Problem
Research questions and friction points this paper is trying to address.
Olympiad reasoning
mathematical problem solving
scientific reasoning
long-horizon reasoning
AI reasoning
Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
🤖 AI Summary
This work proposes a general-purpose reasoning system designed to achieve gold-medal performance in international olympiads such as the International Mathematical Olympiad (IMO) and the International Physics Olympiad (IPhO). By integrating a unified training framework that combines reverse perplexity curricula, two-stage reinforcement learning, and test-time scaling, the system enables stable long-horizon reasoning and cross-disciplinary generalization. The approach synergistically leverages supervised fine-tuning, verifiable reward-driven reinforcement learning, and proof-level optimization to substantially enhance the model’s capacity for rigorous proof search and self-verification. The resulting model, SU-01, demonstrates gold-medal-level performance on recent competitions including IMO 2025, USAMO 2026, and IPhO 2024/2025, while supporting stable inference over sequences exceeding 100,000 tokens.
📝 Abstract
Recent progress in reasoning models has substantially advanced long-horizon mathematical and scientific problem solving, with several systems now reaching gold-medal-level performance on International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO) problems. In this paper, we introduce a simple and unified recipe for converting a post-trained reasoning backbone into a rigorous olympiad-level solver. The recipe first uses a reverse-perplexity curriculum for SFT to instill rigorous proof-search and self-checking behaviors, then scales these behaviors through a two-stage RL pipeline that progresses from RL with verifiable rewards to more delicate proof-level RL, and finally boosts solving performance with test-time scaling. Applying this recipe, we train a 30B-A3B backbone with SFT on around 340K sub-8K-token trajectories followed by 200 RL steps. The resulting model, SU-01, supports stable reasoning on difficult problems with trajectories exceeding 100K tokens, while achieving gold-medal-level performance on mathematical and physical olympiad competitions, including IMO 2025/USAMO 2026 and IPhO 2024/2025. It also demonstrates strong generalization of scientific reasoning to domains beyond mathematics and physics.
Innovation
Methods, ideas, or system contributions that make the work stand out.
reverse-perplexity curriculum
two-stage reinforcement learning
proof-level RL
test-time scaling
olympiad-level reasoning
🔎 Similar Papers
No similar papers found.
