One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

NVIDIA's Nemotron model family was adapted to achieve gold-medal level results in both competitive programming (IOI) and mathematical proofs (IMO). Starting with the Nemotron 3 foundation model, researchers used a recipe involving curated domain-specific data, supervised fine-tuning (SFT), and reinforcement learning (RL). The specialized models were then paired with an inference loop that generates, evaluates, and refines answers, such as the GenCorrect strategy for coding problems. This approach demonstrates that a single, strong foundation model can be effectively specialized for diverse and complex tasks without building new models from scratch. The results highlight the combined power of fine-tuning and sophisticated test-time compute strategies.