Skip to content

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

Oct 12 · Monday

All papers

Found 5 papers
Filter5

Model risks

Reward hacking3

Defense type

Model hardening103Detection and filtering13Monitoring and auditing5

Impact

Oversight failure5

System type

Models and APIs5coding agent0Browser and computer use0

Layer

Model layer17Application layer17Training and data layer5

System component

Guardrails and judges3Monitors3
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.
Defense type：Monitoring and auditingSystem type：Models and APIsSystem type：coding agentSystem type：Browser and computer useLayer：Training and data layerClear all

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

  1. DefensesarXiv 2610.11012

    Redwood Research 提出蒸馏双重困境：用蒸馏让模型坦白隐藏行为或只继承能力

    提出 DFI 与 DFC 蒸馏，定罪隐瞒失配的教师或安全提取能力

    arXiv
    2610.11012
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    Llama-3.3-70B-Instruct (AuditBench SFT/KTO organisms), Llama-3.3-70B-Instruct (DFI distilled student), Llama-3.1-8B-Instruct (DFI distilled student) and others4

    Abstract概括了论文针对隐瞒失配模型提出的 DFI 与 DFC 机制、主实验发现及适用边界。

    Paper analysisRead analysis
  2. DefensesarXiv 2610.02557

    新 AI 辩论协议实现实例级最优：最坏情况正确且诚实为占优策略

    提出 FBS Debate 协议，使诚实为占优策略且最坏情况正确

    arXiv
    2610.02557
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    AbstractFBS 辩论用 fbs 约束敏感分解，给出最坏情况占优策略，并匹配黑盒查询下界。

    FBS Debate 是一套可扩展监督的辩论规则：两名计算方争论一个可递归分解的是非问题，人类只在最后选出的子问题上做黑盒判断。要处理的缺口是，prover-estimator debate 虽能覆盖稳定分解，但只在输入分布的平均意义下成立，而且 Bob 的获胜策略一般要跟着 Alice 的作弊策略变。作者要在同一问题类上，让每个错误实例都能被抓住，并让诚实且正确成为双方的占优策略。

    Paper analysisRead analysis
  3. DefensesarXiv 2609.04665

    HackProbe：面向自演化语言模型奖励作弊的黑盒检测与免疫方法

    提出 HackProbe，黑盒检测并免疫自演化中的奖励作弊

    arXiv
    2609.04665
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Abstract黑盒探针检测自演化奖励黑客，并以带宽受限重选回收真实能力。

    HackProbe 是挂在自演化循环上的黑盒监控器，用来发现奖励黑客并改选候选。

    Paper analysisRead analysis
  4. Risky behaviorarXiv 2606.04923

    CHERRL：可复现、分析与检测评分标准强化学习中的奖励作弊

    提出 CHERRL 注入裁判偏差以复现奖励作弊，并用 RHDA 检测训练起点

    arXiv
    2606.04923
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    Qwen3-4B, Qwen3.5-plus, Qwen3.5-397B-A17B and others4
    AbstractCHERRL 注入已知偏差以观察奖励黑客，并用于分析偏差和评测起点检测器 RHDA。

    CHERRL 是 rubric-based RL 的可控环境，用来复现奖励黑客、观察奖励分歧，并给检测器提供起点参照。

    Paper analysisRead analysis
  5. DefensesarXiv 2602.01750

    Adversarial Reward Auditing: Proactively Detecting and Mitigating Reward Hacking

    提出 ARA，读取奖励模型激活并门控 RLHF 奖励以缓解奖励作弊

    arXiv
    2602.01750
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    Llama-2-7B, Llama-3.1-8B, Qwen-2.5-7B
    AbstractARA 用奖励模型内部表示检出利用，并在 RLHF 里按 Auditor 概率压低代理奖励。

    ARA 是一套两阶段的 RLHF 审计：先对冻结奖励模型找出利用回复，再在策略优化时按检出概率压低奖励。

    Paper analysisRead analysis
You have reached the end