Meta FAIR's h1: a minimal Lean feedback loop beats hand-designed theorem-proving workflows, hits SOTA
burny_tech · x · 2026-10-13
A Meta FAIR / EPFL / NYU paper asks the bitter-lesson question for formal math: instead of hand-designing proof workflows (decomposition, planning, agent orchestration), what if a frontier model just gets a good Lean 4 interface and verifier feedback?
- h1 is a minimal harness: no imposed decomposition or search structure, just efficient Lean interaction plus self-reflection.
- SOTA on PutnamBench, FATE-X, and FormalConjectures100-Solved; settles 8 open conjectures in FormalConjectures.
- At matched models it matches or beats specialized workflows like Goedel-Architect and AlphaProof Nexus; a single long attempt (pass@1) often scales better than many retries.
- Ablations (vs Claude Code, AxProverBase) show simplicity isn't enough: an efficient Lean interface largely determines cost and solve rate.
- Recommendation: evaluate frontier models with efficient execution feedback before prescribing workflows.
Related event: Meta FAIR's h1 Sets New SOTA in Automated Theorem Proving(2 posts)→
More from coding & agent
- Running tens of millions of experiments on useless programs with LLMs — bushibuilds · 2026-10-13
- Xiryn: Open-Source Personal Agent with Docker Sandboxes and Telegram Control — Scobleizer · 2026-10-13
- Witness Tracks Everything You Do and See So Your Agents Always Have Context — Scobleizer · 2026-10-13
- MecAgent 2.0 launches as its maker claims 'CAD is about to get solved' — Scobleizer · 2026-10-13
- Dev finds ~25% of AI-generated code introduces regressions in .NET 8+ — lucasmeijer · 2026-10-13
- Investor says coding agents left most software firms badly overstaffed, cut now or be disrupted — doodlestein · 2026-10-13