Simple AI Loop Beats Engineered Proof Systems
The research-stage h1 system matched or beat hand-built Lean workflows on reported benchmarks at lower cost per proof.
Updated on Nov. 1, 2026 in Mathematics
Meta and New York University researchers released h1, a minimal agentic loop for automated theorem proving, in a paper published October 25, 2026. The work is a research-stage system described in a paper, not a shipped product.
Why it matters
h1 connects frontier large language models to the Lean 4 proof assistant and feeds verifier results back in real time, a design that skips predefined multi-step workflows. The approach builds on Meta's earlier HTPS neural theorem prover introduced in 2022.
The paper reports h1 scoring 670 of 672 on PutnamBench and 82.5% on Erdős problem resolutions, against AlphaProof Nexus at 76%. Costs per successful proof were $121 for h1 and $317 for AlphaProof Nexus, and the figures come from the authors' own evaluations rather than third-party testing.
The players
Meta
Meta is a major technology company with large-scale AI research programs, and it introduced the HTPS neural theorem prover in 2022.
New York University
New York University is a research university whose researchers co-developed the h1 system with Meta.
The details
h1 pairs frontier language models with Lean 4, a formal proof language in which every step is machine-checked, so the compiler's verdict arrives immediately after each edit. The system keeps a persistent proof file that carries its state between steps, letting the model edit, check, and reflect in a loop rather than relying on a pipeline that splits a problem into subtasks. The design bets that a simple loop with strong feedback will outperform hand-engineered workflows as underlying models improve.
Timeline
2022: Meta introduced the HTPS neural theorem prover, an earlier step in the research line.
October 25, 2026: The h1 paper was released.
November 1, 2026: Coverage of the release was published on cryptobriefing.com.
The Tech Race
h1 is presented as a test of the bitter lesson, the argument that general methods with more computation tend to beat hand-built structure, applied to automated theorem proving. It positions the work against engineered Lean proof workflows and against AlphaProof Nexus on the Erdős problem benchmark.
h1 is a research system described in a paper, with no announced availability, pricing, or release date for general use. Its near-term relevance is to researchers working on formal verification and automated proof search rather than to everyday users.
The takeaway
The central question is whether minimal feedback loops keep beating engineered workflows as models improve. Watch for independent replication of the PutnamBench and Erdős results.
Further reading
For broader context on formal proof and the mathematics behind it, explore Mathematics coverage.
