arXiv 2609.38847 · cs.LG · 2026-09-30
Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics
Maoqi Liu, Junwei He, Bowen Zhang, Feiran Li, Wentao Ma, Rongyi Lin, Shuhan Zhong, Quan Fang
ByteDance
在最新榜 #160 本站收录于 2026-10-21
一句话要点
指出基于评分细则的 RL 用加权求和会被“刷分”（以临床问诊为例），并提出协议级细则加以缓解。
Shows weighted-sum rubric rewards can be hacked (e.g. in clinical consultation) and proposes protocol-level rubrics to fix it.
阅读 arXiv 原文 PDF Hugging Face 讨论 代码仓库
主题与学科
同主题：强化学习
- Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation ▲218
- Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight ▲189
- LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches ▲152
- Sharpening Tax in Post-Training ▲111
- RULER: Instance-aware Rubric Rewards for SVG Generation ▲104
数据说明
元数据来自 arXiv API（已核实）；点赞来自 Hugging Face 每日论文；引用来自 Semantic Scholar。最近刷新 2026-10-22。
要点由本站根据摘要自写，不转载摘要原文，请阅读原文。