Illustration, AI-generated from the story text. Tap a panel to enlarge.

An off-the-shelf model proves 669 of 672 Putnam problems in Lean, and half of real code

A paper posted to arXiv this week describes EasyProver, which uses an unnamed frontier model with no extra training, run as recursive agents, to write machine-checked proofs in Lean. It reports 669 of 672 PutnamBench problems, all six IMO 2026 problems, and 189 of 189 on the VERINA code-verification set. On a new set of 2,000 statements about real Rust programs it proved 1,054, or 52.7%, within 30 minutes each.

The competition-math benchmark is effectively finished by scaffolding alone. Real software is where half the work still fails, which is the more useful number. The paper lists nine authors and no affiliations, and does not say whether the code is released. SourcesA

Sources

• AScaling Lean Theorem Proving for Real-World Code Verification (arXiv)arxiv.org

This story ran in the Morning Brief of October 30th 2026.