MaxProof framework: MiniMax M3 surpasses human gold medal line on IMO 2025 and USAMO 2026
Original title:MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Evolutionary Search
AI summary
MiniMax M3 uses the MaxProof framework to surpass the human gold medal line on two math olympiad benchmarks, IMO 2025 and USAMO 2026. The framework trains in three stages: Proof RL uses a generative verifier to provide rewards and performs long-horizon reinforcement learning to improve proof generation; Verifier Alignment aligns verification as an error localization task; Refinement Augmentation uses erroneous proofs and verification analysis pairs generated during training to fine-tune repair capabilities via rejection sampling. The three are merged into the general M3 model. The system filters noise through a low false-positive rate to ensure RL stability.
Key Takeaways
- MiniMax M3 uses the MaxProof framework to surpass the human gold medal line on two math olympiad…
- The framework trains in three stages: Proof RL uses a generative verifier to provide rewards and…
- Verifier Alignment aligns verification as an error localization task
- Refinement Augmentation uses erroneous proofs and verification analysis pairs generated during…
- The three are merged into the general M3 model.
Quick Facts
- Published
- Updated
- Source
- MiniMax:Blog(网页)
- Category
- Papers & Research
- AI score
- 76
M3在数学奥赛上搞定人类金牌,靠的是用生成验证器做RL和进化搜索,这套组合对复杂推理任务的普适性可能比提高一个benchmark分数更有价值。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: MaxProof framework: MiniMax M3 surpasses human gold medal line on IMO 2025 and USAMO 2026. June 9, 2026. https://new.nat6.net/en/news/cmqadpjwv0kjzslldk3axe95f/