Reproduce “JoMoHa: Joint Model--Harness Optimization with Reinforcement Learning” 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
Xudong Wu, Guangchen Lan, Mingqi Yuan, Jiayu Chen
JoMoHa trains an Engineer and Solver together using shared task outcomes to update each policy from its own generated tokens. It evaluates on ALFWorld, WebShop, and DBBench, reporting success rates of 97.0%, 45.8%, and 68.0%, respectively. Its equally weighted mean is 70.3%, 7.2 percentage points above WHALE among evaluated methods, not establishing universal superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.