ARBOR: Online process rewards improve LLM search agents

Arxiv CS.CL (NLP/language models)

ARBOR introduces a reusable rubric buffer to provide online process-level rewards for LLM-based search agents, addressing the degeneration of outcome-only reward on outcome-homogeneous groups. The method enables finer-grained supervision during the search process.

reward-modeling

Arxiv CS.CL (NLP/language models)·Zheng LiuLongxiang ZhangXintong WangZhiang XuShaoxiong ZhanXin ShanWen HuangTao DaiShu-Tao XiaChengfu HuoLiang Ding

8 days ago