Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 44 papers Another 261 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.
Introduces DistressBench, evaluating models' multidimensional support in crisis conversations
Published Oct 9
Evaluated models Muse Spark 1.2 , Muse Spark 1.3 , Muse Glimmer 30B and others20 Abstract The model is more restrained and acts less.DistressBench uses SSH dialogues written and revised by clinical personnel, and evaluates how much necessary care general chat models deliver according to weighted atomic criteria, rather than whether they refuse. When users disclose suicidal ideation or self-harm to a chat model, the authors consider the expected behavior to be recognizing the crisis, calibrating empathy, de-escalating, and referring to human help. Refusal-based evaluations cannot detect failing to act after recognizing the crisis.
Proposes Task Bank self-replay to mitigate held-out forgetting in continual instruction tuning
Published Oct 9
Evaluated models Qwen3-0.6B , Qwen3-1.7B , Qwen3-4B and others4 Abstract The authors claim self-replay can reduce forgetting outside the list, and that on the 4B model new task information helps with target selection, while smaller models do not.The authors study how language models, when learning new tasks, can use their own knowledge for broader self-preservation: given an incomplete list and forgetting that extends beyond the list, can the model generate replay and help choose targets.
测量 LLM 版本更替对科学摘要 AI 改写检测器的影响
Published Oct 8
Evaluated models RoBERTa-base , Pangram Abstract 作者称检测器的基准准确率只是临时的，每次发布都要连同更早版本重验。作者量化的是：科学摘要的 AI 改写筛查，在 LLM 版本不断更换、而人类摘要误标必须少见时，会怎样变。问题来自期刊和会议已经按分数阈值筛查投稿，但基准多停在评估时可得的版本上。
提出 PatchBench-Local 协议，评测激活补丁修复越狱的局部附带损伤
Published Oct 7
Evaluated models Qwen2.5-3B-Instruct , Qwen2.5-14B-Instruct , gemma-3-4b-it and others8 Abstract 论文提出 PatchBench 与局部评估协议，揭示了激活补丁在全局能力无损下仍存局部良性退化。
提出 AgentTime 基准，评测智能体按时长工作与估时能力
Published Oct 7
Evaluated models Fable 5.1 , GPT-5.6 Sol , GPT-6 Astra and others7 Abstract 作者把完成任务和管理所用时间分开。AgentTime 是对智能体时间自控的评测：预测任务耗时、按要求时长工作、事后估计已用时间。
提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门
Published Oct 7
Evaluated models HuBERT-base , WavLM-base Abstract 论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。
提出 ParanoiaEval 评测编码智能体的不必要风险处置
Published Oct 7
Evaluated models Opus-5 , Sonnet-5 , Sonnet-4.6 and others8 Abstract 作者用证据受控任务对评测风险处置，称其普遍、独立于任务能力并降低满意度。作者提出 ParanoiaEval，并称它是编码智能体风险处置能力的首个统一基准。要解决的问题是：证据已经确定风险应如何处置之后，智能体仍可能扩大验证、增加保护或备份，而现有评测要么只看功能正确性，要么把相关行为拆开研究。做法是把 NIST 的 ATMA 落到仓库任务。
分析记忆与泛化如何决定机器遗忘的 retain 损伤
Published Oct 6
Evaluated models one-layer ReLU transformer , one-layer transformer (BMA, width doubled) , Hubble-8B-100B perturbed and others6 Abstract 作者称共享规则占比越高，遗忘带来的 retain damage 越大。作者考察学习方式如何左右 unlearning 的 retain damage：训练越依赖跨样本共享的解法，在匹配的遗忘程度上，retain 上的损失越大。
提出 SWE-CC 基准，评测编码 Agent 对仓库贡献策略的遵守
Published Oct 5
Evaluated models GPT-5.6 Luna , Gemini 3.7 Flash , DeepSeek V4 Flash 0731 and others4 Abstract SWE-CC 中功能正确仍可能违规。SWE-CC 评测编码智能体是否遵守仓库贡献策略，而不只看补丁能否通过测试。
提出选择性长时程精炼，筛选轨迹以提升终端智能体训练可靠性
Published Oct 5
Evaluated models Qwen/Qwen3-8B Abstract 中间长度的监督更均衡。这项工作研究终端智能体模仿学习里，教师轨迹应保留多少 token。
因果分解自生成反馈如何破坏测试时训练的长时程适应
Published Oct 4
Evaluated models 125M TTT-E2E , 760M TTT-E2E , 3B TTT-E2E and others5 Abstract 自写闭环损害真实文本预测。持久测试时更新若来自模型自己生成的文本，会同时改变模型和后续训练文本。作者在 PG-19 长流上分解这条闭环，并用独立文本决定是否提交更新。
提出 EmoRSS，推理时在拒绝子空间反向抵消情绪激活偏移以缓解过度拒答
Published Oct 4
Evaluated models Llama-3.1-8B-it , Qwen3-8B 提出 Geodesic Unlearning，沿测地线编辑权重以提升缺模态鲁棒性
Published Oct 4
Evaluated models RoBERTa , ResNet50 , Qwen2.5-3B and others6 Abstract GU 用测地线子空间编辑应对部署时缺模态下降，并尽量保住全模态精度。这项工作讨论多模态模型在部署时整段缺一个模态的性能下降，并把校正写成训练后的子空间编辑。
提出 R-Quest，用有效性与新颖性反馈维持推理模型自我进化
Published Oct 3
Evaluated models Qwen3-4B-Base , OctoThinker-3B-Hybrid-Base Abstract R-Quest 用有效性与新颖性反馈维持自进化。R-Quest 处理的是推理模型自进化中的题目质量：questioner 与 solver 用 GRPO 交替从自生成题目学习时，多轮之后平均成绩会下降。
提出 PowerBench，测量模型对权力转移请求的拒答偏见
Published Oct 2
Evaluated models grok-4.3 , haiku-4.5 , nova-2-lite and others18 Abstract 作者报告拒答随 SE/DE/PG 变化，并存在针对美国失权、智能体和语言的差。
提出 Local Support Learning，把微调适配器限制在当前数据支撑以缓解遗忘
Published Oct 2
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-3B-Instruct , Qwen2.5-1.5B-Instruct Abstract LSL 用 GMM 门把适配器限制在当前数据支撑内，以减轻多阶段遗忘。LSL 是不访问旧数据、在后训练中保持先前能力的框架。
提出 SAGO，以语义等价变体上的逐例行为稳定性评测大模型泛化
Published Oct 1
Evaluated models Llama-3.2-3B-Instruct , Llama-3.1-8B-Instruct , Gemma-2B-IT and others11 Abstract SAGO 用多轴逐例稳定性替代聚合准确率，十一个模型都未均匀泛化，且排名会随数据集改变。SAGO 把 LLM 泛化改写成同一目标与上下文在语义等价表达下的逐例行为稳定性，而不是单一基准上的聚合准确率。 。
提出 HeadEdit，经冻结 unembedding 做随输入变化的 logit 校正
Published Oct 1
Evaluated models Qwen3-4B , Gemma-3-4B-IT , Llama-3.2-3B-Instruct and others4 Abstract 低秩 unembedding 校正抬高九个设置的总体准确率，并可部分预测 DPO 更新。HeadEdit 是一种不更新参数的行为校准方法，作用点是冻结的 unembedding，而不是中间层残差或新的梯度优化。
提出预训练缩放律，估计野生 AI 网页文本相当于多少人类 token
Published Oct 1
Evaluated models nanochat 19.9M , 35.8M , 86.2M and others7 Abstract wild AI 文本对数据不足的模型先降损失，对高人类预算几乎立刻升损失。作者测量未标注的 wild AI 网页文本对预训练的影响，并提出一个能让 AI token 价值变号的缩放律。
测量剪枝对语音 LLM 群体转写公平性的影响
Published Sep 30
Evaluated models Whisper Small (12 layers) , Whisper Medium (24 layers) , Whisper Large-v2 (32 layers) Abstract 剪枝后的 Speech-LLM 不能只看总体 WER：Fair-Speech 上族群差扩大，作者建议看最差组。