The authors trained LLM-based search agents on synthetic multi-hop question-answering data that includes annotations for the factual evidence needed to answer each question.
Evidence annotations improve step-level credit for deep search agents
EG-GRPO rewards retrieval steps when they first uncover required evidence, while retaining supervision from final-answer accuracy.
Chinese Tech
Sijia Luo · Mengjie Gao · Yuxiao He · Mengshu Sun · Cuiping Li · Hong Chen · +1 more
Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, MOE · Ant Group · Beijing Key Laboratory of Intelligent Database Systems and Applications · Engineering Research Center of Database and Business Intelligence, MOE
Research Digest··3 min read
Luo and colleagues propose EG-GRPO, a reinforcement-learning method that turns evidence annotations in synthetic search data into rewards for individual retrieval steps.
§