Overconfident Errors Need Stronger Correction: Asymmetric Confidence Penalties for Reinforcement LearningDate: 2026-02-27Fetched: 2026-02-28T01:46:40.445633+00:00AuthorsYuanda Xu, Hejian Sang, Zhengze Zhou, Ran He, Zhipeng WangLinksHFarXivPDF1Abstract中文摘要English可验证奖励的强化学习因统一错误惩罚而面临推理多样性降低的问题，对此，一种置信度感知的不对称错误惩罚方法通过基于 rollout 置信度动态调节优势加以解决。
