Reproduce “SOAR: Source-Aware Reinforcement Learning with Precise Credit Assignment for Search Agents” 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
Shengze Huang, Zihao Tang, Jinluan Yang, Xin Yu, Zhenghao Lin, Xiao Liu, Zengxuan Wen, Yeyun Gong and 6 more
SOAR trains search agents to distinguish internally recalled facts from retrieved evidence and to match answers to a declared source. It elicits the two information streams separately and rewards source-answer consistency alongside answer correctness. On Qwen2.5-3B-Base, it reaches 0.414 mean exact-match accuracy across seven benchmarks and a 4.05% judged fabrication rate; this grounding measure does not establish truth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.