MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling
Summary
MaxProof introduces a population-level test-time scaling framework for mathematical proofs using a defense-in-depth generative verifier. The approach combines generation, verification, and refinement to search a pool of candidate proofs and selects a final proof, achieving top scores on IMO 2025 and USAMO 2026. The paper demonstrates an advancement in automated theorem proving and AI-assisted mathematical reasoning.