Maxproof (arxiv.org) AI

The arXiv paper “MaxProof” proposes a population-level test-time scaling approach for generating mathematical proofs, using an M3 model trained on proof generation, verification, and critique-conditioned repair. At test time, it treats the model as generator/verifier/refiner/ranker, searches over many candidate proofs, and selects the final result via tournament selection—reporting 35/42 on IMO 2025 and 36/42 on USAMO 2026.