Meta, NYU Release h1 Automated Theorem Prover
The open research system connects large language models to the Lean 4 proof assistant and feeds them real-time compiler feedback.
Updated on Nov. 1, 2026 in Mathematics
Meta and New York University researchers have released h1, a minimal agentic loop for automated theorem proving that connects frontier large language models to Lean 4. The h1 paper, titled "h1: A Taste of the Bitter Lesson in Automated Theorem Proving," was released on October 25, 2026.
Why it matters
h1 matched or beat engineered Lean proof workflows at lower cost per successful proof, suggesting simpler, general-purpose setups may outperform hand-built pipelines. The research builds on Meta's earlier neural theorem prover work, including the HTPS system introduced in 2022.
The h1 system resolved eight open problems in the FormalConjectures collection, according to benchmark results reported in the paper. Results are self-reported by the research team, and independent replication of the figures has not been described.
The players
Meta
Meta is the technology company whose AI research group developed h1 and earlier neural theorem proving work.
New York University
New York University is the research university that co-developed the h1 system with Meta researchers.
The details
h1 connects frontier large language models to Lean 4 and feeds them verifier results in real time, using persistent file editing, immediate compiler feedback after every edit, and model reflection. Rather than relying on predefined workflows that split problems into pieces and route subtasks, h1 works on a proof file that keeps its state between steps.
Timeline
2022: Meta introduced the HTPS neural theorem prover
October 25, 2026: The h1 paper was released
The Big Picture
h1 extends Meta's HTPS neural theorem prover line from 2022 toward a simpler agentic design. Its results suggest that general-purpose language model loops may challenge the engineered Lean workflows that have dominated automated proof research.
Automated proof tools that cut cost per verified result could make formal verification of software and mathematical claims cheaper over time. The h1 results are research benchmarks, so practical deployment timelines remain unclear.
The takeaway
Lower-cost automated proof systems may let researchers verify more claims without building specialized pipelines for each problem. Readers should watch for independent replications of h1's benchmark figures.
Further reading
Explore more Mathematics coverage to see how automated proof systems are changing formal verification.
