Meta and NYU present h1, an AI that lowers theorem-proving costs
BROBROK News · Based on reporting by DiarioBitcoin ·
Researchers at Meta and NYU introduced h1, a theorem-proving system that pairs language models with Lean 4 and immediate verifier feedback. In the study, it posted strong results on formal math benchmarks and a lower cost per successful proof on Erdős problems than AlphaProof Nexus.
Key points
- h1 links language models to Lean 4 without a complex agent orchestration stack.
- The system scored 670 out of 672 on PutnamBench and 90 out of 100 on FATE-X.
- The paper says it solved eight open problems from the FormalConjectures collection.
- On Erdős problems, h1 reported an 82.5% success rate and a cost of $121 per successful proof.
- The study frames simplicity of design as a key factor in automated theorem proving.
Why it matters
This matters because formal reasoning systems are becoming more capable and more efficient, which can affect how researchers build reliable AI tools. The findings are benchmark-specific, but they suggest that simpler architectures can still compete at a high level.
Original headline: “Meta y NYU presentan h1, una IA que demuestra teoremas con menor costo”
Read the original on DiarioBitcoin
Summary and analysis written by BROBROK News from the original report. Read the full story on the source website. Not investment advice.