One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
Two high-performing models, Nemotron 3 Nano-CC and Nemotron 3 Ultra-CC, achieved gold-level results in the International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO) competitions. Nemotron, a foundation model, was fine-tuned using supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference.
The IOI result was achieved through a live run with the same constraints as human contestants, while the IMO system's proofs were graded by official graders. Four key components were used to fine-tune Nemotron for these demanding tasks: extensive training and inference runs, curating 22,000 competitive programming problems and generating synthetic reasoning traces for Nano-CC, and using SFT for Ultra-CC.
The specialization of Nemotron for each competition allowed it to outperform general models. For IMO, a complementary approach of using both SFT and RL checkpoints, along with a natural language system for generating, verifying, and refining proofs, resulted in a final score of 30 out of 42 points, surpassing the gold-medal threshold.
The findings demonstrate the importance of fine-tuning a strong foundation model for specific domains and the benefits of combining different training techniques, data curation, and inference strategies.
Rewritten by urgent.news; may contain errors — check the original.