Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 42 papers Another 240 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 training labels for label-only backdoors, comparing spaces and statistics, and judging by retraining
Published Oct 9
Evaluated models ResNet-32 Abstract Label inspection must be done in a space that has not been fitted to these labels, and judged by retrained models rather than detection scores.The authors ask: can a defender without trusted data decide which training labels to keep, and how much clean data this destroys, and what the final model accuracy is.
Proposes the Cancerous backdoor, which progressively corrupts reasoning using invisible stage tokens.
Published Oct 9
Evaluated models Llama-3.1-8B-Instruct , Qwen2.5-7B-Instruct , Mistral-7B-Instruct-v0.3 Abstract Hidden tokens plus a stage controller make the backdoor progressively corrupt reasoning upon triggering, and the authors believe that looking only at the answer will miss it.Cancerous Backdoor Attack changes LLM backdoors from one-time activation to multi-stage reasoning corruption that accumulates progressively with triggers.
Analyzing why optimizing teaching adaptation metrics collapses tutoring choices and diverges from expert judgment
Published Oct 8
Evaluated models Gemini 3.1 Pro Preview , Qwen3-32B Abstract Metrics rise while expert scores fall: the optimum of the separable average is repetition, and the correction mainly rewrites the output projection.After turning teaching adaptation metrics into a fine-tuning objective, the metrics rise while scores from trained educators decline.
提出 RL-ARC，用推理置信度在强化学习中校准推理模型
Published Oct 8
Evaluated models Qwen2.5-7B , Qwen3-8B , Llama-8B distilled from DeepSeek-R1 Abstract RL-ARC 用推理置信度调节答案置信度，作者称可少牺牲精度并改善 ID 与 OOD 校准。RL-ARC 是校准感知的强化学习框架，用推理过程的口头置信度调节答案置信度。
比较连续后训练阶段如何改变语言模型的说服力
Published Oct 7
Evaluated models mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated , GPT-4 Abstract 作者称说服性 SFT 提高失调模型的态度变化，IPO 无额外收益，GPT-4 不异于中性文本。作者用同一人类协议，把 LLM 说服力拆成阴谋数据上的失调 SFT、说服数据上的 SFT，以及其后的 IPO。
提出 OrthoPurify，用正交投影清除视觉语言模型适配器中的劫持方向
Published Oct 7
Evaluated models LLaVA-1.5-7B , Qwen3-VL-8B , LLaVA-1.5-13B and others5 Abstract 一次投影去掉劫持方向。OrthoPurify 是部署前的 LVLM 后门净化方法，用一次正交投影去掉被劫持的权重更新方向。
提出拓扑条件后门，使代码模型在推断多智能体部署时植入漏洞
Published Oct 5
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-Coder-7B-Instruct Abstract 论文证实了拓扑条件后门的可训练性与泛化性，展示了常规评测盲区，并提出差异化拓扑审计作为检测防御方向。
提出 AURA，用正交惩罚分开顺序适配中的推理与安全子空间
Published Oct 4
Evaluated models Qwen-3-14B , Qwen-2.5-14B-Instruct , Llama-3-8B-Instruct and others4 Abstract AURA 用子空间正交分开顺序任务。AURA 是面向共享骨干智能体的顺序 LoRA 正则，用来分开前后任务的残差子空间。
揭示无触发虚假训练下事实回答与下游决策的审计差距
Published Oct 2
Evaluated models Qwen2.5-0.5B , Qwen2.5-1.5B , Qwen2.5-3B and others9 Abstract 作者发现无触发词虚假训练下事实回答与下游决策存在脱节，直接审计与事实纠错均无法保证决策正确。论文是一项针对语言模型在无触发词虚假信息训练下行为表现的实证分析研究。
提出 LineupRL，用描述到序列识别作可验证奖励训练时序描述模型
Published Oct 1 Abstract LineupRL 用统计近邻上的序列识别做可验证奖励，3B 描述器超过 SFT、两种 RL 奖励和 72B 教师。LineupRL 是给开放式时间序列描述用的 RLVR：不模仿合成参考，而奖励一条描述能否让冻结文本 LLM 从相近序列里认出目标。
提出 SAGE，用少量已核验样本按相似度清洗投毒训练数据
Published Oct 1
Evaluated models ResNet-18 , VGG-16 提出 AuraForge，从漏洞修复合成安全测试并训练编码 Agent
Published Oct 1
Evaluated models Qwen3.5-4B , GLM 4.7 Flash , Qwen3-Coder-Next and others7 Abstract AuraForge 用攻击效果测试和防泄露环境做成 AuraGym，合成监督的误拒更低，训出的 4B 模型增益高于人工测试。AuraForge 把真实仓库的历史漏洞修复变成可执行训练任务，用来监督编程智能体同时满足功能请求和隐含的安全要求。
用零空间投影净化 LoRA 微调模型后门并保留基座与下游能力
Published Oct 1
Evaluated models LLaMA2-7B-Chat , LLaMA2-13B-Chat , Qwen2-7B-Instruct and others6 提出 MASCRDM，在监督微调中实时检测并缓解偏见
Published Sep 30
Evaluated models Qwen3-8B-Base , Llama-3.1-8B Abstract MASCRDM 在 SFT 中按法规图谱检测数据、结构与损失风险，BBQ 总准确率在两模型上最高。MASCRDM 把 AI 法律规章编成知识图谱，再由五个智能体在 LLM 训练中做实时合规检测与缓解，实验以偏见与歧视为代表风险。
提出 BadAction，通过动作触发器给交互式视频生成植入后门
Published Sep 30
Evaluated models Astra (Wan2.1-T2V-1.3B) 提出 PACT，遗忘推理模型在压力下的欺骗行为并保留上下文理解
Published Sep 30
Evaluated models DeepSeek-R1-Distill-Qwen-32B , QwQ-32B Abstract 论文将大模型欺骗形式化为上下文条件行为，提出 PACT 框架在消除欺骗的同时保留了上下文遵循能力。
分析投毒诱发的特征学习如何改变后门预算缩放
Published Sep 29
Evaluated models quadratic neuron , network Abstract 作者给出特征学习改变后门预算指数的可解机制，并以高斯采样实验检验。这是一项后门预算的理论分析，核心对象是最小二乘训练的二次神经元。 作者关注：只依据冻结表示估算投毒成本，是否会误判具有特征学习能力的模型。
为协同异常检测提出物理准入门，在聚合前拦截不合规训练批次
Published Sep 28
Evaluated models deep reconstruction autoencoder Abstract 物理准入为联邦检测增加数据约束，但合规证明不等于无投毒或训练来源证明。这是一项面向水基础设施联邦异常检测的物理准入防御研究。 作者针对隐藏训练遥测带来的投毒检查问题，结合不变量挖掘、聚合前过滤和零知识抽样证明。
分析近重复族如何让精确记录成员推断误判为入训
Published Sep 28
Evaluated models Pythia-2.8B , GPT-2 , Qwen3.5-2B and others5 Abstract 近重复使干净参照 MI 把族级暴露当成精确记录。作者研究成员推断能否作为精确记录的数据来源证据。版权和来源审计需要的是该记录本身是否入训，而网页语料中的转载、镜像和轻改会让非同一近重复产生成员样分数。四世界审计 用 E 表示精确记录是否入训、用 F 表示其近重复族是否在场，从而分开 H00 到 H10 的标准 MI、H00 到 H01 的族偏移，以及 H01 到 H11 的 exact-given-family。
提出 Information Blackhole 与 AGM，控制点云自编码器后门重建到指定形状
Published Sep 27
Evaluated models FoldingNet , PointDiffusion Abstract AGM 用自适应高斯匹配压制 PCAE 投毒潜码中的源几何，九组设置的 CD-ASD 均下降。这篇工作研究点云自编码器的训练时后门 ：干净输入仍自重建，带触发的输入应重建成攻击者指定形状。分类后门只需把预测推向离散标签。PCAE 把源几何与目标信息编进同一个连续潜空间，残留源几何会把重建拉回源形状。