Simple Lean 4 harness h1 matches specialized prover workflows
Original title: h1: A Taste of the Bitter Lesson in Automated Theorem Proving
With matched models, a single attempt plus verifier feedback solved eight open FormalConjectures problems and led three formal benchmarks.
The paper introduces h1, a harness that gives a model an interface to Lean 4 and continues via verifier feedback and self-reflection without a fixed decomposition or search structure. The authors say h1 reaches state-of-the-art results on PutnamBench, FATE-X, and FormalConjectures100-Solved and settles eight open problems in FormalConjectures. At matched models it matches or beats Goedel-Architect and AlphaProof Nexus, and pass@1 often solves open problems. Ablations against Claude Code and AxProverBase attribute cost and solve rate largely to the Lean interface.
Unconfirmed: Benchmark leadership and the eight solved problems are the authors' claims from their comparisons.
- h1 does not impose planning, decomposition, or a search structure.
- It reports state-of-the-art results on PutnamBench, FATE-X, and FormalConjectures100-Solved.
- Eight open FormalConjectures problems were settled.
- Authors say an efficient Lean interface largely determines cost and solve rate.
- Products
- Lean 4
From the original article
1 min read at huggingface.co
While progress in informal mathematics has come largely from scaling the chain-of-thought and agentic reasoning of frontier language models, formal theorem proving has increasingly relied on human-designed workflows around these models, such as planning, decomposition, and orchestrating specialized provers.
By Skander Moalla, Ahmad Rammal, Niket Patel · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.