h1 paper: minimal Lean loop hits SOTA on PutnamBench and solves 8 open conjectures
rohanpaul_ai · x · 2026-11-02
The arXiv page (2610.35962) for the Meta/NYU paper h1: A Taste of the Bitter Lesson in Automated Theorem Proving, authored by Skander Moalla, Rémi Munos, Julia Kempe and others.
- h1 is a simple harness providing an efficient Lean 4 interface, iterating via verifier feedback and self-reflection without imposed decomposition or search
- State of the art on PutnamBench, FATE-X and FormalConjectures100-Solved; settles eight open problems in FormalConjectures, often with a single pass@1 attempt
- Matches or outperforms specialized workflows like Goedel-Architect and AlphaProof Nexus at matched models
- Ablations show simplicity alone isn't enough: an efficient Lean interface largely determines cost and solve rate
More from Research
- AAAI NeuroAI workshop 2nd edition opens submissions, papers due Nov 20 — tyrell_turing · 2026-11-02
- Kaggle and IBM launch ScarfBench: 204 tasks testing coding agents on enterprise Java migration — gabriberton · 2026-11-02
- ThreeJS-3D-Scenes dataset trends on Hugging Face with text, image and 3D modalities — ccore · 2026-11-02
- Testing randomized algorithms: 14.5M Monte Carlo runs can miss bugs that exact tracing proves instantly — xtimv · 2026-11-02
- Karpathy: my arXiv tool was too much work, alphaXiv is the one to use — vhanagwal · 2026-11-02
- An AI security agent that stops attacks by breaking the app should count as failure — krishnan · 2026-11-02