SARL adds a review phase to the usual cycle of reasoning, searching and answering during training.
Training search agents to assess missing evidence improves research performance
Across ten benchmarks, the authors report stronger retrieval-based answers after training models to review incomplete searches, without requiring explicit self-assessment at inference time.
Big Tech
Taewhoo Lee · Minju Song · Jungwoo Park · Chanwoong Yoon · Junha Jung · Hyeon Hwang · +3 more
Amazon · Korea University · AIGEN Sciences · University of Massachusetts Amherst · Meridian Intelligence Global Inc
Research Digest··2 min read
Lee and colleagues introduce Self-Assessment Reinforcement Learning (SARL), a training framework that teaches language-model search agents to recognize gaps in their retrieved evidence.
§