Meta paper: branching harness search lifts Olympiad math +34.8% over Meta-Harness
dair_ai · x · 2026-10-02
A new Meta paper improves on Meta-Harness-style agent harness optimization.
- The baseline uses one development set and one proposal policy, so every edit follows a single path and can get stuck in a local optimum.
- The new approach splits the search into branches: each branch keeps dev cases its harnesses solve better than other branches, drops cases every branch already solves, and rewrites its own proposal policy from its history. A router picks one branch's harness per new input before it runs.
- Gains over Meta-Harness: +34.8% on Olympiad-level math, +11.6% on Terminal-Bench 2.0, +3.8% on SWE-bench Lite.
- Harness selection and the router rely only on development data.
More from coding & agent
- Three reasons vibe-coded software is still far from production grade, with ReactBench data — aidenybai · 2026-10-02
- GitHub Copilot adds computer use to control desktop apps in public preview — PaulShellDev · 2026-10-02
- Dev builds her first Game Boy-style game entirely with OpenAI agents — craigsdennis · 2026-10-02
- When the Trace Looks Fine but the Agent Output Is Wrong — Sensitive-Parsnip-12 · 2026-10-02
- iPad + Tailscale + Codex is this dev's new favorite way to work outside the office — flavioAd · 2026-10-02
- Will Cloud AI Agents Need Residential IPs? Amazon Blocked Meta's Muse — SignificantFail3632 · 2026-10-02