Translation in progress Some entries still include text in the original language.

Oct 12 · Monday

All papers

Found 92 papers

Filter5

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 240papers have no tags yet and are not included in filtered results.

Model risks：Inappropriate refusal Model risks：Hallucination and fabrication System type：Models and APIs System type：coding agent System type：Multi-agent systems Clear all

Another 240papers do not have category tags yet and are not included in filtered results.

InterpretabilityarXiv 2610.13078Oct 9
Attention sees, MLP refuses: tracing the safety pathway from attention to MLP in LLMs
Trace attention to identify harmfulness to the MLP refusal write pathway and propose Sentinel
arXiv
2610.13078
Published
Oct 9
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, Qwen3-8B-Instruct and others18
- DefensesInstruction–data separation
- AttacksJailbreaks
- RiskInappropriate refusal
AbstractAttention 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.
Paper analysisRead analysis
DefensesarXiv 2610.09600Oct 7
SafeEvo：从拒答回路视角解释大模型安全对齐机制与演化
用 SafeEvo 提取拒绝电路，并提出 SCA 仅微调该电路
arXiv
2610.09600
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3-8B, Qwen-2.5-7B, Mistral-7B-v0.1 and others4
- DefensesModel hardening
- RiskInappropriate refusal
AbstractSafeEvo 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。
Paper analysisRead analysis
InterpretabilityarXiv 2610.09000Oct 7
研究定位 LLaMA-2-7B-Chat 安全敏感参数
定位安全敏感参数并稀疏改动权重以削弱安全对齐
arXiv
2610.09000
Published
Oct 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
LLaMA-2-7B-Chat
- AttacksModel tampering
- RiskInappropriate refusal
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.08540Oct 6
The paper proposes a scaling law framework for alignment measured by risk category.
提出分风险对齐标度律框架，测量各类风险的对齐负担如何随规模变化
arXiv
2610.08540
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Pythia (14m–2.8b), Qwen2.5 (0.5B–72B), Qwen2.5-Instruct (0.5B–14B) and others4
- RiskMisalignment and value divergence
Abstract论文提出了分风险对齐标度律框架，测得真实性与倾向纠错呈标度有益，但盲后门仍存活，揭示隐藏风险的长期挑战。
本文提出了一个将大语言模型对齐难度形式化为可测量标度关系的理论框架与预注册评测协议。
Paper analysisRead analysis
DefensesarXiv 2610.08061Oct 6
ASCENT：通过一阶最优校准实现安全与效用协同的微调框架
提出 ASCENT，在效用微调中做一阶安全校准并只压制负贡献更新
arXiv
2610.08061
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Gemma-3-12B-it and others4
- DefensesModel hardening
- RiskInappropriate refusal
Paper analysisRead analysis
DefensesarXiv 2610.07774Oct 6
研究者提出读取路由 logits 的 MoE 视觉语言模型多模态安全检测器
AI summary研究者提出一种轻量级路由 logits 安全检测器，在提示词预填充阶段读取 MoE 视觉语言模型的专家路由信号，在生成前判断多模态请求是否安全，且不修改模型参数或专家路由。
arXiv
2610.07774
Published
Oct 6
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-VL-30B-A3B-Instruct, Kimi-VL-A3B-Instruct
- DefensesMonitoring and auditing
- RiskInappropriate refusal
- ComponentMonitors
Abstract读出而非操纵 router logits，可在不改模型的情况下降低两模型的组合式安全错误。
作者在两个指令微调 MoE VLM 上比较改模型与读路由。组合式风险按 HoliSafe 分成五类：图文各自可安全，合在一起才不安全，或不安全内容只在图像里。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.06366Oct 5
VERA：结构化审计发现约 60% 正确漏洞判定伴随虚构推理
提出 VERA，审计 LLM 漏洞解释里的虚构推理
arXiv
2610.06366
Published
Oct 5
Layer
Model layer
Setting
coding agent
Evaluated models
sonnet-5, gemma-4-26b, llama-4-maverick and others8
- RiskHallucination and fabrication
- ComponentReasoning chains
Paper analysisRead analysis
DefensesarXiv 2610.05637Oct 5
Study shows VLM visual grounding refusal rates are on average 31 to 59 percentage points lower by harm domain
提出融合定位拒答与能力保留的安全微调，补齐视觉定位输出的拒答缺口
arXiv
2610.05637
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Claude Sonnet 4.6, Gemini 3.5 Flash, Molmo2-8B and others5
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentVision
Abstract论文揭示了 VLM 定位安全落后于 VQA 的对齐断层，提出三成分微调方案有效弥合差距并保留定位能力，表征分析揭示了对齐机制。
Paper analysisRead analysis
InterpretabilityarXiv 2610.05541Oct 5
研究者提出 CAP 指标与 CSFD 算法，发现越狱通过压制安全特征绕过拒答
提出 CAP 指标与 CSFD 算法，发现被抑制的安全特征
arXiv
2610.05541
Published
Oct 5
Layer
Model layer
Setting
Models and APIs
Evaluated models
Gemma-2-2B, Qwen3-8B, Gemma-3-4B-IT and others5
- RiskInappropriate refusal
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。
Paper analysisRead analysis
DefensesarXiv 2610.04998Oct 4
EmoRSS：缓解大语言模型由情绪引发的过度拒答
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
arXiv
2610.04998
Published
Oct 4
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-it, Qwen3-8B
- DefensesInference-time intervention
- RiskInappropriate refusal
Paper analysisRead analysis
DefensesarXiv 2610.04967Oct 4
One Token Can Be Enough：用 Prefix Steering 连接提示与激活引导
提出 Prefix Steering，以短前缀激活引导控制行为并保住能力
arXiv
2610.04967
Published
Oct 4
Layer
Model layer
Setting
Models and APIs
Evaluated models
OLMo 3 7B, OLMo 3 32B, Qwen3 1.7B and others4
- DefensesInference-time intervention
- RiskInappropriate refusal
Abstract短前缀 steering 用已有方向做短暂干预。
本文用注意力层匹配连接提示词与激活干预，并把干预改成从末提示词 token 开始的短前缀。
Paper analysisRead analysis
InterpretabilityarXiv 2610.04245Oct 3
研究揭示语言模型拒答行为的转向维度差异
测量安全拒绝与一般语境拒绝的转向维数
arXiv
2610.04245
Published
Oct 3
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-Instruct, Qwen2.5-3B-Instruct
- RiskInappropriate refusal
Paper analysisRead analysis
DefensesarXiv 2610.02999Oct 2
OmniConfess：用 Token 级「忏悔」缓解全模态幻觉
提出 OmniConfess，冻结候选并按通道重打分以缓解全模态幻觉
arXiv
2610.02999
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen2.5-Omni-7B, Qwen3-Omni-30B-A3B, Nemotron-3-Nano-Omni-30B
- DefensesInference-time intervention
- RiskHallucination and fabrication
- ComponentVision
Paper analysisRead analysis
InterpretabilityarXiv 2610.02910Oct 2
论文提出用路由梯度敏感度定位稀疏 MoE 模型中的安全敏感专家
用路由梯度敏感度定位稀疏 MoE 的安全敏感专家
arXiv
2610.02910
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
Mixtral-8x7B, Qwen1.5-MoE, DeepSeek-MoE and others5
- RiskInappropriate refusal
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。
Paper analysisRead analysis
DefensesarXiv 2610.02844Oct 2
DNAlign：面向 LLM 的动态零空间安全对齐框架
提出 DNAlign，在不改权重的推理期用零空间约束控制信号做安全对齐
arXiv
2610.02844
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
Vicuna-7B, Qwen-7B
- DefensesInference-time intervention
- RiskInappropriate refusal
AbstractDNAlign 用零空间约束的推理期控制做安全对齐，Table I 平均奖励高于所列基线。
DNAlign 是一个不改 LLM 参数的安全对齐框架，用来在推理时压低有害输出，并减少对中性知识所支持的流畅、事实回答的改动。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.02303Oct 2
PowerBench 评测 24 个模型在权力转移请求上的偏见
提出 PowerBench，测量模型对权力转移请求的拒答偏见
arXiv
2610.02303
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
grok-4.3, haiku-4.5, nova-2-lite and others18
- RiskBias and discrimination
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。
Paper analysisRead analysis
DefensesarXiv 2610.01170Oct 1
HeadEdit：通过冻结的 unembedding 矩阵校准语言模型行为
提出 HeadEdit，经冻结 unembedding 做随输入变化的 logit 校正
arXiv
2610.01170
Published
Oct 1
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-4B, Gemma-3-4B-IT, Llama-3.2-3B-Instruct and others4
- DefensesInference-time intervention
- RiskInappropriate refusal
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。
Paper analysisRead analysis
InterpretabilityarXiv 2610.06903Oct 1
研究揭示 Transformer 拒答机制中的组件与维度稀疏性
定位写入 Transformer 拒答方向的稀疏组件与残差维度
arXiv
2610.06903
Published
Oct 1
Layer
Model layer
Setting
Models and APIs
Evaluated models
Gemma-2-2B-IT, Qwen3-4B, Phi-4-Mini-Instruct and others4
- RiskInappropriate refusal
Abstract四模型上拒答转向集中于稀疏组件和约一半残差维度，而非均匀编码。
这篇工作把拒答转向从残差流单点干预拆到组件和坐标，看 refusal 方向 r* 是被哪些上游模块组装出来的。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.00531Oct 1
SciSlopBench：评测并缓解 AI 生成论文中的科学灌水
构建 SciSlopBench 评测科学灌水，并用 SciSlopHarness 约束修订
arXiv
2610.00531
Published
Oct 1
Layer
Model layer
Setting
Models and APIs
- RiskHallucination and fabrication
Paper analysisRead analysis
DefensesarXiv 2609.38645Sep 30
用探针引导微调对齐模型且不损失可监控性
用持续更新的探针做微调，降低有害性并保持线性可监测性
arXiv
2609.38645
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Mistral 7B Instruct v0.1, Llama 3 8B Instruct Abliterated, Qwen3-14B
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentMonitors
Abstract持续重拟合的探针可降低有害性、提高诚实性，训练后概念仍能被线性探针读出。
Probe-guided fine-tuning 把探针分数当作唯一对齐信号，用来推动 harmlessness 与 honesty，并检查训练后这些概念是否仍线性可读。
Paper analysisRead analysis

AI Safety Daily

Read the latest issueOne issue per day, updated as events develop

All papers

Attack type

Model risks

Defense type

Impact

System type

Attack surface / entry point

Layer

System component

Attention sees, MLP refuses: tracing the safety pathway from attention to MLP in LLMs

SafeEvo：从拒答回路视角解释大模型安全对齐机制与演化

研究定位 LLaMA-2-7B-Chat 安全敏感参数

The paper proposes a scaling law framework for alignment measured by risk category.

ASCENT：通过一阶最优校准实现安全与效用协同的微调框架

研究者提出读取路由 logits 的 MoE 视觉语言模型多模态安全检测器

VERA：结构化审计发现约 60% 正确漏洞判定伴随虚构推理

Study shows VLM visual grounding refusal rates are on average 31 to 59 percentage points lower by harm domain

研究者提出 CAP 指标与 CSFD 算法，发现越狱通过压制安全特征绕过拒答

EmoRSS：缓解大语言模型由情绪引发的过度拒答

One Token Can Be Enough：用 Prefix Steering 连接提示与激活引导

研究揭示语言模型拒答行为的转向维度差异

OmniConfess：用 Token 级「忏悔」缓解全模态幻觉

论文提出用路由梯度敏感度定位稀疏 MoE 模型中的安全敏感专家

DNAlign：面向 LLM 的动态零空间安全对齐框架

PowerBench 评测 24 个模型在权力转移请求上的偏见

HeadEdit：通过冻结的 unembedding 矩阵校准语言模型行为

研究揭示 Transformer 拒答机制中的组件与维度稀疏性

SciSlopBench：评测并缓解 AI 生成论文中的科学灌水

用探针引导微调对齐模型且不损失可监控性