arXiv · AI· Pu Li, Zhigang Lin, Qiang Wu, Yongxuan Lv, Fei Wang, Shan You·· 2026-10-14AI 评分28
Reward as an Agent 与 DynDiff-GRPO 提升具身世界模型 RL 性能
Reward as An Agent for Embodied World Models
AI 导读
该研究提出 Reward as an Agent 框架与 DynDiff-GRPO 方法,通过可靠验证和动态感知采样多元化,在多个开源具身世界模型上显著降低 reward hacking 并提升准确率。
来源:arXiv · AI · arxiv.org