Reproduce “RRSI: Regularized Recursive Self-Improvement of 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
Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhuang, Yoonho Lee, Chengsong Huang, Han Yu and 6 more
RRSI evolves a fixed language model’s harness while keeping its backbone unchanged. It regularizes how edits are proposed and selected, using search history, noise-adjusted floors, cost checks, and pruning. In Claude Opus 4.8’s workspace experiments, it reaches a 43.6 OOD average across three benchmarks, versus 39.7 for the unevolved harness; this shows transfer on tested splits, not universal reuse.
Key insights
- Harness evolution is itself an adaptive optimization process: repeatedly selecting against the evolve set can produce overfitting even when model weights remain fixed.
- Regularizing both proposal generation and candidate acceptance improves transfer in the reported ablations; removing either set of controls lowers the out-of-distribution average.
- An evolve-set score is not a reliable stand-alone measure of harness quality: some prior methods score well on the search benchmark but show little transfer to other tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.