Reproduce “Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Taewhoo Lee, Minju Song, Jungwoo Park, Chanwoong Yoon, Junha Jung, Hyeon Hwang, Sanghoon Lee, Minbyul Jeong and 1 more
SARL trains search-augmented question answering agents to track which question constraints retrieved evidence resolves and what remains missing. During training, it reviews search trajectories, assesses evidence coverage, and uses incomplete assessments to guide further searches; the assessment is not used at inference. Across ten benchmarks, it averages 0.429 versus ASearcher’s 0.345, under a 32-turn maximum.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.