h1: A Taste of the Bitter Lesson in Automated Theorem Proving (arxiv.org)

A recent study highlights a notable advancement in automated theorem proving, addressing how formal verification can keep pace with rapid developments in AI, particularly with language models. Researchers introduced a new tool called h1, which provides an efficient interface to the Lean 4 theorem prover without relying on complex human-designed workflows. This approach yielded state-of-the-art results on several benchmarks, including PutnamBench and FormalConjectures100-Solved, demonstrating that a simple, agentic loop can match or even outperform specialized workflows such as Goedel-Architect and AlphaProof Nexus. The significance of this finding lies in the potential for streamlined theorem proving processes amidst growing demands for formal verification in AI research. The study concludes that while simplicity is important, the efficiency of the interaction with Lean plays a critical role in cost and problem-solving rates. This calls for future efforts in the AI/ML community to focus on optimizing execution feedback from frontier models rather than defaulting to complex workflows—highlighting that a refined approach can enhance automated theorem proving's effectiveness while avoiding past pitfalls.