as open-source resources. The resulting student supports end-to-end ABSA and emotion detection at substantially lower inference cost than the teacher. On a manually annotated 500-tweet gold set, the 5-task Qwen2.5-7B student improves over the untuned model by 7.9 F1 points on aspect extraction, 17.0 points on emotion accuracy, and 6.5 points on sentiment accuracy. On the external ABEA benchmark, CrisisKD improves the same-model Qwen2.5-7B ICL baseline by 2.8 F1 points on ATE and 3.8 F1 points on joint ATE+AEC.

中文摘要

摘要：在危机情境下，尤其是与健康相关的危机中，识别情绪词或短语的目标，对于理解跨文化和语言环境下的公众关切非常重要。我们提出了 CrisisKD，一种五阶段的教师－学生知识蒸馏框架，用于对未标注的社交媒体数据进行层面级情感和情绪分析。教师大语言模型（LLM）生成层面级标签和推理轨迹，用于监督较小的学生模型在层面提取、句法解析、观点提取、情感分类和情绪分类上的学习。利用该框架，我们构建并发布了一个包含 50,615 个层面级标签的数据集，同时提供标注和微调脚本作为开源资源。所得学生模型支持端到端的 ABSA 和情绪检测，其推理成本远低于教师模型。在手工标注的 500 条推文黄金数据集上，进行五任务训练的 Qwen2.5-7B 学生模型在层面提取上比未调优模型提升了 7.9 个 F1 分，在情绪准确率上提升了 17.0 个百分点，在情感准确率上提升了 6.5 个百分点。在外部 ABEA 基准测试上，CrisisKD 比相同模型的 Qwen2.5-7B ICL 基线在 ATE 上提升了 2.8 个 F1 分，在联合 ATE+AEC 上提升了 3.8 个 F1 分。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05757 (HTTP 429)

Authors: Marko Haralović, Onat Akca, Salih Eren Yücetürk, Minsi Li, Mariët Theune

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05757.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05757

Published: 2026-09-10T01:26:51.917Z

10. Recall Is Not Protection: Evaluating Safety Monitors Against Model Compliance

Abstract:Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt \emph{elicitable} if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer.

中文摘要

摘要：安全监控器会筛查发送到已部署语言模型的提示，标记有害请求，以确保这些请求永远不会得到回答。它们通过与有害性标签的召回进行评估，但只有在模型本来会遵从时，拦截才会防止伤害。我们直接测量这种差异：我们从目标模型中抽样重复响应，如果模型至少有一次遵从，我们将有害提示称为“可引发的”，并在可引发提示和不可引发提示上分别报告监控器的召回率。在六种监控器配置和三个模型家族中，涵盖激活探针、微调文本守卫以及一个120B策略条件推理分类器，在固定的假阳性率下，可引发提示的召回率比不可引发提示低0.22到0.38。监控器遗漏的提示比它捕获的提示更有可能被遵从，概率高2.8到5.6倍。这个差距在三个模型家族中重复出现，并且在完全独立于目标模型的纯文本监控器中也出现。这表明，标准召回率可能高估了监控器在实际中提供的保护，并且监控器应根据其模型实际上会回答的内容进行评估。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05797 (HTTP 429)

Authors: Sripad Karne

Categories: cs.CL

PDF URL: https://arxiv.org/pdf/2609.05797.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05797

Published: 2026-09-10T01:26:51.917Z

Agent Domain Papers

Abstract:Previous AI alignment efforts have focused primarily on first-order social norms — teaching models what is socially acceptable or unacceptable (e.g., `do not steal’). However, social intelligence depends not only on norm recognition, but also on anticipating who will enforce it and how (e.g., public shame or even imprisonment). These second-order expectations, known as metanorms, govern how people respond when social rules are broken. We introduce a novel framework for evaluating metanorm reasoning in Large Language Models (LLMs) along two dimensions: emotional appraisal and behavioral response, and propose new classification tasks, namely, predicting self-regulation in violators, and other-regulation in observers. We release a multi-perspective dataset, NormReact, of 450 norm violation scenarios, hand-annotated for emotions and behavioral responses across norm violators’ gender and observers’ social closeness. Current LLMs portray a harsher social world: across six models, they overpredict negative sanctions where humans would expect inaction, and alignment with human judgments deteriorates as social distance increases. These findings suggest that AI systems in norm-sensitive domains from conflict mediation to policy simulation, may risk producing a distorted picture of social regulation: one that over-represents punishment and under-represents the tolerance, restraint, and relational calibration that characterize actual norm enforcement in real world.

中文摘要

摘要：以往的AI对齐工作主要集中在一级社会规范——教模型什么是社会上可接受或不可接受的行为（例如，“不要偷窃”）。然而，社会智力不仅依赖于规范识别，还依赖于预测谁会执行规范以及如何执行（例如，公开羞辱甚至监禁）。这些二级预期，被称为元规范，决定了人们在社会规则被违反时的反应。我们提出了一个用于评估大型语言模型（LLM）元规范推理的新框架，从两个维度进行评估：情绪评价和行为反应，并提出了新的分类任务，即预测违规者的自我调节，以及观察者的他人调节。我们发布了一个多视角数据集NormReact，包含450个规范违规场景，对违规者的性别和观察者的社会关系亲密度下的情绪和行为反应进行了手工标注。当前的LLM描绘了一个更严苛的社会世界：在六个模型中，它们在本应无行动的人类预期下过度预测负面惩罚，且随着社会距离的增加，与人类判断的一致性下降。这些发现表明，在从冲突调解到政策模拟的规范敏感领域中，AI系统可能会产生对社会规范执行的扭曲认知：即过度表现惩罚，低估现实世界中实际规范执行所体现的宽容、自我约束和关系校准。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05437 (HTTP 429)

Authors: Sunny Rai, Jinyi Kuang, Reyhan Jamalova, Annie Lou, Cristina Bicchieri, Niyati Malhotra, Victor Hugo Orozco-Olvera, Ana Maria Munoz-Boudet, Lyle H Ungar, Sharath C Guntuku

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05437.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05437

Published: 2026-09-10T01:27:48.385Z

2. CriticGen: Generation-Aware Evaluation as Actionable Feedback

Abstract:Current evaluation methods for large language models are coarse-grained and decoupled from generation, producing generic explanations that fail to provide actionable feedback for model improvement. We propose CriticGen, a fine-grained, generation-aware evaluation framework that turns evaluation into actionable control for answer improvement. CriticGen first generates sample-specific evaluation dimensions and scoring criteria under high-level categories such as subjective, objective, and self-derived constraints. These criteria then serve as a dynamic rubric for jointly producing a score, a reason, an executable refinement suggestion, and a refined answer. This rubric-conditioned refinement process enables models to diagnose flaws and perform targeted answer improvement. Experimental results show that fine-grained evaluation should be both instance-specific and actionable. CriticGen induces higher-quality rubrics, improving relevance/coverage from 3.33/4.03 to 3.97/4.24. CriticGen also achieves the best score correlations, with 0.9556 Pearson and 0.9560 Spearman, and raises the F1 of criterion-grounded reasons and executable suggestions from 0.6369/0.5994 to 0.7554/0.7900. Crucially, its feedback translates into reliable answer improvement, improving 73.17% of answers with a 93.28% non-degradation rate.

中文摘要

摘要：当前大型语言模型的评估方法粗粒度且与生成过程脱节，产生的解释通常过于笼统，无法为模型改进提供可操作的反馈。我们提出了CriticGen，一种细粒度、生成感知的评估框架，将评估转化为可执行的答案改进控制。CriticGen首先在主观、客观和自主衍生约束等高层类别下，生成针对具体样本的评估维度和评分标准。随后，这些标准作为动态评分表，用于联合生成分数、原因、可执行的改进建议以及改进后的答案。这一基于评分表的改进过程使模型能够诊断缺陷并进行针对性的答案优化。实验结果表明，细粒度评估应既针对具体实例又具有可操作性。CriticGen引导出更高质量的评分表，将相关性/覆盖率从3.33/4.03提升至3.97/4.24。CriticGen还实现了最佳的分数相关性，皮尔逊相关系数为0.9556，斯皮尔曼相关系数为0.9560，同时将基于标准的理由和可执行建议的F1从0.6369/0.5994提升至0.7554/0.7900。关键是，其反馈能够转化为可靠的答案改进，提升了73.17%的答案，并保持了93.28%的非下降率。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05439 (HTTP 429)

Authors: Huifang Du, Zecheng Zuo, Sen Wang, Chenghao Fan, Haofen Wang, Yehui Yang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05439.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05439

Published: 2026-09-10T01:27:48.385Z

3. When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

Abstract:Long-term memory for LLM agents is evaluated today by conversational recall benchmarks (LoCoMo, LongMemEval), which measure question answering over dialogue history, not whether remembered facts change what a tool-using agent does. We present MERIT (Memory Evaluation for Realistic Instrumented Tasks), a benchmark and harness that measures the marginal utility of memory for task-executing agents under explicit cost accounting. MERIT provides episodic tool-use tasks in three domains whose dependence on earlier-episode facts is verified by an automated leak check; a difficulty ladder ending in updated-fact recall; controlled memory corruption; and full token and dollar metering of every memory operation. Across 23,440 scored episodes ($42.57), a two-generation pilot on gpt-4.1-mini and a preregistered 3-model x 3-seed grid (GPT-4.1, Claude Haiku 4.5; memory side held fixed), memory lifts dependent-task success from a leak-verified floor of 0.00 to 0.55-1.00. On updated facts, embedding retrieval collapses unpredictably (0.30-0.95 across models; max seed gap 0.45), and agents act on a correctly retrieved value only 55% of the time, while update-on-write stores (a structured fact store and, notably, LLM summarization) remain at 0.70-1.00; the hybrid is worse than the fact store alone. A latest-generation spot-check (Claude Sonnet 5, gated on a clean full-replay control) reproduces the pattern. Swapping a memory’s implementation moves task success by up to 60 points, and full replay is never economical: the best condition per domain delivers 2.7-3.9x its marginal utility per dollar. We release the benchmark, harness, and all traces.

中文摘要

摘要：目前对大型语言模型（LLM）智能体的长期记忆评估主要通过对话回忆基准（LoCoMo, LongMemEval）进行，这些基准衡量的是对话历史的问答能力，而非记忆事实是否会改变使用工具的智能体的行为。我们提出了 MERIT（现实工具任务记忆评估），这是一个基准和框架，用于在明确成本核算下衡量记忆对执行任务智能体的边际效用。MERIT 提供三个领域的情景工具使用任务，其对早期情节事实的依赖性通过自动泄漏检查进行验证；一个以更新事实回忆为终点的难度阶梯；受控记忆损坏；以及对每个记忆操作进行完整的代币和美元计量。在 23,440 个评分情节（$42.57）中，对 gpt-4.1-mini 进行的两代试点，以及预注册的 3 模型 x 3 种随机种子网格（GPT-4.1、Claude Haiku 4.5；记忆边保持固定），记忆将依赖任务的成功率从经过泄漏验证的 0.00 提升至 0.55-1.00。在更新事实方面，嵌入检索的表现无法预测地崩溃（各模型间为 0.30-0.95；最大种子差距为 0.45），且智能体仅在 55% 情况下根据正确检索的值采取行动，而写入时更新的存储（结构化事实存储，尤其是 LLM 摘要）维持在 0.70-1.00；混合模式的表现还不如单独的事实存储。最新一代的抽查（Claude Sonnet 5，在干净的完全重放控制下）复制了这一模式。更换记忆实现方式会使任务成功率变化高达 60 个点，完全重放从未经济：每个领域的最佳条件每美元提供的边际效用为 2.7-3.9 倍。我们发布该基准、框架及所有追踪数据。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05441 (HTTP 429)

Authors: Shweta Mishra, Shashank Mishra

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05441.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05441

Published: 2026-09-10T01:27:48.385Z

4. AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

Abstract:We introduce AutoFyn, an agent harness inspired by the Expert Iteration algorithm, adapting a frozen model across many rounds by updating persistent state from verified reward signals rather than model weights. Each round begins from a fresh model session, and durable information is reintroduced only through explicit interfaces such as persistent memory files, reports, and repository state. Within a round, an orchestrator explores, plans and builds many alternative approaches with specialized agents, while a task-grounded verifier verifies the work and supplies an objective reward for measuring progress. This reward is distilled back into the persistent state, which updates the effective policy for the next round. In this technical report, we formalize this loop and describe its persistent state and verification interfaces. We then demonstrate its use in three domains, namely olympiad mathematics, data science, and cybersecurity. On the six fresh problems of the 2026 International Mathematical Olympiad, every model with room to improve scores higher under AutoFyn than in its provider’s own coding agent. AutoFyn also built the top-ranked agent on the Spider 2.0 dbt benchmark, and has produced $16$ maintainer-confirmed vulnerability advisories in this http URL, MetaMask, pnpm, Warp, LiteLLM, Langflow, and Open WebUI.

中文摘要

摘要：我们介绍了 AutoFyn，这是一种受专家迭代算法启发的代理工具，通过从已验证的奖励信号而非模型权重更新持久状态，在多轮中适应一个冻结模型。每一轮都从一个全新的模型会话开始，持久信息仅通过显式接口（如持久内存文件、报告和存储库状态）重新引入。在一轮中，协调器会探索、规划并使用专门的代理构建多种替代方法，而基于任务的验证器会验证工作并提供客观奖励以衡量进展。这个奖励会被提炼回持久状态，从而更新下一轮的有效策略。在本技术报告中，我们形式化了这一循环，并描述了其持久状态和验证接口。随后，我们展示了它在三个领域的应用，即奥林匹克数学、数据科学和网络安全。在 2026 年国际数学奥林匹克的六个全新问题上，每个有改善空间的模型在 AutoFyn 下的得分都高于其提供方自身的编码代理。AutoFyn 还构建了 Spider 2.0 dbt 基准上排名最高的代理，并在此 http URL、MetaMask、pnpm、Warp、LiteLLM、Langflow 和 Open WebUI 中产生了 16 条维护者确认的漏洞公告。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05446 (HTTP 429)

Authors: Adib Hasan, Daniel Schaffield, Akashnil Dutta, Tarik Adnan Moon

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05446.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05446

Published: 2026-09-10T01:27:48.385Z

5. Damage-Aware Bandit Pruning for Vision and Language Transformers

Abstract:Structured post-training pruning of transformers requires selecting complete functional units whose suppression causes limited degradation. We formulate structured-unit selection for language and vision transformers as a damage-aware multi-armed bandit problem under a fixed candidate-evaluation budget. Attention heads and MLP channel groups are temporarily masked on calibration batches. Paired damage is the masked loss minus the base loss on the same batch, reducing batch-to-batch variation. A smooth bounded reward drives either a UCB-style policy or fractional-Beta Thompson Sampling, and the final mask is constructed sequentially by adding one unit at each step. The selected units are functionally zeroed in the original dense checkpoint; therefore, the reported parameter effects represent effective structural suppression rather than physical compression or measured speedup. Experiments on WikiText-2, LAMBADA, and Imagenette cover GPT-2, OPT, Pythia, Qwen2.5, SmolLM2, ViT-B/16, DeiT-Tiny, and Swin-Tiny, with comparisons against random, magnitude, static-saliency, and budgeted-greedy selection. Across five seeds, the bandit methods usually reduce degradation relative to budgeted greedy in the paired language-model comparisons. Of 28 comparisons highlighted in the paper, 23 bootstrap confidence intervals exclude zero and 11 paired tests have p < 0.05; six have q < 0.05 after Benjamini-Hochberg correction across the full family of 116 dataset-wise tests. Matched-evaluation results for ViT-B/16 and Swin-Tiny indicate that their gains are not explained solely by a larger candidate-evaluation budget.

中文摘要

摘要：对变换器进行结构化训练后剪枝需要选择完整的功能单元，其抑制会导致有限的性能下降。我们将语言和视觉变换器的结构单元选择问题表述为在固定候选评估预算下的损害感知多臂老虎机问题。在校准批次上临时屏蔽注意力头和MLP通道组。成对损害被定义为屏蔽损失减去同一批次的基础损失，从而减少批次间的波动。一个平滑有界的奖励驱动UCB风格策略或分数Beta汤普森采样，最终掩码通过每步添加一个单元逐步构建。在原始密集检查点中，所选单元功能上被置零；因此，报告的参数效果代表有效的结构抑制，而非物理压缩或测量的加速。针对WikiText-2、LAMBADA和Imagenette的数据集进行的实验涵盖了GPT-2、OPT、Pythia、Qwen2.5、SmolLM2、ViT-B/16、DeiT-Tiny和Swin-Tiny，并与随机选择、幅值选择、静态显著性选择以及预算贪心选择进行了对比。在五个随机种子的实验中，在成对语言模型比较中，老虎机方法通常相对于预算贪心方法减少了性能下降。本文重点的28个比较中，23个自助法置信区间排除了零，11个成对检验的p值小于0.05；在全体116个数据集维度检验的Benjamini-Hochberg校正后，六个q值小于0.05。ViT-B/16和Swin-Tiny的匹配评估结果表明，其收益不能仅用更大的候选评估预算来解释。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05448 (HTTP 429)

Authors: Salem Ameen, Sunil Vadera

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05448.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05448

Published: 2026-09-10T01:27:48.385Z

6. Compiling VGDL into Causal Models

Abstract:Reinforcement learning and large language models often struggle to accurately capture the causal mechanics of game environments. Standard reinforcement learning agents tend to rely on spurious correlations, while large language models are prone to hallucinating game rules. Although causal reinforcement learning improves interpretability, there is currently no formal methodology to map complex game mechanics directly into causal models. To address this, we propose a deterministic framework that compiles games specified in the Video Game Description Language into Dynamic Structural Causal Models. Rather than inferring causal structures from gameplay traces or noisy large language models’ outputs, our methodology directly translates game components, including sprite dynamics, interaction rules, and termination conditions, into explicit structural equations. Each game tick represents a causal transition from state variables at time $t$ to $t+1$. By establishing this grounded mapping, the approach guarantees absolute causal fidelity to the ground-truth game mechanics. The resulting models offer transparent causal pathways that support counterfactual reasoning, causal reinforcement learning agent training, and procedural content validation. This framework provides a principled bridge between symbolic game descriptions and causally grounded game AI.

中文摘要

摘要：强化学习和大型语言模型往往难以准确捕捉游戏环境的因果机制。标准强化学习代理倾向于依赖虚假的相关性，而大型语言模型容易对游戏规则产生幻觉。尽管因果强化学习提高了解释性，但目前尚无将复杂游戏机制直接映射到因果模型的正式方法。为了解决这一问题，我们提出了一种确定性框架，将用视频游戏描述语言（Video Game Description Language）指定的游戏编译为动态结构因果模型。我们的方法不是从游戏过程轨迹或噪声较多的大型语言模型输出中推断因果结构，而是直接将游戏组件，包括精灵动态、交互规则和终止条件，转化为明确的结构方程。每一游戏步都代表从时间 $t$ 到 $t+1$ 的状态变量因果转变。通过建立这种有根基的映射，该方法保证了与实际游戏机制的绝对因果一致性。生成的模型提供了透明的因果路径，支持反事实推理、因果强化学习代理训练以及程序化内容验证。该框架为符号性游戏描述与具有因果基础的游戏人工智能之间提供了一个原则性桥梁。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05459 (HTTP 429)

Authors: Mohit Jiwatode, Bodo Rosenhahn, Alexander Dockhorn

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05459.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05459

Published: 2026-09-10T01:27:48.385Z

7. ARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models

Abstract:Reward-free latent world models plan by scoring candidate actions with distances in a frozen latent space: an action is preferred if its predicted future embedding lands closer to the goal embedding. This silently assumes that latent closeness is action-rankable, i.e., that ordering candidates by latent distance agrees with ordering them by true cost. We audit this assumption directly. We introduce ARC-Bench, a no-leak, fixed-candidate protocol that measures whether frozen JEPA-style objectives rank candidate actions correctly, and apply it to official released JEPA-WM checkpoints across navigation and manipulation-style control. The assumption fails, severely and structurally: on the official manipulation audits the top-scored candidate is almost always suboptimal, and the same inversion appears in the maze domains. A controlled visual-backbone extension shows that the defect persists when DINOv2 is replaced by video-pretrained V-JEPA 1 and V-JEPA 2 encoders at ViT-L/ViT-G scale. Provenance, undertraining, matched-budget backbone controls, and metric-circularity controls rule out trivial explanations. We then explain why this defect has stayed invisible: closed-loop replanning masks it. When we reduce the planner’s replanning frequency, success collapses in both a navigation and a manipulation domain, and the episodes rescued by frequent replanning are enriched for severe first-plan ranking failures in the PointMaze first-plan diagnostic. Closed-loop success rates therefore systematically overstate the rankability of frozen latent representations. ARC-Bench supplies the measurement, and the masking mechanism the explanation, for methods that adapt, amortize, or replan around latent-space planners without directly auditing released JEPA-WM action rankability.

中文摘要

摘要：无奖励的潜在世界模型通过在冻结的潜在空间中使用距离对候选动作进行评分来进行规划：如果动作预测的未来嵌入更接近目标嵌入，则该动作被优先选择。这隐含地假设潜在空间的接近度是可以对动作进行排序的，即按潜在距离对候选动作排序与按真实成本排序一致。我们直接审查了这一假设。我们引入了ARC-Bench，一种无泄漏、固定候选协议，用于测量冻结的JEPA风格目标是否正确排序候选动作，并将其应用于导航和操作型控制任务中官方发布的JEPA-WM检查点。该假设失败，并且是严重且结构性的：在官方的操作审计中，得分最高的候选动作几乎总是不优的，这种反转现象在迷宫域中也出现。一个受控的视觉骨干扩展表明，当将DINOv2替换为视频预训练的V-JEPA 1和V-JEPA 2编码器（ViT-L/ViT-G规模）时，该缺陷仍然存在。来源、训练不足、匹配预算的骨干控制以及指标循环控制排除了简单解释。然后，我们解释了为什么这个缺陷一直未被发现：闭环重新规划掩盖了它。当我们降低规划器的重新规划频率时，无论在导航域还是操作域，成功率都会崩溃，而通过频繁重新规划拯救的实验集中表现出PointMaze首计划诊断中严重的首计划排序失败。因此，闭环成功率系统性地高估了冻结潜在表示的可排序性。ARC-Bench提供了测量手段，而掩盖机制则解释了那些围绕潜在空间规划器进行适应、分摊或重新规划的方法而没有直接审计发布的JEPA-WM动作可排序性的原因。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05461 (HTTP 429)

Authors: Zhengshu Zhang, Zhiyuan Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05461.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05461

Published: 2026-09-10T01:27:48.385Z

8. RAPID: Reliability-Aware Pair Importance Distillation

Abstract:Inter example relational distillation transfers a teacher’s representation geometry by matching relations among examples within a mini batch. Computing all pairs has quadratic complexity in the batch size, whereas uniform subsampling may use a limited relation budget inefficiently. We introduce Reliability Aware Pair Importance Distillation, or RAPID, which separates a reliability gated relational target from a full support adaptive pair proposal. Reliability determines which teacher relations are emphasized, while calibrated teacher entropy and detached student-teacher residuals determine which relations are evaluated. Exact inverse proposal correction makes the loss and gradient estimators conditionally unbiased with respect to the gated mini batch target. We evaluate RAPID in two text classification settings: AG News with BERT-to-DistilBERT distillation using three paired seeds and a relation budget of 256, and SST-2 with DistilBERT to DistilBERT distillation using three paired seeds and a relation budget of 64. Reliability gated relational distillation achieves the highest observed mean student accuracy on both datasets: 94.285 plus or minus 0.054 percent on AG News and 88.800 plus or minus 0.532 percent on SST-2. RAPID ranks second, achieving 94.241 plus or minus 0.025 percent and 88.685 plus or minus 0.462 percent, respectively, compared with 94.154 plus or minus 0.124 percent and 87.271 plus or minus 0.162 percent for the cross entropy baseline. Pilot evaluations are counted toward the same total budget as the main relation evaluations. Across both settings, the gated target yields the highest mean accuracy, while the adaptive proposal remains within seed-level variation. These results support the modular view that target reliability and evaluation priority are separable design dimensions.

中文摘要

摘要：抽象示例关系蒸馏通过匹配微批次内样本之间的关系，将教师的表示几何信息传递给学生模型。计算所有样本对的复杂度在批次大小上是二次的，而均匀子采样可能会低效地使用有限的关系预算。我们提出了可靠性感知的样本对重要性蒸馏（Reliability Aware Pair Importance Distillation，RAPID），它将可靠性门控的关系目标与全支持自适应样本对提议分离开来。可靠性决定了哪些教师关系被强调，而校准的教师熵和分离的学生-教师残差决定了哪些关系被评估。精确的逆向提议校正使损失和梯度估计器在条件上对门控微批次目标无偏。我们在两个文本分类场景中评估了RAPID：使用三组配对种子和256关系预算的AG News数据集上的BERT到DistilBERT蒸馏，以及使用三组配对种子和64关系预算的SST-2数据集上的DistilBERT到DistilBERT蒸馏。可靠性门控关系蒸馏在两个数据集上实现了观察到的最高学生模型平均准确率：AG News为94.285±0.054%，SST-2为88.800±0.532%。RAPID排名第二，分别实现94.241±0.025%和88.685±0.462%，而交叉熵基线分别为94.154±0.124%和87.271±0.162%。试点评估计入与主关系评估相同的总预算。在两个场景中，门控目标产生了最高的平均准确率，而自适应提议仍在种子级变化范围内。这些结果支持了模块化观点，即目标可靠性和评估优先级是可分离的设计维度。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05481 (HTTP 429)

Authors: Ali Mahdavi, Azadeh Zamanifar, Amirfarhad Farhadi, Omid Kashefi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05481.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05481

Published: 2026-09-10T01:27:48.385Z

9. PGP-Clinical-TimeKAN: Prior-Guided Joint Probabilistic Forecasting of Clinical Trajectories

Abstract:Clinical deterioration unfolds through coupled, partially observed trajectories, not a single diagnostic label. We introduce PGP-Clinical-TimeKAN, a trajectory-first framework for joint probabilistic forecasting of multivariate physiology. It combines missingness-aware temporal encoders, a soft organ-system prior, patient-specific relations, nonlinear Kolmogorov-Arnold messages, and a low-rank multivariate Student-t head. We evaluate 24-hour histories and six-hour forecasts on a frozen MIMIC-IV-derived cohort of 6,882 patients and 54,694 windows. Across five seeds and 13 models, PGP-Clinical-TimeKAN obtains the second-lowest normalized MAE (0.37727 +/- 0.00029) and the lowest RMSE (0.52656 +/- 0.00034). It reduces MAE by 0.52% relative to deterministic TimeKAN. For probabilistic forecasting, it reaches a marginal NLL of 0.66380 and a CRPS of 0.27301. Empirical coverage is 0.533, 0.831, and 0.958 for nominal 50%, 80%, and 95% intervals. Removing relational structure causes the largest ablation loss. Increasing covariance rank improves joint likelihood but has little effect on point accuracy. A trajectory-derived risk score remains weaker than a dedicated GRU-D classifier (AUROC 0.603 versus 0.650), which limits the present clinical claim. Joint trajectory forecasting therefore provides an inspectable intermediate task, but accurate physiology forecasts alone do not ensure a calibrated event detector.

中文摘要

摘要：临床恶化是通过耦合的、部分观察到的轨迹展开的，而不是单一的诊断标签。我们提出了PGP-Clinical-TimeKAN，这是一个以轨迹为先的多变量生理联合概率预测框架。它结合了对缺失敏感的时间编码器、软性器官系统先验、患者特异性关系、非线性Kolmogorov-Arnold消息以及低秩多变量Student-t头。我们在一个派生自MIMIC-IV的冻结队列上评估了24小时的历史记录和6小时的预测，该队列包含6,882名患者和54,694个时间窗口。在五个随机种子和13个模型中，PGP-Clinical-TimeKAN获得了第二低的标准化MAE（0.37727 +/- 0.00029）和最低的RMSE（0.52656 +/- 0.00034）。与确定性TimeKAN相比，它将MAE降低了0.52%。在概率预测方面，它达到了边际NLL 0.66380和CRPS 0.27301。经验覆盖率对于名义的50%、80%和95%区间分别为0.533、0.831和0.958。去除关系结构会导致最大的消融损失。增加协方差秩会提高联合似然，但对点预测精度几乎没有影响。基于轨迹的风险评分仍然弱于专门的GRU-D分类器（AUROC 0.603 对比 0.650），这限制了当前的临床声明。因此，联合轨迹预测提供了一个可检查的中间任务，但仅准确的生理预测并不能确保事件检测器的校准。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05488 (HTTP 429)

Authors: Weizhi Nie, Rihao Chang, Weijie Wang, Yuting Su

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05488.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05488

Published: 2026-09-10T01:27:48.385Z

10. SciLitBench: Benchmark and Design Principles for LLM-Powered Systematic Literature Reviews

Abstract:Systematic reviews require sustained human judgment across thousands of records, yet existing evaluations of large language models (LLMs) typically examine review stages in isolation. We introduce SciLitBench, a multi-stage benchmark spanning title and abstract screening, full-text screening, and schema-guided data extraction, with 42,981 retrieved records, 1,012 full texts, and annotations for 888 included papers. Across 22 open-weight LLMs from six model families, explicit inclusion and exclusion criteria improve title and abstract screening $F_2$ by 28.8\%, while researcher-authored rationales improve full-text screening by 15\%. Data extraction reveals a different reliability regime: performance declines from 0.97 accuracy for publication year to 0.37 Jaccard overlap for computational approach, while the strongest models recover only 30\% of annotated evaluation evidence and 25\% of limitations. SciLitBench identifies a practical boundary between high-recall screening and evidence-complete extraction and provides a reproducible resource for evaluating LLM-assisted evidence synthesis.

中文摘要

摘要：系统综述需要对数千条记录持续进行人工判断，而现有对大型语言模型（LLM）的评估通常只单独考察综述的各个阶段。我们引入了 SciLitBench，这是一个跨越标题与摘要筛选、全文筛选以及模式引导数据提取的多阶段基准，包含 42,981 条检索记录、1,012 篇全文以及 888 篇纳入论文的标注。在来自六个模型家族的 22 个开放权重 LLM 中，明确的纳入和排除标准使标题与摘要筛选的 $F_2$ 提升了 28.8%，而研究者撰写的理由则使全文筛选提升了 15%。数据提取表现出不同的可靠性状态：从发表年份的 0.97 精确度下降到计算方法的 0.37 Jaccard 重叠，而性能最强的模型仅恢复了 30% 的标注评估证据和 25% 的限制。SciLitBench 确定了高召回率筛选与证据完整提取之间的实际界限，并提供了一个可复现的资源用于评估 LLM 辅助的证据综合。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05505 (HTTP 429)

Authors: Miguel Zabaleta, Baihan Lin

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05505.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05505

Published: 2026-09-10T01:27:48.385Z

Evaluation Domain Papers

中文摘要

摘要：以往的AI对齐工作主要集中在一级社会规范——教模型什么是社会可接受或不可接受的行为（例如，“不要偷窃”）。然而，社会智力不仅依赖于规范识别，还依赖于预测谁会执行规范以及如何执行（例如，公开羞辱甚至监禁）。这些被称为元规范的二级期望，决定了人们在社会规则被违反时的反应。我们提出了一个评估大型语言模型（LLMs）元规范推理的新框架，通过两个维度：情绪评估和行为反应，并提出了新的分类任务，即预测违反者的自我调节行为，以及观察者的他人调节行为。我们发布了一个多视角数据集NormReact，包括450个规范违规场景，手工标注了违规者的性别和观察者的社会亲近度下的情绪和行为反应。当前的LLMs描绘了一个更严厉的社会世界：在六个模型中，它们在本应无动作的人类预期场景中过度预测负面制裁，并且随着社会距离的增加，与人类判断的一致性下降。这些发现表明，在从冲突调解到政策模拟等对规范敏感的领域，AI系统可能产生扭曲的社会规范执行图景：这种图景过度表现惩罚，而低估了宽容、克制和关系调控，这些特征实际上在现实世界的规范执行中更为常见。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05437 (HTTP 429)

Authors: Sunny Rai, Jinyi Kuang, Reyhan Jamalova, Annie Lou, Cristina Bicchieri, Niyati Malhotra, Victor Hugo Orozco-Olvera, Ana Maria Munoz-Boudet, Lyle H Ungar, Sharath C Guntuku

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05437.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05437

Published: 2026-09-10T01:28:08.954Z

2. CriticGen: Generation-Aware Evaluation as Actionable Feedback

中文摘要

摘要：当前大型语言模型的评估方法粗粒度且与生成过程脱节，产生的解释通常过于笼统，无法为模型改进提供可执行的反馈。我们提出了CriticGen，一种细粒度、生成感知的评估框架，将评估转化为可执行的答案改进控制。CriticGen首先在主观、客观和自主衍生约束等高层类别下，生成针对具体样本的评估维度和评分标准。随后，这些标准作为动态评分表，用于联合生成分数、原因、可执行的改进建议以及改进后的答案。这一基于评分表的改进过程使模型能够诊断缺陷并进行针对性的答案优化。实验结果表明，细粒度评估应既针对具体实例又具有可执行性。CriticGen引导出更高质量的评分表，使相关性/覆盖率从3.33/4.03提升至3.97/4.24。CriticGen还实现了最佳的分数相关性，皮尔逊相关系数为0.9556，斯皮尔曼相关系数为0.9560，同时将基于标准的理由和可执行建议的F1从0.6369/0.5994提升至0.7554/0.7900。关键是，其反馈能够转化为可靠的答案改进，提升了73.17%的答案，并保持了93.28%的非下降率。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05439 (HTTP 429)

Authors: Huifang Du, Zecheng Zuo, Sen Wang, Chenghao Fan, Haofen Wang, Yehui Yang

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05439.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05439

Published: 2026-09-10T01:28:08.954Z

3. When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

中文摘要

摘要：目前对大型语言模型（LLM）智能体的长期记忆评估主要通过对话回忆基准（LoCoMo, LongMemEval）进行，这些基准衡量的是对话历史的问答能力，而非记忆事实是否会改变使用工具的智能体的行为。我们提出了 MERIT（现实工具任务记忆评估），这是一个基准和框架，用于在明确成本核算下衡量记忆对执行任务智能体的边际效用。MERIT 提供三个领域的情景工具使用任务，其对早期情景事实的依赖性通过自动泄漏检查进行验证；一个以更新事实回忆为终点的难度阶梯；受控的记忆破坏；以及对每次记忆操作的完整代币和美元计量。在 23,440 个评分情景（$42.57）中，对 gpt-4.1-mini 进行的两代试点，以及预注册的 3 模型 x 3 种随机种子网格（GPT-4.1、Claude Haiku 4.5；记忆边保持固定），记忆将依赖任务的成功率从经过泄漏验证的 0.00 提升至 0.55-1.00。在更新事实方面，嵌入检索的表现无法预测地崩溃（各模型间为 0.30-0.95；最大种子差距为 0.45），且智能体仅在 55% 情况下根据正确检索的值采取行动，而写入时更新的存储（结构化事实存储，尤其是 LLM 摘要）维持在 0.70-1.00；混合方式的表现还不如单独的事实存储。最新一代的抽查（Claude Sonnet 5，在干净的完全重放控制下）复制了这一模式。更换记忆实现方式会使任务成功率变化高达 60 个点，且完全重放从未经济可行：每个领域的最佳条件的每美元边际效用为其 2.7-3.9 倍。我们发布了该基准、框架以及所有追踪数据。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05441 (HTTP 429)

Authors: Shweta Mishra, Shashank Mishra

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05441.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05441

Published: 2026-09-10T01:28:08.954Z

4. AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05446 (HTTP 429)

Authors: Adib Hasan, Daniel Schaffield, Akashnil Dutta, Tarik Adnan Moon

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05446.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05446

Published: 2026-09-10T01:28:08.954Z

5. Damage-Aware Bandit Pruning for Vision and Language Transformers

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05448 (HTTP 429)

Authors: Salem Ameen, Sunil Vadera

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05448.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05448

Published: 2026-09-10T01:28:08.954Z

6. Compiling VGDL into Causal Models

中文摘要

摘要：强化学习和大型语言模型往往难以准确捕捉游戏环境的因果机制。标准强化学习代理倾向于依赖虚假的相关性，而大型语言模型容易对游戏规则产生幻觉。尽管因果强化学习提高了解释性，但目前尚无将复杂游戏机制直接映射到因果模型的正式方法。为了解决这一问题，我们提出了一种确定性框架，将用视频游戏描述语言（Video Game Description Language）指定的游戏编译为动态结构因果模型。我们的方法不是从游戏过程轨迹或噪声较多的大型语言模型输出中推断因果结构，而是直接将游戏组件，包括精灵动态、交互规则和终止条件，转化为明确的结构方程。每一游戏步都代表从时间 $t$ 到 $t+1$ 的状态变量因果转变。通过建立这种有根映射，该方法保证了对真实游戏机制的绝对因果忠实性。生成的模型提供了透明的因果路径，支持反事实推理、因果强化学习代理训练以及程序化内容验证。该框架为符号化游戏描述与基于因果的游戏人工智能之间提供了一个有原则的桥梁。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05459 (HTTP 429)

Authors: Mohit Jiwatode, Bodo Rosenhahn, Alexander Dockhorn

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05459.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05459

Published: 2026-09-10T01:28:08.954Z

7. ARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05461 (HTTP 429)

Authors: Zhengshu Zhang, Zhiyuan Li

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05461.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05461

Published: 2026-09-10T01:28:08.954Z

8. RAPID: Reliability-Aware Pair Importance Distillation

中文摘要

摘要：抽象示例关系蒸馏通过匹配微型批次中示例之间的关系，将教师的表示几何结构传递给学生。计算所有成对关系在批量大小上具有二次复杂度，而均匀子采样可能会低效地使用有限的关系预算。我们提出了可靠性感知成对重要性蒸馏（RAPID），它将可靠性门控的关系目标与全面支持的自适应成对提议分离。可靠性决定了强调哪些教师关系，而校准的教师熵和分离的师生残差决定了评估哪些关系。精确的逆提议校正使损失和梯度估计在条件上对门控的微型批次目标保持无偏。我们在两个文本分类设置中评估RAPID：使用三组配对随机种子且关系预算为256的AG News数据集上的BERT到DistilBERT蒸馏，以及使用三组配对随机种子且关系预算为64的SST-2数据集上的DistilBERT到DistilBERT蒸馏。可靠性门控关系蒸馏在两个数据集上都实现了观察到的最高平均学生准确率：AG News为94.285±0.054%，SST-2为88.800±0.532%。RAPID排名第二，分别实现94.241±0.025%和88.685±0.462%，而交叉熵基线分别为94.154±0.124%和87.271±0.162%。试点评估被计入与主关系评估相同的总预算。在两个设置中，门控目标产生了最高平均准确率，而自适应提议保持在种子级差异范围内。这些结果支持模块化观点：目标可靠性和评估优先级是可分离的设计维度。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05481 (HTTP 429)

Authors: Ali Mahdavi, Azadeh Zamanifar, Amirfarhad Farhadi, Omid Kashefi

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05481.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05481

Published: 2026-09-10T01:28:08.954Z

9. PGP-Clinical-TimeKAN: Prior-Guided Joint Probabilistic Forecasting of Clinical Trajectories

中文摘要

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05488 (HTTP 429)

Authors: Weizhi Nie, Rihao Chang, Weijie Wang, Yuting Su

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05488.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05488

Published: 2026-09-10T01:28:08.954Z

10. SciLitBench: Benchmark and Design Principles for LLM-Powered Systematic Literature Reviews

中文摘要

摘要：系统综述需要对数千条记录持续进行人工判断，而现有对大型语言模型（LLM）的评估通常只单独考察综述的各个阶段。我们引入了 SciLitBench，这是一个跨越标题与摘要筛选、全文筛选以及模式引导数据提取的多阶段基准，涵盖 42,981 条检索记录、1,012 篇全文以及 888 篇纳入论文的标注。在来自六个模型家族的 22 个开放权重 LLM 中，明确的纳入和排除标准使标题与摘要筛选的 $F_2$ 提升了 28.8%，而研究人员撰写的理由使全文筛选提升了 15%。数据提取表现出不同的可靠性状态：从发表年份的 0.97 精确度下降到计算方法的 0.37 Jaccard 重叠，而性能最强的模型仅恢复了 30% 的标注评估证据和 25% 的限制。SciLitBench 确定了高召回率筛选与证据完整提取之间的实际界限，并提供了一个可复现的资源用于评估 LLM 辅助的证据综合。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05505 (HTTP 429)

Authors: Miguel Zabaleta, Baihan Lin

Categories: cs.AI

PDF URL: https://arxiv.org/pdf/2609.05505.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05505

Published: 2026-09-10T01:28:08.954Z

VLM Domain Papers

1. Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

Abstract:Deep learning registration methods routinely stack two kinds of enhancement on a base network: architectural additions such as affine pre-alignment stages, and training-objective additions such as regularization losses. Papers tend to adopt both at once, so it is unclear which is doing the work. I ran a controlled ablation to separate them. Using the OASIS brain MRI dataset (394 training subjects, 20 test subjects), I trained four variants of the same registration pipeline: a baseline 3D U-Net with basic similarity losses, the same U-Net with a full regularization suite, an affine-plus-deformable architecture with basic losses, and the affine architecture with the full suite. I evaluated registration accuracy (MSE, NCC, SSIM), deformation quality (Jacobian determinant preservation, displacement statistics, an anatomical plausibility score), and computational cost. Regularization alone accounted for most of the gain: a 21.3% relative gain on the MSE-improvement metric (1.78% to 2.16%, P<.001) and a 21.8% relative gain in NCC improvement, while cutting maximum deformation from 53.1 to 0.51 units, a 99.0% reduction, at essentially no computational cost (-0.06% inference time). The combined model produced the largest accuracy gain, 25.8% (1.78% to 2.24%), and raised anatomical plausibility from 0.596 to 0.930, at a moderate +9.8% inference-time cost. Gradient correlation rose from 0.742 at baseline to 0.980 for the fully enhanced model. All enhanced variants reached sub-voxel accuracy under plausible deformation constraints. Regularization losses are the primary driver in this setting, delivering the accuracy gains and almost all of the deformation control for free at inference time, while the affine architecture adds a smaller complementary benefit at acceptable cost. The 99% reduction in unrealistic deformations addresses a known barrier to clinical deployment.

中文摘要

摘要：深度学习配准方法通常在基础网络上叠加两种增强：架构上的补充（如仿射预对齐阶段）和训练目标上的补充（如正则化损失）。论文往往同时采用两者，因此无法明确哪一部分真正起作用。我进行了一个受控消融实验来将它们区分开来。使用OASIS脑部MRI数据集（394名训练对象，20名测试对象），我训练了同一配准管线的四个变体：一个带有基本相似性损失的基线3D U-Net，带有完整正则化套件的同一U-Net，带有基本损失的仿射加可变形结构，以及带有完整套件的仿射结构。我评估了配准精度（MSE, NCC, SSIM）、变形质量（雅可比行列式保持、位移统计、解剖学合理性评分）以及计算成本。仅正则化就占据了大部分增益：在MSE改进指标上相对增益为21.3%（从1.78%提升到2.16%，P<.001），NCC改进的相对增益为21.8%，同时将最大变形从53.1单位降到0.51单位，减少了99.0%，而几乎没有计算成本（推理时间减少0.06%）。联合模型产生了最大的精度提升，为25.8%（从1.78%提升到2.24%），并将解剖学合理性从0.596提高到0.930，推理时间成本中等，增加了9.8%。梯度相关性从基线的0.742上升到完全增强模型的0.980。所有增强变体在合理变形约束下均达到了亚体素精度。正则化损失是在这种设置下的主要驱动因素，在推理阶段几乎免费地提供了精度提升和几乎所有的变形控制，而仿射架构则以可接受的成本提供了较小的互补性收益。99%的不现实变形减少解决了临床应用的已知障碍。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05484 (HTTP 429)

Authors: Nabira Rashid

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05484.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05484

Published: 2026-09-10T01:28:29.052Z

2. A Survey on Adversarial Attacks and Defenses for Diffusion Models Across Multiple Modalities

Abstract:Diffusion models have become the dominant family of generative models in the visual domain. However, their widespread public availability enables misuse at scale, motivating a rapidly growing body of research on adversarial attacks and defenses. This survey provides, to our knowledge, the first unified review of this literature across three visual modalities: image, video, and 3D. We introduce a comprehensive, task-centric taxonomy: we first divide the literature by modality; within each modality, we separate methods into attacks and defenses, and then group them by the generative task they target, presenting them chronologically within each task. Moreover, we provide an in-depth analysis of their evaluation settings, consolidating the datasets, metrics, and benchmarks used to assess them. We conclude by identifying several open challenges and outlining concrete future research directions. Project Webpage: this https URL

中文摘要

摘要：扩散模型已经成为视觉领域中主流的生成模型家族。然而，它们的广泛公开可用性使得大规模滥用成为可能，从而推动了关于对抗性攻击与防御的快速增长的研究。根据我们的了解，本综述首次提供了跨三种视觉模态（图像、视频和三维）的统一文献回顾。我们引入了一个全面、以任务为中心的分类法：首先按模态划分文献；在每种模态中，将方法分为攻击和防御，然后按它们针对的生成任务进行分组，并在每个任务内按时间顺序呈现。此外，我们提供了对其评估设置的深入分析，整合了用于评估它们的数据集、指标和基准测试。最后，我们指出了若干未解决的挑战，并概述了具体的未来研究方向。项目网页：这个 https URL

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05503 (HTTP 429)

Authors: Ozgur Kara, Tarik Can Ozden, Furkan Horoz, Zeqian Long, Haotian Xue, Yipu Chen, Oguzhan Akcin, Yongxin Chen, James Matthew Rehg

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05503.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05503

Published: 2026-09-10T01:28:29.052Z

3. An Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception

Abstract:Unified perception enables autonomous driving systems to perform object detection, drivable-area segmentation, and lane segmentation within a single network, improving efficiency and reducing deployment complexity. Jointly optimizing multiple perception tasks remains challenging because tasks exhibit different convergence rates, loss scales, and optimization stability. Existing task-weighting methods use loss magnitude, learned uncertainty, short-term loss changes, or gradient statistics; here, we explore the frequency structure of a recent loss-history window as a complementary signal. We implement and examine Frequency-aware Task Weighting (FTW), a dynamic task-balancing rule that estimates a loss-trajectory stability proxy from the low-frequency energy ratio of recent loss histories. FTW assigns larger weights to tasks whose mean-centered loss trajectories contain a larger proportion of low-frequency power. We document FTW and two baselines under full-network static training and progressive freezing using a unified YOLOv8-based perception framework with three task-specific heads. Experiments on Mapillary Vistas compare FTW with fixed and uncertainty-based weighting under both configurations. Final holdout metrics are reported for the checkpoint with the lowest per-epoch validation loss in each run. Across six single-run configurations, static FTW has the largest derived overall score and lane mIoU, progressive FTW has the largest detection mAP, and static uncertainty weighting has the largest drivable-area mIoU. Without repeated-seed estimates, single-task baselines, or FTW ablations, these rankings are descriptive. The evidence supports the feasibility of loss-frequency-based weighting in this pipeline, but does not establish improvement over the baselines or generalization beyond the reported runs.

中文摘要

摘要：统一感知使自动驾驶系统能够在单个网络中执行物体检测、可行驶区域分割和车道分割，从而提高效率并减少部署复杂性。联合优化多个感知任务仍然具有挑战性，因为各任务具有不同的收敛速度、损失尺度和优化稳定性。现有的任务加权方法使用损失大小、学习到的不确定性、短期损失变化或梯度统计；在这里，我们探索了近期损失历史窗口的频率结构，作为一种补充信号。我们实现并研究了频率感知任务加权（FTW），这是一种动态任务平衡规则，通过近期损失历史的低频能量比率来估计损失轨迹稳定性代理。FTW 给损失轨迹中低频功率比例较大的任务分配更大的权重。我们在一个基于 YOLOv8 的统一感知框架中对 FTW 和两个基线方法进行了全网络静态训练和渐进冻结的实验，这个框架包含三个任务专用头。在 Mapillary Vistas 数据集上的实验比较了 FTW 与固定加权和基于不确定性的加权在两种配置下的表现。每次运行中，最终保留的指标来自每轮验证损失最低的检查点。在六种单次运行配置中，静态 FTW 拥有最高的综合得分和车道 mIoU，渐进 FTW 拥有最高的检测 mAP，而静态不确定性加权拥有最高的可行驶区域 mIoU。在没有重复随机种子估计、单任务基线或 FTW 消融实验的情况下，这些排名仅具描述性。证据支持基于损失频率的加权在此流程中是可行的，但并未证明其优于基线方法或在报告的运行之外具有泛化能力。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05516 (HTTP 429)

Authors: Zhiyuan Nie, Zixi Zhou, Xianbin Gu

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05516.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05516

Published: 2026-09-10T01:28:29.052Z

4. Emergent Goal-Directed Attention in Large Vision-Language Models

Abstract:Human observers prioritize visual information according to task goals. Most computational models of naturalistic viewing are gaze-trained for free viewing, leaving open whether goal-directed attention can emerge in systems without gaze supervision. We tested two off-the-shelf vision-language models (VLMs), Qwen3-VL-32B-Thinking and Gemma-4-26B-A4B-it, on 4,887 naturalistic scenes under visual-search and free-viewing instructions. Model predictions were compared with human fixations on the same images under corresponding tasks. Both models aligned more closely with human fixations under matching goals than under mismatched goals. This crossover persisted in target-absent scenes, where alignment could not be explained by simple visual grounding, and appeared in decoder-layer readouts. Furthermore, model-thinking traces were grounded in target semantics during search and in visual prominence during free viewing. These findings show that general-purpose VLMs can generate human-aligned, goal-directed spatial priorities without gaze-specific training, informing theories of goal-directed attention and offering scalable tools for predicting where people look across tasks.

中文摘要

摘要：人类观察者会根据任务目标优先处理视觉信息。大多数自然观看的计算模型是在自由观看的条件下进行注视训练的，因此尚不清楚在没有注视监督的系统中，是否能够出现目标导向的注意力。我们在4,887个自然场景下测试了两种现成的视觉-语言模型（VLMs），Qwen3-VL-32B-Thinking和Gemma-4-26B-A4B-it，分别在视觉搜索和自由观看指令下进行实验。模型预测结果与同一图像在人类对应任务下的注视点进行了比较。两种模型在目标匹配的情况下与人类注视点的对齐度高于目标不匹配的情况。这种交叉现象在目标缺失的场景中仍然存在，此时对齐度无法通过简单的视觉联系来解释，并出现在解码层的读取结果中。此外，模型思维轨迹在搜索任务中基于目标语义，在自由观看中则基于视觉显著性。研究结果表明，通用视觉-语言模型能够在无需特定注视训练的情况下生成与人类一致的目标导向空间优先级，这为目标导向注意力理论提供了参考，并为预测人们在不同任务中的注视位置提供了可扩展的工具。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05517 (HTTP 429)

Authors: Han Zhang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05517.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05517

Published: 2026-09-10T01:28:29.052Z

Abstract:Despite the strong capabilities of multimodal large language models (MLLMs), their parametric knowledge remains incomplete and difficult to update, motivating multimodal retrieval-augmented generation (RAG) to ground responses in external text and images. However, existing benchmarks face two major limitations: (i) they typically emphasize single-hop retrieval or reasoning over a small set of provided contexts rather than open-domain evidence discovery; and (ii) they provide fragmented coverage of cross-modal reasoning paths, leaving complex multi-hop and multi-image reasoning underexplored. In this paper, we introduce CrossModalQA, an open-domain benchmark for evaluating multimodal retrieval and reasoning over heterogeneous corpora. CrossModalQA contains 1,863 question-answer pairs constructed from 4,987 Wikipedia articles and 4,431 Wikimedia Commons images. It covers five complementary reasoning paths: vision-to-text, text-to-vision, vision-to-text-to-vision, multi-image intersection, and image-set reasoning. Every question requires retrieving and composing distributed textual and visual evidence, with an average reasoning depth of 3.50 hops. We construct the benchmark through multimodal knowledge graph-guided subgraph sampling and apply rule-based consistency checking and LLM verification to ensure multimodal dependence and traceable evidence. Extensive experiments demonstrate that existing multimodal RAG systems struggle to recover complete evidence chains and can underperform closed-book models when incomplete retrieval introduces distracting context. Further analysis reveals that complete cross-modal retrieval contributes more to answer accuracy than generator scaling, while multi-image retrieval and reasoning remain the primary bottlenecks limiting end-to-end performance.

中文摘要

摘要：尽管多模态大语言模型（MLLMs）具备强大的能力，但其参数化知识仍不完整且难以更新，这促使多模态检索增强生成（RAG）方法被提出，用于在外部文本和图像中获取响应依据。然而，现有基准存在两个主要局限性：（i）它们通常侧重于单跳检索或在提供的小规模上下文集上进行推理，而非开放域证据发现；（ii）它们对跨模态推理路径的覆盖零散，使复杂的多跳和多图像推理尚未充分探索。本文提出 CrossModalQA，一个用于评估异构语料库上多模态检索与推理的开放域基准。CrossModalQA 包含 1,863 个问答对，由 4,987 篇维基百科文章和 4,431 张维基共享资源图片构建。它涵盖五种互补的推理路径：视觉到文本、文本到视觉、视觉到文本再到视觉、多图像交集以及图像集推理。每个问题均需要检索和组合分布式的文本与视觉证据，平均推理深度为 3.50 跳。我们通过多模态知识图引导的子图采样构建该基准，并应用基于规则的一致性检查及大语言模型验证，以确保多模态依赖性和可追踪证据。大量实验表明，现有多模态 RAG 系统难以恢复完整的证据链，当不完整的检索引入干扰上下文时，甚至可能表现得不如闭卷模型。进一步分析显示，完整的跨模态检索对答案准确性贡献大于生成器规模的扩大，而多图像检索与推理仍然是限制端到端性能的主要瓶颈。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05518 (HTTP 429)

Authors: Jiacheng Cai, Zijin Hong, Zheng Yuan, Huachi Zhou, Qinggang Zhang, Xiao Huang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05518.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05518

Published: 2026-09-10T01:28:29.052Z

6. Contrastive Knowledge Distillation for Anomaly Detection in Multi-Illumination/Focus Display Images

Abstract:In this paper, we tackle automatic anomaly detection in multi-illumination and multi-focus display images. The minute defects on the display surface are hard to spot out in RGB images and by a model trained with only normal data. To address this, we propose a novel contrastive learning scheme for knowledge distillation-based anomaly detection. In our framework, Multiresolution Knowledge Distillation (MKD) is adopted as a baseline, which operates by measuring feature similarities between the teacher and student networks. Based on MKD, we propose a novel contrastive learning method, namely Multiresolution Contrastive Distillation (MCD), which does not require positive/negative pairs with an anchor but operates by pulling/pushing the distance between the teacher and student features. Furthermore, we propose the blending module that transforms and aggregate multi-channel information to the three-channel input layer of MCD. Our proposed method significantly outperforms competitive state-of-the-art methods in both AUROC and accuracy metrics on the collected Multi-illumination and Multi-focus display image dataset for Anomaly Detection (MMdAD).

中文摘要

摘要：在本文中，我们研究了多光照和多焦点显示图像的自动异常检测。显示表面上的微小缺陷在RGB图像中以及仅使用正常数据训练的模型中都很难被发现。为了解决这一问题，我们提出了一种基于知识蒸馏的异常检测新型对比学习方案。在我们的框架中，采用多分辨率知识蒸馏（Multiresolution Knowledge Distillation, MKD）作为基线，其通过测量教师网络和学生网络之间的特征相似性来工作。在MKD的基础上，我们提出了一种新型对比学习方法，即多分辨率对比蒸馏（Multiresolution Contrastive Distillation, MCD），该方法不需要带锚点的正/负样本对，而是通过拉近/推开教师和学生特征之间的距离来操作。此外，我们提出了融合模块，将多通道信息转换并汇聚到MCD的三通道输入层中。我们提出的方法在收集的多光照与多焦点显示图像异常检测数据集（Multi-illumination and Multi-focus display image dataset for Anomaly Detection, MMdAD）上，在AUROC和准确率指标上均显著优于现有的竞争性最先进方法。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05520 (HTTP 429)

Authors: Jihyun Lee, Hangil Park, Yongmin Seo, Taewon Min, Joodong Yun, Jaewon Kim, Tae-Kyun Kim

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05520.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05520

Published: 2026-09-10T01:28:29.052Z

7. Diffusion models for eye-gaze trajectory generation using position and velocity representations

Abstract:Eye-tracking data are expensive to collect, requiring specialized hardware and controlled laboratory conditions, and difficult to share because of privacy constraints. We address this using two complementary denoising diffusion probabilistic models (DDPMs) for unconditional generation of eye-gaze dynamics from visual-search data. Both use an identical FiLM-conditioned one-dimensional U-Net with self-attention (19.35,M parameters), trained on 8,s sliding-window sequences from 28 participants. One model generates raw two-dimensional gaze-position sequences, while the other generates two-component velocity sequences; each uses representation-specific preprocessing, training settings, data partitions, and evaluation protocols. Both are evaluated across three independent training seeds, with aggregated metrics reported as mean,$\pm$,SD. The position-space model achieves a mean Jensen-Shannon (JS) divergence of $0.016\pm0.004$ across nine kinematic features, with the highest feature-wise mean below $0.030$, fixation duration within 2% of real data, and a Fr’echet Gaze Distance more than an order of magnitude below statistical and Markovian baselines. Under a Train-on-Synthetic-Test-on-Real protocol, synthetic-only training achieves $R^2=0.66\pm0.02$, or 82.7% of the real-data $R^2$ point estimate. The velocity-space model achieves a mean JS divergence of $0.0065$ across velocity components, speed, log-speed, and turning angle, with a maximum of $0.015\pm0.005$. Reconstructed path length is less accurate ($0.21\pm0.02$ versus $0.03\pm0.01$ in position space), although the protocols differ. Overall, unconditional diffusion captures local gaze kinematics and short-range temporal and directional structure, while long-range properties such as saccade counts and cumulative path geometry remain targets for future conditioned models.

中文摘要

摘要：眼动追踪数据的收集成本高，需要专用硬件和受控实验室环境，并且由于隐私限制难以共享。我们通过使用两种互补的去噪扩散概率模型（DDPMs）来解决这个问题，以便从视觉搜索数据中无条件生成眼动动态。两者都使用相同的FiLM条件一维U-Net并配备自注意力机制（19.35M参数），在28名参与者的8秒滑动窗口序列上进行训练。其中一个模型生成二维原始注视位置序列，而另一个生成两分量速度序列；每个模型使用特定于表示的预处理、训练设置、数据划分和评估协议。两者均在三个独立训练随机种子上进行评估，并报告聚合指标的均值±标准差。位置空间模型在九个运动学特征上的平均Jensen-Shannon（JS）散度为0.016±0.004，各特征平均值最高低于0.030，注视持续时间与真实数据相差在2%以内，Fréchet注视距离比统计和马尔可夫基线低一个数量级以上。在“以合成训练、以真实测试”协议下，仅合成训练实现R²=0.66±0.02，相当于真实数据R²点估计的82.7%。速度空间模型在速度分量、速度、对数速度和转向角上的平均JS散度为0.0065，最大为0.015±0.005。重建路径长度准确性较低（0.21±0.02，而位置空间为0.03±0.01），尽管两者协议不同。总体而言，无条件扩散能够捕捉局部注视运动学以及短程时间和方向结构，而长程特性如扫视次数和累计路径几何仍是未来条件模型的目标。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05522 (HTTP 429)

Authors: Laxman Basnet, Alexander Szorkovszky, Pedro G. Lind, Anis Yazidi, Shailendra Bhandari

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05522.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05522

Published: 2026-09-10T01:28:29.052Z

8. Infrastructure-based Monocular 3D Vehicle Localization Framework with Experimental Validation

Abstract:This paper presents a one-stage learning framework that maps monocular roadside-camera images directly to vehicle states in a ground-fixed coordinate frame. Unlike conventional approaches that first detect vehicles in the image plane and subsequently apply geometric post-processing, the proposed method leverages features from a pretrained object detector to jointly estimate each vehicle’s ground-plane position, dimensions, and yaw angle. The framework therefore uses visual features not only for vehicle detection but also for direct spatial and orientation estimation. To support model training and evaluation, we develop a data-collection and label-generation pipeline based on synchronized video from a roadside camera and an unmanned aerial vehicle (UAV). Acting as a temporary top-view sensing platform, the UAV provides vehicle trajectories, dimensions, and orientations, which are transformed into the ground-fixed coordinate frame and temporally aligned with the roadside-camera images to generate ground-truth labels. The framework is evaluated using data collected during multiple experiments at the Mcity Test Facility. Results show that the proposed method can recover vehicle trajectories and orientations from monocular roadside imagery without a separate geometric post-processing stage, demonstrating its potential as a scalable approach to infrastructure-based perception at urban intersections.

中文摘要

摘要：本文提出了一种单阶段学习框架，将单目路边摄像机图像直接映射到地面固定坐标系中的车辆状态。与传统方法需要先在图像平面上检测车辆，然后再进行几何后处理不同，所提出的方法利用预训练目标检测器的特征，联合估计每辆车的地面平面位置、尺寸和偏航角。因此，该框架不仅使用视觉特征进行车辆检测，还直接用于空间位置和方向估计。为了支持模型训练和评估，我们开发了一个数据采集和标签生成流程，基于路边摄像机与无人机（UAV）同步拍摄的视频。作为临时的俯视感知平台，无人机提供车辆轨迹、尺寸和方向，这些数据被转换到地面固定坐标系，并与路边摄像机图像在时间上对齐，以生成真实标签。该框架使用在Mcity测试设施进行的多次实验中收集的数据进行了评估。结果表明，所提出的方法能够从单目路边图像中恢复车辆轨迹和方向，无需单独的几何后处理阶段，展示了其在城市路口基于基础设施的感知中作为可扩展方法的潜力。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05523 (HTTP 429)

Authors: Akos T. Kopeczi-Bocz, Tian Mi, Gabor Orosz, Denes Takacs

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05523.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05523

Published: 2026-09-10T01:28:29.052Z

9. DIVA: Exploiting Cross-Step Conditional Propagation for Visual Jailbreaks in Discrete Diffusion Vision-Language Models

Abstract:Large vision-language models (VLMs) are increasingly deployed in safety-critical settings, yet existing visual jailbreak research has focused almost exclusively on autoregressive architectures, leaving an important emerging family unstudied: multimodal discrete diffusion vision-language models (dVLMs). We identify a vulnerability specific to diffusion generation: because the visual embedding conditions every reverse denoising step rather than acting as a one-time prefix, adversarial visual semantics are repeatedly propagated and amplified across the generation trajectory, a phenomenon we term cross-step conditional propagation. We provide empirical evidence through stage-sensitivity analysis, prompt-level switch rates, and pairwise denoising-bin disagreement metrics, confirmed by bootstrap resampling. We propose DIVA (Discrete-diffusion Vision-language model Attack), a white-box visual jailbreak framework using cross-modal intent obfuscation and diffusion-aware multi-timestep adversarial optimization. Across three dVLMs, DIVA reaches 58.8%, 67.7%, and 69.1% HADES ASR under the Beaver reward-model metric, outperforming visual jailbreak baselines designed for autoregressive models. Code: this https URL

中文摘要

摘要：大型视觉-语言模型（VLMs）正越来越多地部署在安全关键的环境中，然而现有的视觉越狱研究几乎完全集中在自回归架构上，留下了一个重要的新兴类别未被研究：多模态离散扩散视觉-语言模型（dVLMs）。我们发现了一种特定于扩散生成的漏洞：由于视觉嵌入条件作用于每一步反向去噪，而不是仅作为一次性前缀，对抗性视觉语义会在生成轨迹中重复传播并被放大，这种现象我们称之为跨步骤条件传播。我们通过阶段敏感性分析、提示级切换率以及成对去噪-bin不一致性指标提供了实证证据，并通过自助重采样方法进行了验证。我们提出了DIVA（离散扩散视觉-语言模型攻击），这是一个白盒视觉越狱框架，利用跨模态意图混淆和扩散感知多步对抗优化。在三个dVLMs上，DIVA在Beaver奖励模型指标下达到了58.8%、67.7%和69.1%的HADES ASR，优于为自回归模型设计的视觉越狱基线。代码：此 https URL

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05525 (HTTP 429)

Authors: Guorui Song, Runqing Tang, Jingye Zhang, Luyuan Zhang, Feice Huang, Cong Ray, Guocun Wang, Dake Zhong, Choo Sin Wai, Bingquan Dai, Chuming Wang, Tongxu Lin, Wanyu Guo, Haoqian Wang

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05525.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05525

Published: 2026-09-10T01:28:29.052Z

10. A Specialized Large Multimodal Model for Interpreting PET/CT in Head and Neck Cancer

Abstract:Background: Diagnosing head and neck cancer using PET/CT is clinically challenging and time-consuming due to the anatomical complexity of the region, motivating computer-aided diagnosis (CAD). Generalist Large Multimodal Models (LMMs) remain limited in medical contexts by insufficient domain-specific knowledge, privacy and security concerns, and verbosity, motivating specialized standalone LMMs. Purpose: We evaluated the feasibility of a specialized LMM for automated PET/CT interpretation in head and neck cancer using a large-scale multi-institutional PET/CT dataset, a tailored training curriculum, and autoregressive training. Methods: LLaVA-NeXT was fine-tuned using a two-level curriculum with image-conversation pairs curated by two radiologists from public data. The dataset included clinically important annotations such as primary tumor presence and metastatic lymph node location. Level 1 used 28,000 image-conversation pairs to learn basic information, including modality type and hypermetabolism. Level 2 used 12,975 pairs to learn primary tumor presence and the existence and anatomical location of cervical lymph node metastases. External validation included four institutions with diverse imaging devices. Results: The specialized LMM substantially outperformed ChatGPT and LLaVA-NeXT. In Level-2 external validation, ROUGE-L, ROUGE-S, Cosine Similarity, Precision, Recall, and F1 were 0.8751, 0.8794, 0.8324, 0.8794, 0.8711, and 0.8751, while generalist models consistently scored below 0.1. Primary tumor classification accuracy was 83.14 +/- 1.15% internally and 69.03 +/- 0.81% externally. For lymph node localization, the corresponding scores were 0.6389, 0.6257, 0.5287, 0.5782, 0.6371, and 0.6648. Conclusion: Specialized LMMs show promising results for fast, accurate PET/CT-based diagnostic support and medical education, highlighting their potential for clinical translation.

中文摘要

摘要：
背景：由于头颈部区域的解剖结构复杂，使用PET/CT诊断头颈癌在临床上具有挑战性且耗时，这推动了计算机辅助诊断（CAD）的发展。通用大型多模态模型（LMMs）在医疗环境中仍受限于行业特定知识不足、隐私和安全问题以及冗长性，这促使研究者开发专门的独立LMMs。
目的：我们评估了一种专门LMM在头颈癌PET/CT自动解读中的可行性，使用了大型多机构PET/CT数据集、定制化训练课程以及自回归训练的方法。
方法：LLaVA-NeXT通过两阶段课程进行微调，使用由两名放射科医生从公开数据中策划的图像-会话对。数据集包含临床重要标注，如原发肿瘤存在性及转移性淋巴结位置。第一阶段使用28,000个图像-会话对学习基础信息，包括成像模态类型和高代谢区。第二阶段使用12,975个图像-会话对学习原发肿瘤存在性及颈部淋巴结转移的存在与解剖位置。外部验证涵盖四个拥有不同成像设备的机构。
结果：该专门LMM显著优于ChatGPT和LLaVA-NeXT。在第二阶段外部验证中，ROUGE-L、ROUGE-S、余弦相似度、精确率、召回率和F1分别为0.8751、0.8794、0.8324、0.8794、0.8711和0.8751，而通用模型得分一直低于0.1。原发肿瘤分类准确率在内部验证中为83.14 +/- 1.15%，在外部验证中为69.03 +/- 0.81%。对于淋巴结定位，相应的分数为0.6389、0.6257、0.5287、0.5782、0.6371和0.6648。
结论：专门LMM在快速、准确的PET/CT诊断支持和医学教育中显示出有前景的结果，突显了其在临床转化中的潜力。

LLM Analysis

LLM Analysis Failed: Error: Kimi内容未就绪或抓取失败: 2609.05532 (HTTP 429)

Authors: Haengbok Chung, SunGyu Kim, Joo hyun Lee, Sangjin Bae, Min Jeong Cho, Minseok Suh, Jae Sung Lee

Categories: cs.CV

PDF URL: https://arxiv.org/pdf/2609.05532.pdf

CoolPaper URL: https://papers.cool/arxiv/2609.05532

Published: 2026-09-10T01:28:29.052Z

ArXiv Domain 2026-09-10

LLM Domain Papers

1. When Agent Governance Helps

2. TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

3. Better Together: Complementary Query Rewriting Under a Strong RAG Baseline

4. Who Maintains Agent Skills? A Longitudinal Study of Human-Governed, AI-Assisted Skill Maintenance

5. A Rubric-Guided Large Language Model Solution for Opioid Use Disorder Computable Phenotyping

6. Intra-Prompt Parallel Decoding for Common-Context Question Answering

7. MedWER: A Reproducible, Model-Free Evaluation Protocol for Medical Speech Recognition

8. Some Tokens Behave like Magnets: Revealing Linguistic Organization in the Layers of Language Models

9. CrisisKD: Five-Stage Knowledge Distillation for Aspect-Level Sentiment and Emotion Analysis in Crisis Discourse

10. Recall Is Not Protection: Evaluating Safety Monitors Against Model Compliance

Agent Domain Papers

1. Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models

2. CriticGen: Generation-Aware Evaluation as Actionable Feedback

3. When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

4. AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

5. Damage-Aware Bandit Pruning for Vision and Language Transformers

6. Compiling VGDL into Causal Models

7. ARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models

8. RAPID: Reliability-Aware Pair Importance Distillation

9. PGP-Clinical-TimeKAN: Prior-Guided Joint Probabilistic Forecasting of Clinical Trajectories

10. SciLitBench: Benchmark and Design Principles for LLM-Powered Systematic Literature Reviews

Evaluation Domain Papers

1. Beyond Right and Wrong: Evaluating Second-order Social Reasoning in Large Language Models

2. CriticGen: Generation-Aware Evaluation as Actionable Feedback

3. When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

4. AutoFyn Technical Report: Non-Parametric Expert Iteration for Long-Horizon Agents

5. Damage-Aware Bandit Pruning for Vision and Language Transformers

6. Compiling VGDL into Causal Models

7. ARC-Bench: Closed-Loop Replanning Masks Broken Action Ranking in Frozen JEPA World Models

8. RAPID: Reliability-Aware Pair Importance Distillation

9. PGP-Clinical-TimeKAN: Prior-Guided Joint Probabilistic Forecasting of Clinical Trajectories

10. SciLitBench: Benchmark and Design Principles for LLM-Powered Systematic Literature Reviews

VLM Domain Papers

1. Architectural and Regularization Components in Deep Learning Medical Image Registration: Systematic Ablation Study

2. A Survey on Adversarial Attacks and Defenses for Diffusion Models Across Multiple Modalities

3. An Exploratory Study of Frequency-Aware Task Weighting for YOLOv8-Based Unified Driving Perception

4. Emergent Goal-Directed Attention in Large Vision-Language Models

5. CrossModalQA: A Cross-modal and Multi-hop Benchmark for Multimodal Retrieval-augmented Generation

6. Contrastive Knowledge Distillation for Anomaly Detection in Multi-Illumination/Focus Display Images

7. Diffusion models for eye-gaze trajectory generation using position and velocity representations

8. Infrastructure-based Monocular 3D Vehicle Localization Framework with Experimental Validation

9. DIVA: Exploiting Cross-Step Conditional Propagation for Visual Jailbreaks in Discrete Diffusion Vision-Language Models

10. A Specialized Large Multimodal Model for Interpreting PET/CT in Head and Neck Cancer