One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO (huggingface.co)

Recent advancements with the Nemotron AI model have led to gold-level performances in both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO) 2026. Leveraging Nemotron 3 as a strong foundation, researchers employed supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference to create specialized models that not only outperformed human competitors but demonstrated the model's adaptability for high-stakes challenges. Notably, the Nemotron-3-Ultra-CC model scored 535.4 out of 600 in the IOI, surpassing the gold threshold, while another variant achieved 30 out of 42 points in the IMO competition, highlighting the effectiveness of targeted model adaptation. The significance of these results lies in their demonstration that a capable foundation model can be efficiently specialized for specific domains without the need for entirely new architectures. The approach emphasizes a systematic training process involving domain-specific problem curation, iterative data refinement, and strategic pairing of SFT and RL techniques. This strategy not only yielded exceptional performance in competitions but also suggested that co-designing the model and its inference process could yield superior outcomes. The team has made their methodologies and trained models available via Hugging Face, encouraging the AI/ML community to build upon these findings for future advancements.