HuggingFace Daily Papers·2026-10-02 08:00· 3 天前AI 评分42

MetaRubric：面向基于评分标准的强化学习，用证据感知奖励解决 Vacuous Credit 问题

MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning

AI 导读

MetaRubric 通过交替进行证据感知的策略优化与响应引导的评分标准自适应，解决评分模型在响应缺失所需信息时仍给出高分的 Vacuous Credit 问题。