Meta paper: branching self-improving harness search lifts Olympiad math accuracy to 62%

rohanpaul_ai · x · 2026-10-04

A new paper from Meta, Duke and University of California introduces a branched approach to self-improving agent harness optimization:

A harness is the code around an LLM controlling tools, retrieval and self-checks. Prior work (Meta-Harness) rewrites it iteratively against a fixed dev set, so search converges to one path and risks local optima.
The new method splits search into branches, each keeping the dev cases it solves better than others and writing its own notes on what worked — in math, one branch learned answer verification while another learned full derivations.
A router selects the best-fit branch head per new input before execution.
Results: on Gemini 3 Flash, Olympiad-level math accuracy rose from 46.0% to 62.0% (34.8% relative gain), plus +11.6% on Terminal-Bench 2.0 and +3.8% on SWE-bench Lite, beating Meta-Harness in all 4 settings.

Practical takeaway: when auto-tuning agents, keep several specialized harnesses and route between them rather than betting on a single winner.

More from coding & agent