Meta and NYU paper: simple agentic loop h1 matches hand-built theorem-proving workflows
rohanpaul_ai · x · 2026-11-02
A Meta/NYU paper, h1: A Taste of the Bitter Lesson in Automated Theorem Proving, shows a minimal agentic loop can match or beat hand-designed theorem-proving workflows.
- h1 gives the model an efficient Lean 4 interface and keeps iterating via verifier feedback and self-reflection, with no imposed planning or decomposition
- At matched models, h1 beat Goedel-Architect on PutnamBench 598 to 597 problems, at $1.08 vs $1.47 per problem
- With Opus 4.8, h1 solved 83 of 100 hard algebra problems vs 63 for Claude Code
- Takeaway echoing the Bitter Lesson: gains come from a simple loop plus compute; the efficiency of the Lean interface largely determines cost and solve rate
More from coding & agent
- Stardock CEO uses AI agents to animate game characters without rigging, runs 101 agents — draginol · 2026-11-02
- Stardock CEO runs 101 AI agents like a studio: inside his Clairvoyance workflow — draginol · 2026-11-02
- DeepSeek's SuperAgent architecture questioned over risky multi-worker event design — teortaxesTex · 2026-11-02
- Dev wires watch button to record, upload, and let Grok Bot delegate voice tasks — johnlindquist · 2026-11-02
- Matt Pocock: Redux's single-source-of-state idea may be exactly what agents need — mattpocockuk · 2026-11-02
- Game2Player: builders and players co-develop 568 games to train a 9B visual agent — ManlingLi_ · 2026-11-02