In a new research paper, NVIDIA presents an end-to-end specialization pipeline for training competitive programming models that achieve gold-medal performance in international coding competitions. The work focuses on the International Olympiad in Informatics (IOI) and the International Collegiate Programming Contest (ICPC), widely regarded as the most challenging settings for AI reasoning.
The pipeline combines large-scale problem curation, synthetic reasoning traces, supervised fine-tuning (SFT), and reinforcement learning (RL). Using 22,000 curated problems, the team trained two models: Nemotron-3-Nano-CC (30B-A3B) with SFT and RL, and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone. They also introduced GenCorrect, a feedback-driven test-time compute strategy that iteratively generates, evaluates, and refines diverse solutions.
On the IOI 2025 benchmark, Nano-CC improved from 130 points to 291 after post-training, and to 468 with GenCorrect—exceeding the gold threshold of 438.3. Ultra-CC reached 502 points. Guided by these results, the team developed a competition-specific Ultra-CC system and evaluated it prospectively during IOI 2026, under the same time, internet-access, and submission constraints as human contestants.
The system scored 535.4 out of 600, surpassing both the gold threshold of 361.12 and the top human score of 498.27. According to the authors, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set.