Reproduce “Heterogeneous Co-Evolution Learning for Tool-Efficient Search Agents with Zero Labels” 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

Qin-Wen Luo, Sheng Ren, Xiang Chen, Rui Liu, Jun Fang, Naiqiang Tan, Jun Wang, Hua Chai and 1 more

ZEST-Search trains a search agent with tasks generated from unlabeled reference questions and retrieved evidence. A proposer targets entities and builds evidence-supported questions, while a solver receives rewards for correct answers and lower task-relative search use. For a Qwen2.5-7B-Base solver, seven-benchmark macro-average normalized exact match rose from 21.8 to 38.2, while mean parsed search calls fell from 1.59 to 1.16; these are benchmark and call-count results, not measured time savings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.