Post-Training Language Models for Gold-Medal Performance in Coding Competitions
An arXiv paper describes an end-to-end specialization pipeline for competitive programming using 22,000 curated problems, synthetic reasoning traces, supervised fine-tuning (SFT), and reinforcement learning (RL). It trains Nemotron-3-Nano-CC (30B-A3B) with SFT and RL, and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone. It introduces GenCorrect, a feedback-driven test-time compute strategy. On IOI 2025, Nano-CC improves from 130 to 291 points after post-training and to 468 with GenCorrect, exceeding the gold threshold of 438.3, while Ultra-CC reaches 502. A competition-specific Ultra-CC system evaluated prospectively during IOI 2026 scores 535.4 out of 600, exceeding the gold threshold of 361.12 and the top human score of 498.27.
A research paper presents a post-training pipeline for large language models aimed at competitive programming. Using 22,000 curated problems, the authors train Nemotron-3-Nano-CC (30B-A3B) with SFT and RL, and Nemotron-3-Ultra-CC (550B-A55B) with SFT alone. They introduce GenCorrect, a test-time strategy that iteratively generates, evaluates, and refines solutions. On IOI 2025, Nano-CC improves from 130 to 291 points after post-training and to 468 with GenCorrect, exceeding the gold threshold of 438.3, while Ultra-CC reaches 502. A competition-specific Ultra-CC system evaluated prospectively during IOI 2026 scores 535.4 out of 600, exceeding the gold threshold of 361.12 and the top human score of 498.27, reportedly the first AI system to outscore the highest-scoring human contestant.
The pipeline combines large-scale problem curation, synthetic reasoning traces, SFT, and RL. GenCorrect is a feedback-driven test-time compute strategy that iteratively generates, evaluates, and refines diverse solutions. The paper reports that post-training alone raises Nano-CC from 130 to 291 points on IOI 2025, and adding GenCorrect raises it to 468, indicating significant gains from test-time refinement. Ultra-CC, trained with SFT only, reaches 502 on IOI 2025, suggesting that larger models may benefit from SFT without RL in this domain. The IOI 2026 system scores 535.4 out of 600 under human contestant constraints.
The reported results suggest that specialized post-training and test-time compute strategies can enable language models to surpass top human performance in competitive programming. The use of a 30B-A3B model with RL and a 550B-A55B model with SFT alone indicates a trade-off between model size and training method. The introduction of GenCorrect highlights the growing importance of test-time compute for reasoning tasks. The paper's claim of being the first AI system to outscore the highest-scoring human at IOI may influence benchmarks and expectations for AI coding capabilities.
The demonstrated capability could be applied to automated code generation, debugging, and competitive programming assistance. The pipeline's use of synthetic reasoning traces and curated problems may reduce reliance on human-annotated data. The test-time compute strategy GenCorrect could be integrated into developer tools to improve solution quality. The reported performance may attract interest from companies developing AI coding assistants or seeking to benchmark their models against top human performance.
Observable next signals include whether the models or methods are released publicly, whether the results are replicated in other competitions like ICPC, and whether the GenCorrect approach is adopted in other reasoning domains. Further signals include the publication of peer-reviewed versions, the release of the curated problem set or training data, and any commercial deployment of the models for coding assistance.