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

This paper presents an end-to-end specialization pipeline for competitive programming, combining curation of 22,000 executable problems, synthetic reasoning traces from DeepSeek-V4-Flash, supervised fine-tuning, executable-reward reinforcement learning with GRPO, and GenCorrect, a feedback-driven test-time strategy that generates diverse solutions, evaluates selected submissions, and iteratively refines them. NVIDIA’s Nemotron-3-Nano-CC uses 1.2M training traces plus reinforcement learning, while the larger Nemotron-3-Ultra-CC relies on supervised fine-tuning alone. On IOI 2025, Nano-CC improved from 130 points at Score@1 to 291 after post-training and 468 after five GenCorrect rounds, surpassing the 438.3 gold threshold despite having only 3B active parameters. Ultra-CC reached 502 after the same iterative inference. For a prospective IOI 2026 run, the system combined GLM-5.2-generated training data, expanded final-round sampling, execution-based candidate selection, and NVFP4 quantization; under the same time, internet, and submission constraints as contestants, it scored 535.4 out of 600, exceeding the 361.12 gold threshold and the 498.27 top-human score. The result is a system-level demonstration that post-training and test-time compute can transform general-purpose language models into gold-medal coding systems, although it required substantial computational resources and was not an official contestant.