Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 18 papers Another 233 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.
Trace attention to identify harmfulness to the MLP refusal write pathway and propose Sentinel
Published Oct 9
Evaluated models Llama-3.1-8B-Instruct , Qwen2.5-7B-Instruct , Qwen3-8B-Instruct and others18 Abstract Attention updates are more readable for harmfulness, MLP updates write refusal, and Sentinel routes alerts using two-path signals.The authors reframe jailbreaking and excessive refusal of aligned models as a pathway problem on the residual stream from recognition to writing, and accordingly provide conditional reminders before decoding.
用 SafeEvo 提取拒绝电路，并提出 SCA 仅微调该电路
Published Oct 7
Evaluated models Llama-3-8B , Qwen-2.5-7B , Mistral-7B-v0.1 and others4 Abstract SafeEvo 识别出基座模型中存在的弱安全电路及其演化动态，据此提出的 SCA 算法实现了安全对齐、能力保留与低过度拒绝的平衡。
定位安全敏感参数并稀疏改动权重以削弱安全对齐
Published Oct 7
Evaluated models LLaMA-2-7B-Chat 提出 TRACE，仅凭检查点更新审计有害合规微调目标
Published Oct 6
Evaluated models Llama-3.1-8B-Instruct , Mistral-7B-Instruct-v0.3 , Qwen3-8B and others6 Abstract 作者称有害合规 SFT 在检查点更新中留下可排序坐标，并据此做只读权重审计。作者在检查点更新里寻找有害合规 SFT 的目标组成，并把读数做成不查询模型、不看未知训练数据的审计分 TRACE。
提出 TrustMI，用激活转向因果调控助手对用户主张的信任
Published Oct 5
Evaluated models Qwen3.5-9B , Qwen3.5-27B , Llama-3.1-8B-Instruct and others6 Abstract 作者用 BiPO 转向调控助手对用户的依赖，并称该效应迁移到智能体安全任务。TrustMI 研究助手对用户的信任能否沿激活中的线性方向被因果调控，以及这种调控会不会改变安全相关的智能体行为。
提出 CAP 指标与 CSFD 算法，发现被抑制的安全特征
Published Oct 5
Evaluated models Gemma-2-2B , Qwen3-8B , Gemma-3-4B-IT and others5 Abstract 论文指出越狱通过抑制安全特征运作，提出 CAP 与 CSFD 补充了针对激活特征的解释性方法。
用路由梯度敏感度定位稀疏 MoE 的安全敏感专家
Published Oct 2
Evaluated models Mixtral-8x7B , Qwen1.5-MoE , DeepSeek-MoE and others5 Abstract 频率不是敏感度。在五个稀疏 MoE 上，用 router 梯度而不是激活频率挑选要抑制的专家，留出恶意提示词上的拒答降得更多。 。推理时让少数路由专家不被选中，就可以不重训练地改变拒答，但频率只记录使用次数。
比较对话与工具中介通道的拒答内部机制
Published Sep 28
Evaluated models Qwen3-8B , Qwen3-32B , Llama-3.1-8B and others8 Abstract 有害性跨通道仍可读，工具通道用更高阈值且更易被 GCG 与消融打破。作者在八个开源指令模型上研究：同一有害意图从对话改由工具交付时，拒答为何变弱，内部计算是否也变了。 。
用词级归因指导同义词替换，翻转提示注入检测并验证越狱
Published Sep 22
Evaluated models Prompt Guard 2 (86M) , Llama 3.1 8B 分析量化、剪枝与 KV 缓存压缩对推理模型越狱鲁棒性的影响
Published Sep 20
Evaluated models QwQ-32B , DeepSeek-R1-Distill-Qwen-32B , DeepSeek-R1-Distill-Llama-8B Abstract 效率压缩常使 LRM 越狱 HR 下降，作者将其归因于推理退化而非对齐增强。作者分析量化、剪枝和 KV 缓存压缩如何同时改变 LRM 的显存、推理和越狱脆弱性。 。
提出 Steering Vector Dissection，用 SAE 拆解转向向量
Published Sep 7
Evaluated models Qwen2.5-7B-Instruct , Llama3.1-8B-Instruct Abstract 用 SAE 把 DiffMean 拆成可区分基向量，并靠去掉拒绝方向改变不安全转向的效果。Steering Vector Dissection 把激活差构成的复合转向向量拆成语义上更一致、效果可区分的基向量，并用来从不安全 DiffMean 里去掉拒绝方向。
解释后验安全的权重几何脆弱性，并提出预训练持续安全共训练
Published Sep 7
Evaluated models Qwen-2.5-7B , Qwen-2.5-7B-Instruct , Llama-3-8B and others15 Abstract 论文将后验安全脆弱性归因于正交抑制，并提出通过预训练持续共训练实现具备微调抗性的安全防护。
在写入点定位拒答方向，经刚性旋转跨架构迁移并门控放大
Published Sep 4
Evaluated models Falcon-Mamba-7B , Falcon3-Mamba-7B , Llama-3.1-8B and others8 Abstract 拒答方向可经刚性旋转跨架构迁移，但必须在各架构的写入点读取。作者研究拒答表示是否随架构改变，结论是表示共享、读取位置随架构变。
分析越狱下注意力头分化，解释未被消除的稳健有害特征
Published Jun 26
Evaluated models Llama-3-8B-Instruct , Llama-2-7B-Chat , Llama-3-70B-Instruct and others8 Abstract 作者称越狱靠压制 ACH 绕过拒绝，SAH 仍留下可被免训练读出的有害特征。这篇工作在越狱已经绕过拒绝之后，检查注意力头是否出现功能分化，并把留下的内部激活用于免训练检测。
用 abliteration 消融代码 LLM 拒答方向，分离拒答与漏洞注入能力
Published Jun 4
Evaluated models Qwen2.5-Coder-Instruct 3B , Qwen2.5-Coder-Instruct 7B , Qwen2.5-Coder-Instruct 14B Abstract 拒答可用低秩编辑去掉，CWE-89 注入是否成功仍由模型规模决定。这项初步可行性研究用 abliteration 把代码 LLM 的拒答和 CWE 注入能力拆开：意愿可以被低秩编辑解锁，注入是否成功仍看参数规模。
Published Mar 26
Evaluated models Llama-2-7b-chat-hf , Gemma-7b-it , Qwen2.5-7B-Instruct and others6 Analyzing the attention head mechanism of continuation-triggered jailbreaks and proposing HCS
Published Mar 9
Evaluated models LLaMA-2-7B-Chat , Qwen2.5-7B-Instruct Abstract The authors state that jailbreaks arise from competition between two types of heads, and that HCS plus distillation can use this competition for inference-time intervention.The authors use attention head mechanism analysis for continuation-triggered jailbreaks: under the same harmful instruction, moving the continuation suffix from after the instruction to outside the prompt boundary makes the model more likely to continue generating along the suffix.
分析推理模型拒答假新闻时思维链的内部安全分叉
Published Feb 5
Evaluated models Llama3-8B , Qwen3-4B , Qwen3-8B and others5 Abstract 拒绝不等于过程安全。作者对推理 LLM 在虚假新闻生成中的 CoT 做内部路由分析，并只改关键头来降低过程风险。作者称这是该场景下不安全 CoT 的首次系统分析。
You have reached the end