arXiv: branched self-improving harness optimization beats Meta-Harness by 34.8% on math

rohanpaul_ai · x · 2026-10-04

The arXiv paper "Mixture of Self-Improving Branches For Agent Harness Optimization" (Meta, Duke, UC; Haoyu Dong et al.) proposes:

Background: harness optimization is a practical setting for recursive self-improvement (RSI). Meta-Harness iteratively rewrites the harness with a fixed dev set and proposal policy, channeling search along one trajectory and risking local optima.
Method: organize the improvement process into branches with evolving dev subsets and proposal policies — each branch keeps cases its leading harnesses solve more often than other branches', drops cases solved by every leading harness everywhere, and revises its proposal policy from its own search history. A router selects a dev-selected branch head per new input before execution.
Results: relative gains over Meta-Harness of 34.8% on Olympiad-level math reasoning, 11.6% on Terminal-Bench 2.0, and 3.8% on SWE-bench Lite, with harness selection and router config based solely on dev data.
Conclusion: evolving branch objectives and proposal policies yields complementary harnesses.

More from coding & agent