The Big Picture
Post-training with curated problem data plus a few rounds of feedback-driven refinement lets relatively compact coding models reach and surpass gold-medal performance in real programming contests.
ON THIS PAGE
The Evidence
Careful post-training—using 22,000 curated problems, large-scale synthetic reasoning traces, supervised fine-tuning, and targeted reinforcement learning—dramatically improves contest-style coding. An iterative test-time strategy that generates many diverse solutions and refines them using execution feedback execution feedback provides the largest extra boost under strict submission limits. A small, specialized model with 3 billion active parameters and a larger 55 billion-active-parameter model both crossed gold thresholds; the competition-tuned system scored 535.4/600 in a live run, beating the highest human score.
Data Highlights
1Supervised fine-tuning raised the Nemotron-3-Nano-CC single-sample score from 130 to 280 (a 150-point gain).
2Five rounds of GenCorrect raised the Nano model to 468 points, exceeding the IOI gold threshold of 438.3 under the 50-submission limit.
3A competition-tuned Ultra system scored 535.4/600 in a live IOI 2026 run, surpassing the top human score of 498.27.
What This Means
Engineers building code-generation assistants and leaders choosing training and inference investments should care because the results show post-training and test-time refinement deliver outsized returns compared with raw model scale alone. Researchers studying execution-grounded learning and product teams evaluating resource trade-offs can use these findings to prioritize curated datasets, synthetic reasoning traces, and closed-loop inference strategies. For practitioners focused on coding workflows, this aligns with the needs of Coding Assistants.
Not sure where to start?Get personalized recommendations
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
The approach requires substantial training and test-time compute, so results reflect a system-level, not equal-resource, comparison with human contestants. Reinforcement learning was limited by compute (the largest model did not receive code-specific RL), so gains from RL may be smaller or variable across scales. Results are specific to competitive programming tasks and may not generalize to other domains without similar curated evaluation and execution feedback. Researchers should weigh these aspects when planning deployments, including potential areas for improvement highlighted by execution feedback.
Methodology & More
A practical pipeline focused on competitive programming produced gold-medal performance by combining large-scale problem curation, synthetic reasoning traces, supervised fine-tuning, execution-based reinforcement learning, and an iterative test-time refinement loop called GenCorrect. The dataset includes 22,000 curated problems packaged into runnable evaluation environments runnable evaluation environments, and synthetic traces were generated at scale to teach stepwise reasoning. Two post-trained systems were built: a compact 30 billion-parameter model with 3 billion active parameters and a larger 550 billion-parameter model with 55 billion active parameters. Fine-tuning produced the largest single-sample gains, while reinforcement learning added smaller incremental improvements where applied.
GenCorrect is an inference-time strategy that repeatedly generates diverse candidate solutions, runs them, collects execution feedback, and uses that feedback to refine subsequent generations within a constrained submission budget (50 submissions per problem). In practice, a few rounds of GenCorrect converted the gains from fine-tuning into contest-winning performance: the compact model rose from 130 to 468 points after post-training and five refinement rounds, and the larger model reached over 500 points. The competition-tuned Ultra system achieved 535.4/600 in a live IOI 2026 run under the same time and submission limits as human contestants, exceeding the highest human score. The team plans to release checkpoints and inference recipes inference recipes, though full training data cannot be redistributed because of third-party restrictions.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
At least one author (Somshubra Majumdar) has h-index 24 (established researcher range 20–40). Other authors have moderate h-indices. Venue is arXiv and affiliations not specified, so not top-lab level but overall credible.