Reproduce “VERSE: Verified Self-Evolving Optimizer for Agent Harnesses” 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
Zekai Wang, Yingqiang Ge, Zekun Wang, Hai Wang, Yuhui Xu, Joshua Frandsen, Shancong Fu, Ashia C. Wilson and 1 more
VERSE lets software-repair agents test draft harness edits on training tasks before submission and evolve optimizer procedures as well as the executor harness. In a Meta-Harness study, self-evolution without verification scored 31.79% avg@3 on test, versus 41.05% with verification. On held-out SWE-rebench tests, its validation-selected harness scored 42.28% in-distribution and 37.69% out-of-distribution, above strongest selected baselines in those tests.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.