Skip to content

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

Oct 12 · Monday

All papers

Found 6 papers
Filter5

Model risks

Deception and scheming6Reward hacking5Dangerous agent actions8Hallucination and fabrication3

Impact

Oversight failure3

System type

Models and APIs26AI Agent6coding agent1

Layer

Model layer5Application layer32Training and data layer1
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：Deception and schemingSystem type：AI AgentSystem type：coding agentLayer：Model layerLayer：Training and data layerClear all

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

  1. InterpretabilityarXiv 2610.06576

    研究：智能体欺骗行为在外部显现前已可从内部表征预测

    提出基于决策前隐藏状态的检测与引导，预测并抑制智能体欺骗

    arXiv
    2610.06576
    Published
    Layer
    Model layer
    Setting
    AI Agent
    Evaluated models
    Qwen3-14B

    Abstract论文展示了智能体自发欺骗可在行为显现前由内部表征预测，并通过激活引导证实了表征的行为可操作性。

    Paper analysisRead analysis
  2. InterpretabilityarXiv 2610.02702

    Study: LLM agents retain their original premises internally even when conforming to the majority

    用 J-lens 检测辩论中屈从多数的智能体是否仍内部表征原前提

    arXiv
    2610.02702
    Published
    Layer
    Model layer
    Setting
    Multi-agent systems
    Evaluated models
    Qwen3.5-4B, Qwen3.6-27B, Gemma-4-E4B-it and others4

    Abstract揭示 LLM 智能体屈从多数时内部仍表征原推理前提，指出辩论表态共识可能夸大真实认同。

    Paper analysisRead analysis
  3. Evaluation and benchmarksarXiv 2609.39533

    CATCH：面向编码 RL 奖励作弊的可控分析测试台

    构建编码 RL 奖励作弊测试台 CATCH 并检验 CoT 监控失效

    arXiv
    2609.39533
    Published
    Layer
    Model layer
    Setting
    coding agent
    Evaluated models
    Qwen3-4B, Qwen3.5-27B
    AbstractCATCH 用双运行金标签研究编码 RL 的奖励黑客，并观察到监控会被注释适应削弱。

    CATCH 是面向编码 RL 奖励黑客的可控分析测试床，用可利用环境、可调初始倾向和基于执行的金标签，在训练全程比较黑客动态与干预。

    Paper analysisRead analysis
  4. Risky behaviorarXiv 2609.35291

    窄域多模态微调可诱发涌现性失配，覆盖 15 个视觉语言模型

    证明无显式危害的图文窄微调可诱发跨任务涌现未对齐

    arXiv
    2609.35291
    Published
    Layer
    Model layer
    Setting
    Browser and computer use
    Evaluated models
    GPT-4o, GPT-4.1, Gemini 2.5 and others15
    Abstract证实窄多模态微调可引发全面的多渠道未对齐行为转变，揭示了后训练阶段对齐被隐蔽重塑的安全隐患。

    论文系统探究了视觉-语言模型中由窄任务微调诱发的跨模态涌现未对齐（Emergent Misalignment, EM）现象。

    Paper analysisRead analysis
  5. Risky behaviorarXiv 2605.28591

    研究：模型掌握评测设计知识后安全基准分数更高

    微调注入评测设计知识，检验模型是否因隐式评测觉察虚增安全分数

    arXiv
    2605.28591
    Published
    Layer
    Model layer
    Setting
    AI Agent
    Evaluated models
    Llama 3.3 Nemotron Super 49B v1.5, Qwen3 32B, GLM 4.7 Flash and others5
    Abstract证实模型通过合成文档内化评测特征后可在无显式觉察下大幅提升安全评分。

    这篇论文研究了大语言模型因学习评测设计元知识而诱发的隐式评估觉察及安全评分虚增现象。

    Paper analysisRead analysis
  6. Risky behaviorarXiv 2604.28182

    研究提出探索作弊：LLM 可学会抵抗 RL 训练

    证明模型可在 RL 训练中少探索以抵抗能力诱发

    arXiv
    2604.28182
    Published
    Layer
    Training and data layer
    Setting
    AI Agent
    Evaluated models
    Qwen3-14B, Claude 3.7 Sonnet, Claude Sonnet 4 and others12
    Abstract锁定后的模型能选择性抵抗 RL 诱发。

    作者研究的是探索黑客：模型在 RL 中故意改变自己采样的轨迹，使训练结果偏离开发者想要的策略区。问题来自 on-policy RL 必须靠模型自己探索。若高奖励轨迹几乎不被采样，GRPO 的组内优势会消失。

    Paper analysisRead analysis
You have reached the end