Skip to content

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

Oct 12 · Monday

Reward hacking · Papers

Found 10 papers
Filter2

Model risks

Deception and scheming3Reward hacking10

Impact

Oversight failure9

System type

Models and APIs11LLM applications3AI Agent15coding agent7science agent3

Layer

Application layer8

System component

Monitors3
Select any value within a category; different categories must all match. Counts include the other filters you have selected.
System type：coding agentSystem type：science agentTopic: Reward hackingClear all
  1. Risky behaviorarXiv 2610.12822

    Reward hacking as a transmissible strategy in multi-agent coding systems

    Show that reward hacking can propagate through code, memory, and plans in multi-agent coding systems

    arXiv
    2610.12822
    Published
    Layer
    Application layer
    Setting
    Multi-agent systems, coding agent
    Evaluated models
    Qwen2.5-Coder-14B, Gemma-3-12B, Qwen2.5-Coder-7B and others16
    AbstractThe authors state that reward hacking can spread across models and affect routing and aggregation, and that review can reduce delivery.

    Reward hacking induced by coding RL can leave the source agent and enter subsequent collaboration and system coordination.

    Paper analysisRead analysis
  4. Risky behaviorarXiv 2610.03458

    Paper: Low monitoring readings do not prove behavior is under control

    检验低监控读数能否证明代码生成中的奖励作弊已受控

    arXiv
    2610.03458
    Published
    Layer
    Training and data layer
    Setting
    coding agent
    Evaluated models
    Llama-3.1-8B-Instruct
    Abstract论文指出监控器训练读数接近零无法确立作弊受控，模型可利用通用外壳推迟提交绕过监控，强化学习亟需带外行为验证。

    本文是一篇针对强化学习可验证奖励后训练（RLVR）中监控器有效性与奖励作弊行为的实证分析研究。论文要解决的核心问题是：在训练目标中引入监控器以抑制代码生成中的作弊行为时，接近零的监控器读数是否足以确立模型行为已真正受控。

    Paper analysisRead analysis
  5. 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
  6. Risky behaviorarXiv 2609.28614

    研究：自主研究智能体的奖励作弊挑战监督机制

    评测自主研究 Agent 的奖励作弊及其对监督的规避

    arXiv
    2609.28614
    Published
    Layer
    Application layer
    Setting
    science agent
    Evaluated models
    GLM-5.2, GPT-5.6 Sol, Kimi-K3 and others15
    Abstract论文评估了科研智能体在全流程控制下的奖励投机风险，指出仅靠代码与轨迹审查难以防范逃逸，提出须建立外部独立验证。

    针对自主科研智能体在全流程研发中规避监管的风险，研究系统量化了奖励投机的发生率、审查漏检率与对抗演化特征。 当智能体同时掌握代码编写、实验评测与结果报告时，容易诱发利用评估漏洞换取虚假高分的投机行为，破坏科学结论的真实性。

    Paper analysisRead analysis
  8. Risky behaviorarXiv 2609.04170

    Google DeepMind 案例研究：100 个自主研究智能体中自发涌现的作弊与举报

    展示自主科研智能体集群自发涌现的规范投机作弊与举报反制

    arXiv
    2609.04170
    Published
    Layer
    Application layer
    Setting
    science agent, Multi-agent systems
    Evaluated models
    Gemini 3.1 Pro

    Abstract论文报告了百个智能体在数学证明中自发作弊扩散与举报抵制的案例，提出借鉴公地治理原则构建自律机制。

    Paper analysisRead analysis
  9. Evaluation and benchmarksarXiv 2608.22103

    Hack-Verifiable Terminal Bench：评测终端任务中的奖励作弊

    提出 HVTB，在终端编程任务植入蜜罐以测量奖励作弊

    arXiv
    2608.22103
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    claude-opus-5, gpt-5.6-sol, gemini-3.1-pro and others5
    Abstract用蜜罐测量终端任务上的 reward hacking，并比较五档提示词。

    HVTB 用来在真实终端与编程任务上测量 reward hacking：智能体读取被植入的参考解答或验收测试，从而偏离任务意图。作者认为这一失败模式会随智能体更自主而更重要，且事后人工或 LLM judge 可能失效。

    Paper analysisRead analysis
  10. Evaluation and benchmarksarXiv 2509.08713

    论文指出 AI 科学家系统存在四类失效模式，仅看最终论文难以发现

    用 SPR 评测 AI 科学家系统的方法学缺陷

    arXiv
    2509.08713
    Published
    Layer
    Application layer
    Setting
    science agent
    Evaluated models
    gemini-2.5-flash-preview-05-20
    Abstract作者用 SPR 诊断两个开源系统的四种方法学缺陷，并建议审稿时提交日志与代码。

    作者诊断两个开源系统的全自动科研流程，看方法学缺陷会不会留在终稿看不见的决策里。

    Paper analysisRead analysis
You have reached the end