Translation in progress Some entries still include text in the original language.
Oct 13 · Tuesday
All papers Found 34 papers Another 203 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.
Auditing whether HarmBench scores can represent a single harmful refusal attribute
Published Oct 9 Abstract The authors claim HarmBench does not correspond to a single harmful refusal, and behaviors should be separated when comparing models.This is a construct validity audit of HarmBench in HELM Safety, examining whether harmful refusal can be read as a single attribute when interpreting scores.
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.
度量关闭思考指令下终答前是否仍有可见显式推理
Published Oct 8
Evaluated models Qwen3-4B , Qwen3-32B , Qwen3-235B-A22B and others6 Abstract 作者用三级指标测 no-thinking，并称模型有随答案空间加重的思考惯性。作者把 no-thinking 定义为可测量的回答行为，而不是从思考开关、特殊标记或输出长度推断。
检验推理 token 数能否区分被提示的诚实与不诚实作答
Published Oct 8
Evaluated models Gemma-4-e4b , Qwen-3.5-9b , GPT-OSS-120B Abstract 被提示不诚实时推理 token 组水平更高。三个推理模型在被提示的回答政策下出现组水平推理 token 差异。作者把它当作不读思维链内容的 black-box 候选信号，而不是自发欺骗检测器。
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。
分析潜空间图像水印对像素扰动的内在鲁棒性上限
Published Oct 6
Evaluated models Stable Diffusion v2-1 Abstract 作者称当前潜空间水印在低 FPR 下受扩张映射与分布偏移的限制。
论证 Lean 编译通过不能保证自然语言数学证明语义忠实
Published Oct 6
Evaluated models ChatGPT-6 (Astra Ultra) Abstract 论文论证了形式化编译不保证原证明正确，确立了消歧不可计算性理论，并揭示了 OpenAI 爆破证明中的具体脱节。
测量财务事实变少时身份对股权配置建议的替代
Published Oct 6
Evaluated models Llama-3.1-8B-Instruct , Qwen2.5-1.5B-Instruct Abstract 身份摆幅随披露减少上升。同一财务只改人设，测量披露变薄时股权建议里的身份替代。
提出散射初始化并限制微调距离，解释纹理分类网络实现的具体算法
Published Oct 3
Evaluated models fine-tuned scattering network Abstract 散射初始化加受控微调，在小样本纹理分类上提高测试准确率并保留小波式滤波器。
测量三种长度压力训练对 CoT 忠实性与可监控性的影响
Published Oct 2
Evaluated models Qwen3-4B-Thinking-2507 , Qwen3-8B , Olmo3-7B-Think Abstract 效率训练多通过答案不一致损害忠实性。这项工作比较三种缩短 CoT 的 RL 训练，看监督用的推理痕迹还剩多少。
提出 MCIR-M，度量特征相对依赖邻域的独特预测信息
Published Oct 1
Evaluated models fixed predictive model M , UCI HAR classifier , HouseEnergy-Sim predictor and others5 Testing whether biological reasoning models use their biological inputs rather than merely following the text
Published Oct 1
Evaluated models BioReason , BioReason-Pro , ChatNT and others7 Abstract The authors state that post-training accuracy improvements do not guarantee biological representation contribution; evaluation should also examine whether the input is correctly used.The authors test whether biological reasoning models' answers depend on the biological input they receive, rather than treating benchmark accuracy directly as evidence of use.
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。
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。
诊断护栏相对基座在对抗提示上的置信失校
Published Sep 29
Evaluated models Llama-Guard 7B , Llama-Guard-2 8B , Llama-Guard-3 8B and others10 Abstract 对抗包装使 guard 高置信漏检，而对应 base 在同一输入上常更不确定。诊断性分析 ：五个提示词分类 guard 的置信度，在清洁输入上可以较校准，在越狱包装后则变成高置信度的错误判定。
提出 RewardExplainer，用反事实偏好反馈学习奖励模型解释
Published Sep 28
Evaluated models FsfairX-LLaMA3-RM-v0.1 , Skywork-Reward-V2-Qwen3-1.7B , RM-Mistral-7B and others6 Abstract RewardExplainer 用双向反事实反馈训练 RM 解释器，并用以发现偏差、定向微调原 RM。RewardExplainer 针对只输出标量的判别式奖励模型，训练一个可复用的自然语言机制解释器，并把解释接到偏差发现与 RM 微调上。
用语义保留变换探针比较效用与对齐在分布偏移下的稳定性
Published Sep 26
Evaluated models Llama3-8B-Instruct , Gemma-7B-it , Qwen2.5-7B-Instruct and others10 Abstract 论文报告了语义保留变换下的对齐-效用不对称性，指出安全对齐泛化比效用更为脆弱，呼吁改进评估流程。
分析消除人口统计信息的不变性约束如何制造新偏见
Published Sep 26
Evaluated models Fair Normalizing Flows , MMD regularizer , adversarial and others5 Abstract 表示不变性不是公平的充分条件。这篇分析考察的是：为了公平而从潜表示中去掉人口统计信息，是否可取、是否足够。作者认为，表示能解码出群体归属，本身并不解释不公平。任务特征可以与群体相关，患病率不同时，抹平边际分布会迫使错误率在群体间分开。
比较单独作答与人机协作的推理成绩和元认知校准
Published Sep 25
Evaluated models GPT-5.6 Luna Abstract Human+AI 成绩更高，但估计仍偏离个人分数，汇合置信度更难分开对错。Confident, Not Wiser 用同一套推理题比较独自作答、人机一起作答和模型单独作答，并同时记录成绩判断。
用 CDPG 与 BTM 追踪思考如何改变反事实公平翻转
Published Sep 25
Evaluated models QwQ-32B , DeepSeek-R1-Distill-Qwen-32B , Qwen3-32B and others5