One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
huggingface ·
NVIDIA researchers have demonstrated that the Nemotron model family can achieve gold-medal level performance in both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). By applying supervised fine-tuning, reinforcement learning, and a generate-verify-refine inference loop, they specialized general foundation models for these distinct competitive domains. For IOI 2026, the Nemotron-3-Ultra-CC system scored 535.4 out of 600, surpassing the gold threshold and the top human score. In IMO 2026, a combination of SFT and RL checkpoints enabled the model to score 30 out of 42 points, exceeding the official gold standard of 29. The results highlight that success stems from co-designing specialized training data with transparent inference workflows rather than relying on brute-force sampling alone. NVIDIA has released the associated models, datasets, and recipes on Hugging Face to encourage further community development. These open resources include the Nemotron-3-Ultra-CC coding specialist and the complete collection for the IMO 2026 mathematical proofs.
huggingface