Reproduce “MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution” 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
Zhiwei Shang, Yu Huo, Mingrong Gong, An Yan, Zikun Qu, Junhao Dong, Bryan Kian Hsiang Low, Chenglin Wu and 1 more
MESH-Harness improves agent code by combining interchangeable implementations for harness jobs. It uses a contextual bandit to select combinations from candidate pools, then uses validation scores and failure traces to guide selection and revise pool entries. With 40 candidate evaluations, it scores 56.33 on Text versus Meta-Harness’s 50.63; gains and cost reductions are limited to the tested tasks and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.