Meta and New York University designed h1, a minimal agentic loop connecting large language models to Lean 4 with immediate verifier feedback. The system scored 670 out of 672 on the PutnamBench benchmark with GPT-5.5, 90 out of 100 on the FATE-X assessment, and resolved eight open problems in the FormalConjectures collection. On Erdős problems, h1 achieved an 82.5% success rate, compared with 76% for DeepMind’s AlphaProof Nexus. Its cost of $121 per successful proof, versus $317 for AlphaProof Nexus, enables teams to attempt more problems, although its performance depends heavily on frontier models.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
