Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
Reward hacking · Papers
Found 6 papersShow that reward hacking can propagate through code, memory, and plans in multi-agent coding systems
- Published
- 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.
提出 TESTJACK，动态审计编码基准通过判定是否满足任务要求
- Published
- Evaluated models
- Claude Opus 4.6, Claude Opus 4.7, Claude Opus 4.8 and others7
用犯罪学框架测试编码智能体压力下的奖励作弊与言行脱节
- Published
- Evaluated models
- Claude Opus 5.5, Claude Sonnet 5.5, GPT-5.6 Sol and others7
Abstract从犯罪学视角揭示 AI 奖励作弊规律，发现言行脱节且规范澄清能消除违规。
检验低监控读数能否证明代码生成中的奖励作弊已受控
- Published
- 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。
发布 SWE-Bench Pro Verified，隔离答案通道并修订任务以抑制奖励破解
- Published
- Evaluated models
- GPT-5.6-Sol, Kimi-K3, GLM-5.3 and others7
提出 HVTB，在终端编程任务植入蜜罐以测量奖励作弊
- Published
- 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。
You have reached the end