RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
The paper says agent self-improvement works better when the harness is constrained against benchmark memorization.
RRSI regularizes how an agent harness evolves, limiting bundled edits, encouraging less-explored paths, and screening changes that look too benchmark-specific. Its selector also prunes edits judged too small, too costly, or no longer useful. Across eight coding, workspace, and engineering-design benchmarks, the authors report gains up to 14.1 points on the target split and up to 4.7 points out of distribution. The resulting harness used 30% fewer policy tokens than unregularized evolution. ArXiv · AI/CL/LG's note
RRSI regularizes how an agent harness evolves, limiting bundled edits, encouraging less-explored paths, and screening changes that look too benchmark-specific. Its selector also prunes edits judged too small, too costly, or no longer useful. Across eight coding, workspace, and engineering-design benchmarks, the authors report gains up to 14.1 points on the target split and up to 4.7 points out of distribution. The resulting harness used 30% fewer policy tokens than unregularized evolution. ArXiv · AI/CL/LG's note
score 4