Reproduce “Evo$^2$: Evolving the Evolver through Weight-Harness Co-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
Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews
Evo² connects harness evolution with model training: one shared policy proposes and assesses harness edits, while evaluated outcomes train later searches. With the task executor fixed, its held-out Qwen3.8-27B Polyglot pass@1 is 82.1%, versus 62.6% for GEA. A controlled comparison reports 78.0% for the trained evolver versus 61.8% untrained, a 16.3-point gain; this supports a training contribution, not its sole responsibility for gains.
Key insights
- Harness search can supply training experience for the agent that conducts the search, allowing workflow improvements and policy improvements to reinforce each other.
- A shared policy for proposing and reviewing edits can use evaluation results and execution traces to decide whether to revise, combine, or retain a candidate.
- Some harness changes transfer across benchmarks and domains, but results remain dependent on the evaluated tasks and the reliability of their metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.