Reproduce “Self-Evolving Coding Agents via Harness-Policy Optimization” 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
Zhepei Wei, Wenlin Yao, Baolin Peng, Danqing Wang, Qianhui Wu, Hao Cheng, Yu Meng, Jianfeng Gao
Self-HPO alternates harness search with reinforcement learning to improve a coding agent. In each round, its current policy examines unresolved-task failures and proposes harness changes; the policy is then trained under the selected harness, and its remaining failures guide the next search. On three benchmarks, average resolve rate rises from 57.8% to 64.8%, including transfer to two out-of-distribution benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.