Post-Training Language Models for Gold-Medal Performance in Coding Competitions

The researchers proposed an end-to-end specialized pipeline, combining large-scale problem screening, synthetic reasoning trajectories, supervised fine-tuning, and reinforcement learning training for the Nemotron-3-Nano-CC and Ultra-CC models. In the IOI 2025 competition, Nano-CC scored 291 points after post-training adjustments, and with the introduction of the GenCorrect strategy, its score increased to 468 points, exceeding the gold medal threshold; Ultra-CC scored 502 points. Based on these results, the specialized Ultra-CC system scored 535.4 points in the IOI 2026 pre-competition evaluation, under the same time, network connectivity, and submission restrictions, exceeding both the gold medal threshold and the highest score of human competitors. It is claimed that this is the first AI system to exceed the highest score of human competitors on the IOI problem set.