New 30B-A3B Reasoning Model Achieves Gold-Medal Performance in Olympiads

A recent paper from Hugging Face introduces a 30B-A3B reasoning model that demonstrates gold-medal-level performance in both mathematical and physics competitions, including the International Mathematical Olympiad (IMO) and International Physics Olympiad (IPhO). The model, named SU-01, employs a systematic approach that includes a reverse-perplexity curriculum, two-stage reinforcement learning, and test-time scaling to enhance its proof-search capabilities. Trained on 340,000 sub-8K-token trajectories and refined through 200 reinforcement learning steps, SU-01 can tackle complex problems with trajectories exceeding 100,000 tokens. The authors emphasize the model's strong generalization abilities beyond traditional mathematics and physics domains, and they have open-sourced the code and model for public access.