The authors analyzed Search-R1, where multiple agent rollouts are scored by final-answer correctness and the resulting group-relative advantage is assigned to every generated token.
Separating search and answer credit improves accuracy and evidence grounding
Grounded Credit Assignment rewards evidence retrieval and answer use separately, improving exact match across seven question-answering benchmarks.
Chinese Tech
Qirui Liu · Hao Chen · Yi Wen · Weijie Shi · Yao Zhao · Xianhe Chen · +4 more
South China University of Technology · FiT, Tencent · City University of Hong Kong · The Hong Kong University of Science and Technology · Nanjing University
Research Digest··3 min read
Liu and colleagues show that reinforcement learning based only on final-answer correctness can make search agents more accurate while making their answers less grounded in retrieved evidence.
§