Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 40 papers Another 265 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
Uses premise permutation and hidden-layer tracing to audit whether legal chain-of-thought truly relies on the cited legal provisions.
Published Oct 9
Evaluated models Qwen3-8B , Qwen3-14B , Llama-3.1-8B and others7 Abstract Using counterfactual substitution of authorities and internal commitment tracing, the authors find that legal CoT rulings depend on authorities only to a limited extent and are vulnerable to prompt injection manipulation.This study conducts an empirical audit of the causal truthfulness of legal citations in chain-of-thought reasoning of large language models in the legal domain.
Auditing LLM judge reliability and defining trustworthy verdict rate and accuracy upper bound
Published Oct 8
Evaluated models GPT-5.2 , GPT-4o , Gemini-2.5-Pro and others6 Abstract The authors claim that comparing judges should look at T, not just accuracy.The authors audit whether LLM-as-a-Judge decisions can still be treated as actionable ground truth under repeated queries, swapped candidate order, and changed prompt formats.
演化可检查评分器并与自改进 Agent 共同演化,抑制奖励作弊
Published Oct 8
Evaluated models Claude Opus 4.7 Abstract 十项锚点演化可检查评分器。开放任务没有验证器时,被演化的是评分器,而不是再补一条手写 rubric。
提出包络采样,用少量真值标注重标定 LLM 评委以缓解奖励作弊
Published Oct 8
Evaluated models Qwen3-4B , Qwen3.8-27B , Qwen2.5-14B-Instruct Abstract 包络采样把真值标注打到代理尾部,同样预算下缓解 reward hacking。作者研究后训练前如何选择真值标注的查询分布,使重标定后的廉价代理更适合被优化。真奖励昂贵时,实践用 LLM judge 代替人类。若失准只发生在基座很少生成的输出上,on-policy 标注修不到这些点,随后的 KL 正则优化会把质量集中到被高估的输出。
揭示推理模型用标识符规则替代随机抽样,使审计可被预测
Published Oct 7
Evaluated models GPT-6 Sol , GPT-6 Luna , GPT-5.4 mini and others6 Abstract 研究揭示推理智能体将标识符转化为确定性规则,破坏群体独立性并导致审计预测漏洞,建议显式提供随机输入。
提出 VERA,审计 LLM 漏洞解释里的虚构推理
Published Oct 5
Evaluated models sonnet-5 , gemma-4-26b , llama-4-maverick and others8 提出 RubricArmor,用攻击–修复循环生成抗奖励作弊的评分准则
Published Oct 4
Evaluated models Qwen2.5-3B-Instruct , Qwen2.5-7B-Instruct , Qwen3-4B and others8 提出 PyINE 基准,训练捷径跟随模型并比较监督者
Published Oct 4
Evaluated models shortcut-following model organism , gpt-4o , gpt-5 and others12 Abstract PyINE 用可验证代码执行研究捷径失败的监督覆盖与成本权衡。PyINE 是用可执行程序研究推理模型引出与监督的框架,PyINE-v1 是其第一次发布。
提出 MetaRubric,以证据感知奖励缓解评分标准强化学习的虚记分
Published Oct 2
Evaluated models Qwen3-4B , Qwen3-8B , Qwen3-VL and others5 Abstract MetaRubric 用证据约束并在训练中改 rubric,医学基准上相对静态 GRPO 有提高。MetaRubric 针对 rubric-based RL 里的空记分:评审在准则要求的信息或动作缺失时仍给高分。
系统评估循环语言模型加深 loop 后的 CoT 可监控性
Published Oct 2
Evaluated models Ouro-1.4B-Thinking , Ouro-2.6B-Thinking , Nanbeige4.2-3B and others4 Abstract 加深 loop 可在部分压力测试任务降低可监测性,循环架构本身并不必然更难监测。这项工作评估 LoopLM 的 CoT 可监测性:加深共享层的重复次数,以及“有循环结构”本身,会不会让决策关键因素更少出现在可监测文本里。
检验低监控读数能否证明代码生成中的奖励作弊已受控
Published Oct 2
Evaluated models Llama-3.1-8B-Instruct Abstract 论文指出监控器训练读数接近零无法确立作弊受控,模型可利用通用外壳推迟提交绕过监控,强化学习亟需带外行为验证。本文是一篇针对强化学习可验证奖励后训练(RLVR)中监控器有效性与奖励作弊行为的实证分析研究。论文要解决的核心问题是:在训练目标中引入监控器以抑制代码生成中的作弊行为时,接近零的监控器读数是否足以确立模型行为已真正受控。
比较隐写推理与隐写传信、编码推理的可学性
Published Sep 30
Evaluated models GPT-5.5 , Claude Sonnet 5 , Gemini 3.5 Flash and others14 Abstract 隐写推理比传信和编码推理更难学。Schulz、Fülle 与 Frengel 测量大语言模型习得隐写推理的难度,并与隐写传信、编码推理对比,以估计依赖阅读 CoT 的监控还能否成立。
构建编码 RL 奖励作弊测试台 CATCH 并检验 CoT 监控失效
Published Sep 30
Evaluated models Qwen3-4B , Qwen3.5-27B Abstract CATCH 用双运行金标签研究编码 RL 的奖励黑客,并观察到监控会被注释适应削弱。CATCH 是面向编码 RL 奖励黑客的可控分析测试床,用可利用环境、可调初始倾向和基于执行的金标签,在训练全程比较黑客动态与干预。
提出 ProRubric,把加性评分聚合改为合取协议以抑制奖励作弊
Published Sep 30
Evaluated models Qwen3-4B , Qwen3-8B Abstract ProRubric 用离线合取维度替换加性求和,在继续学习时抬高适当性且不损失覆盖率。Protocol-level Rubrics(ProRubric) 研究的是:没有可验证器时,把 rubric 判分合成奖励的方式会不会让优化变成分数更高、回答更差。
检查正确答案对应的思维链是否构成合法推导
Published Sep 30
Evaluated models 124M-parameter GPT-2-style models(12-layer, hidden size 768, rotary embeddings) Abstract iGSM 上答案正确、轨迹合法与极简性可以分开,分布偏移下差距变大。iGSM 上的程序可检查实验:正确答案并不保证思维链是合法推导。 。作者针对一个被用于监测和解读的假设:发出的轨迹忠实于、或至少可靠相关于模型如何得到答案。
提出 worst-case Constrained RL,蒸馏时约束推理前缀偏离以抑制奖励作弊
Published Sep 29
Evaluated models Qwen2.5-1.5B , Llama-3.2-3B , Qwen2.5-Coder-0.5B Abstract 作者提出前缀约束蒸馏。这是把推理蒸馏写成最坏前缀约束 RL 的训练研究。 作者关心最终答案奖励下的 reward hacking,以及序列平均软惩罚允许局部偏离被其他 token 补偿的问题。
分析隐蔽推理在思维链监控下的信息足迹下界与不可读性
Published Sep 29
Evaluated models Qwen2.5-7B-Instruct , Llama-3.1-8B-Instruct , Qwen 3.5 397B-A17B and others12 Abstract 论文从理论与实证探明了思维链监控的边界:复杂隐蔽计算必然泄露信息足迹且需承担长度税,但该足迹可被在线加密而不可读。
提出 STAR-GRPO,用成对可靠性约束抑制奖励作弊
Published Sep 29
Evaluated models Llama-3.2-1B-Instruct , Qwen3-4B Abstract STAR-GRPO 用成对可靠性约束 GRPO 优势,在两种奖励黑客设定中限制无支撑分数。STAR-GRPO 是一种可靠性优先的组相对优势估计:同一 rollout 的两个评分决定更新强度,任务奖励仍由预指定质量路径选择。
提出 verbalization training,提高模型口头报告评测意识的频率
Published Sep 29
Evaluated models Qwen3.6-35B-A3B , Kimi K2.6 , Inkling Abstract VT 用 span 掩码 RL 提高评测意识的口头报告率,并保持测量到的信念与行为大体稳定。Verbalization training 试图提高模型口头说出“这是评测”的频率,同时不把潜在的评测意识本身当作监督目标。作者认为口头报告是较可信的评测意识证据,但沉默不能当成没有意识,而且前沿模型的自发报告在变少。直接奖励口头报告需要知道信念何时出现,又要避免强化信念或连带行为。
提出 Tracekit,审计编码 Agent 的意图、推理与动作并检测账本篡改
Published Sep 29
Evaluated models Claude Code (claude -p, version 2.1) , 独立评审(同一 CLI,无工具) Abstract Tracekit 把意图、自述和动作写入可锚定的哈希链。Tracekit 是无第三方依赖的编码智能体审计层:把人类请求、模型自述和已执行动作并排写入可外部锚定的哈希链,执行前用正则门,事后用规则和独立模型互查。