A model aimed at improving the mathematical reasoning ability of large language models, strengthening the verifiability of the reasoning process beyond simple final-answer accuracy
Introduces a self-verification mechanism to address the limitations of existing reinforcement learning approaches centered on final-answer rewards
Designed so that the generative model finds and fixes its own errors on problems that need step-by-step logical development, such as theorem proving
Uses a verifier as the reward model and scales verification compute to automatically label hard proof data, enabling continuous performance gains
Records top-level scores on IMO 2025, CMO 2024, Putnam 2024, and more, demonstrating the feasibility of self-verifying math AI