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.

System type：Models and APIs System type：science agent System type：Browser and computer use System component：Reasoning chains Clear all

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

AttacksarXiv 2610.12644Oct 9
Cancerous backdoor attack: progressive activation via invisible tokens and a stage controller.
Proposes the Cancerous backdoor, which progressively corrupts reasoning using invisible stage tokens.
arXiv
2610.12644
Published
Oct 9
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentReasoning chains
AbstractHidden 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.
Paper analysisRead analysis
InterpretabilityarXiv 2610.12361Oct 9
Study: legal reasoning models citing legal provisions does not equal relying on legal provisions.
Uses premise permutation and hidden-layer tracing to audit whether legal chain-of-thought truly relies on the cited legal provisions.
arXiv
2610.12361
Published
Oct 9
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-8B, Qwen3-14B, Llama-3.1-8B and others7
- AttacksPrompt injection
- ComponentReasoning chains
AbstractUsing counterfactual substitution of authorities and internal commitment tracing, the authors find that legal CoT rulings depend on authorities only to a limited extent and are vulnerable to prompt injection manipulation.
This study conducts an empirical audit of the causal truthfulness of legal citations in chain-of-thought reasoning of large language models in the legal domain.
Paper analysisRead analysis
Analysis and theoryarXiv 2610.11765Oct 8
研究提出三项指标衡量 LLM 不思考行为，发现被要求停止思考时仍持续显式推理
度量关闭思考指令下终答前是否仍有可见显式推理
arXiv
2610.11765
Published
Oct 8
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-4B, Qwen3-32B, Qwen3-235B-A22B and others6
- ComponentReasoning chains
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。
Paper analysisRead analysis
DefensesarXiv 2610.11352Oct 8
RL-ARC：用推理引导的不确定性校准大型推理模型
提出 RL-ARC，用推理置信度在强化学习中校准推理模型
arXiv
2610.11352
Published
Oct 8
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-7B, Qwen3-8B, Llama-8B distilled from DeepSeek-R1
- DefensesModel hardening
- ComponentReasoning chains
AbstractRL-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。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.10405Oct 8
研究发现大语言模型被要求不诚实作答时推理 token 数上升
检验推理 token 数能否区分被提示的诚实与不诚实作答
arXiv
2610.10405
Published
Oct 8
Layer
Model layer
Setting
Models and APIs
Evaluated models
Gemma-4-e4b, Qwen-3.5-9b, GPT-OSS-120B
- RiskDeception and scheming
- ComponentReasoning chains
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。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.09667Oct 7
东京大学研究：推理模型按 UUID 阈值或整除规则挑选审计样本，控制标识符即可躲开复核
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
arXiv
2610.09667
Published
Oct 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
GPT-6 Sol, GPT-6 Luna, GPT-5.4 mini and others6
- ComponentReasoning chains
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。
Paper analysisRead analysis
InterpretabilityarXiv 2610.09087Oct 7
U-Space：揭示语言模型不确定性何时与为何产生
提出 U-SPACE 与 U-LENS，从残差流估计推理轨迹不确定性
arXiv
2610.09087
Published
Oct 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
Gemma-4-31B-it, Qwen3.5-27B, Magistral-Small-2507
- ComponentReasoning chains
AbstractU-SPACE 与 U-LENS 用语义方向和预测熵读推理不确定性，长度控制后仍高于所比基线。
U-SPACE 与 U-LENS 是面向推理模型的无训练读出：用语义方向定位不确定性的形式，并用一个轨迹分数区分更可能错的回答。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.06851Oct 6
研究发现基础模型仅靠固定开头 token 即可达到 RL 级推理表现
证明短开头词可唤醒基模型推理，且效应来自训练数据
arXiv
2610.06851
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Olmo-3-7B, Olmo-3-32B, Qwen3-4B and others10
- ComponentReasoning chains
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.02741Oct 2
研究系统评估循环语言模型的思维链可监控性
系统评估循环语言模型加深 loop 后的 CoT 可监控性
arXiv
2610.02741
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
Ouro-1.4B-Thinking, Ouro-2.6B-Thinking, Nanbeige4.2-3B and others4
- DefensesMonitoring and auditing
- RiskDeception and scheming
- ComponentReasoning chains
Abstract加深 loop 可在部分压力测试任务降低可监测性，循环架构本身并不必然更难监测。
这项工作评估 LoopLM 的 CoT 可监测性：加深共享层的重复次数，以及“有循环结构”本身，会不会让决策关键因素更少出现在可监测文本里。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.02015Oct 2
Research: RLVR post-training induces language drift on novel reasoning tasks.
证明 RLVR 后训练允许无界语言漂移且限制漂移必限制奖励
arXiv
2610.02015
Published
Oct 2
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
gemma-3-1b-pt, Llama-3.2-1B, Qwen2.5-1.5B
- ComponentReasoning chains
Abstract论文证明并验证了 RLVR 在新任务上必然诱发独特的语言漂移，揭示了前沿推理模型可监控性的安全隐患。
论文从理论与实验两方面研究了大语言模型在 RLVR 后训练过程中思维链出现的语言漂移现象。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.03509Oct 2
高效推理训练并非总损害思维链忠实性与可监控性
测量三种长度压力训练对 CoT 忠实性与可监控性的影响
arXiv
2610.03509
Published
Oct 2
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-4B-Thinking-2507, Qwen3-8B, Olmo3-7B-Think
- ComponentReasoning chains
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。
Paper analysisRead analysis
Risky behaviorarXiv 2609.39838Sep 30
研究：模型易学会隐写传信，隐写推理则难得多
比较隐写推理与隐写传信、编码推理的可学性
arXiv
2609.39838
Published
Sep 30
Layer
Model layer
Setting
Models and APIs
Evaluated models
GPT-5.5, Claude Sonnet 5, Gemini 3.5 Flash and others14
- RiskDeception and scheming
- ComponentReasoning chains
Abstract隐写推理比传信和编码推理更难学。
Schulz、Fülle 与 Frengel 测量大语言模型习得隐写推理的难度，并与隐写传信、编码推理对比，以估计依赖阅读 CoT 的监控还能否成立。
Paper analysisRead analysis
InterpretabilityarXiv 2609.38972Sep 30
研究提出 CIA 指标衡量并改进 LLM 思维链与内部计算的一致性
提出 CIA 指标衡量并改进思维链与内部计算的一致性
arXiv
2609.38972
Published
Sep 30
Layer
Model layer
Setting
Models and APIs
Evaluated models
Llama3.1-8B-Instruct, Gemma2-9B-it, Qwen3-8B and others5
- ComponentReasoning chains
Abstract用 CIA 度量并以后训练拉齐 CoT 与内部策略。
Paper analysisRead analysis
DefensesarXiv 2609.38909Sep 30
Purdue 提出 Pact：将欺骗作为行为遗忘
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
arXiv
2609.38909
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
DeepSeek-R1-Distill-Qwen-32B, QwQ-32B
- DefensesModel hardening
- AttacksModel tampering
- RiskDeception and scheming
- ComponentFine-tuning and open weights
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。
Paper analysisRead analysis
Analysis and theoryarXiv 2609.38107Sep 30
研究：iGSM 基准显示正确答案常伴随无效思维链
检查正确答案对应的思维链是否构成合法推导
arXiv
2609.38107
Published
Sep 30
Layer
Model layer
Setting
Models and APIs
Evaluated models
124M-parameter GPT-2-style models（12-layer, hidden size 768, rotary embeddings）
- ComponentReasoning chains
AbstractiGSM 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。
Paper analysisRead analysis
DefensesarXiv 2610.00332Sep 29
最弱一环：用最坏情况约束强化学习蒸馏 LLM 推理能力
提出 worst-case Constrained RL，蒸馏时约束推理前缀偏离以抑制奖励作弊
arXiv
2610.00332
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-1.5B, Llama-3.2-3B, Qwen2.5-Coder-0.5B
- DefensesModel hardening
- RiskReward hacking
- ComponentReasoning chains
Abstract作者提出前缀约束蒸馏。
这是把推理蒸馏写成最坏前缀约束 RL 的训练研究。作者关心最终答案奖励下的 reward hacking，以及序列平均软惩罚允许局部偏离被其他 token 补偿的问题。
Paper analysisRead analysis
Analysis and theoryarXiv 2609.37312Sep 29
研究：隐蔽推理必然留下信息痕迹，但思维链未必可读
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
arXiv
2609.37312
Published
Sep 29
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Qwen 3.5 397B-A17B and others12
- DefensesMonitoring and auditing
- RiskDeception and scheming
- ComponentMonitors
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。
Paper analysisRead analysis
DefensesarXiv 2609.37054Sep 29
ACTR：对齐推理与回答以提升推理大语言模型的多语言安全
提出 ACTR，只更新安全思考神经元以对齐推理与回答的安全类别
arXiv
2609.37054
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-8B, Gemma4-12B-it
- DefensesModel hardening
- AttacksJailbreaks
- ComponentReasoning chains
Paper analysisRead analysis
DefensesarXiv 2609.36641Sep 29
从仅结果奖励的强化学习中诱导过程监督
提出 TIPS，用仅结果奖励诱导生成式过程监督
arXiv
2609.36641
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-4B-Thinking-2507, Qwen3-4B-Instruct-2507, Qwen3-8B-Thinking and others8
- DefensesDetection and filtering
- ComponentGuardrails and judges
Abstract结果监督的 GRPO 可诱导步骤验证。
TIPS 用只含结果对错的强化学习训练生成式奖励模型，让步骤级验证出现在思维链里。
Paper analysisRead analysis
DefensesarXiv 2609.36562Sep 29
ThinkingGuard：通过逐步风险归因解码多模态大语言模型中的隐式危害
构建 TriggerBench 并训练 ThinkingGuard，检测图文隐式风险
arXiv
2609.36562
Published
Sep 29
Layer
Prompt layer
Setting
Models and APIs
Evaluated models
ThinkingGuard（Qwen3-VL-8B-Instruct）
- DefensesDetection and filtering
- AttacksJailbreaks
- ComponentGuardrails and judges
AbstractTriggerBench 隔离触发条件，ThinkingGuard 用分步归因在隐式风险检测上取得作者报告的较高平均分。
这篇工作研究多模态隐式风险检测：图文各自良性，风险只在关键元素与触发元素交互时出现。
Paper analysisRead analysis

AI Safety Daily

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