MaxProof introduces a novel approach to mathematical proof generation that treats large language models as multiple agents working in parallel. The framework, built on the MiniMax-M3 series, combines proof generation, verification, and repair capabilities into a single model that can scale its reasoning at test time.

The system uses tournament selection across a population of candidate proofs, allowing it to search through multiple solution paths simultaneously. This defense-in-depth generative verifier achieves a low false-positive rate while systematically exploring mathematical reasoning chains.

On competition benchmarks, MaxProof scored 35/42 on IMO 2025 and 36/42 on USAMO 2026, surpassing the human gold-medal threshold on both tests. These results demonstrate that population-level scaling can match expert mathematician performance on rigorous proof problems.

The achievement suggests test-time scaling techniques could become essential for tackling complex reasoning tasks that require multiple verification passes and iterative refinement.