Meta and NYU Unveil h1 System for Automated Theorem Proving
Researchers at Meta and New York University have developed h1, a streamlined AI system that achieves top benchmark scores in math theorem proving at a lower cost.
Meta and NYU Unveil h1 System for Automated Theorem Proving
Researchers at Meta and New York University have developed h1, a streamlined AI system that achieves top benchmark scores in math theorem proving at a lower cost.
Meta and NYU Advance Automated Theorem Proving
Researchers at Meta and New York University have introduced a streamlined artificial intelligence system designed for automated theorem proving. The framework, called h1, matches or exceeds the performance of elaborately engineered Lean proof workflows while operating at a reduced cost.
Detailed in a paper released on October 25, 2026, and titled "h1: A Taste of the Bitter Lesson in Automated Theorem Proving," the work originates from Meta's FAIR research team alongside NYU collaborators. The system relies on a minimal agentic loop that connects frontier large language models directly to Lean 4, providing verifier results in real time.
Unlike most automated provers that wrap language models in complex scaffolds and predefined routing workflows, h1 skips external machinery. The research indicates that the performance stems from maintaining an efficient interface with Lean rather than relying on heavy orchestration.
Benchmark Scores and Open Problems
According to the published findings, h1 achieved state-of-the-art performance across multiple formal theorem-proving benchmarks. Operating with GPT-5.5, the system scored 670 out of 672 on PutnamBench and reached 90 out of 100 on the FATE-X assessment.
Furthermore, the system successfully resolved eight open problems within the FormalConjectures collection, including variants of Erdős problems associated with mathematician Paul Erdős. When compared directly against DeepMind's AlphaProof Nexus on Erdős problem resolutions, h1 recorded an 82.5% success rate compared to 76% for AlphaProof Nexus.
The project also demonstrated significant cost efficiency. The h1 framework spent $121 per successful proof, whereas AlphaProof Nexus reported a cost of $317 per successful proof, highlighting a substantial gap in operating expenses.
Context and Methodology
The paper's title references the concept of the "bitter lesson," which suggests that scalable, general-purpose methods supported by heavy computation ultimately surpass specialized, handcrafted techniques over time. This approach marks a shift from Meta's earlier work, such as the HTPS neural theorem prover introduced in 2022, which relied on building smarter search mechanisms around the model.
The success of h1 relies heavily on frontier models like GPT-5.5 to achieve its high PutnamBench scores. By lowering the cost per successful proof from $317 to $121, the system alters the economic viability of testing speculative or complex mathematical conjectures.
Despite these results, the researchers note certain limitations. Benchmark scores reflect performance on specific test sets, and PutnamBench is currently near saturation at 670 out of 672 points, while the underlying capability remains tied to the performance of the frontier model deployed.
About The Author
