NVIDIA says AI outscored top human on IOI 2026 problem set in unofficial run

NVIDIA researchers say a coding-focused AI system scored higher than the best human on the 2026 International Olympiad in Informatics problem set, but the result comes with an important caveat: the run was unofficial, not supervised by IOI organizers and not included in the contest’s official standings.

The claim appears in a paper posted to arXiv on Sept. 2, 2026, titled “Post-Training Language Models for Gold-Medal Performance in Coding Competitions,” listed as arXiv:2609.02849 and marked © 2026 NVIDIA. The authors are Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar and Boris Ginsburg. In the paper, the researchers say they built a “competition-specific Ultra-CC system” and evaluated it prospectively during IOI 2026, meaning they say it was run live on that year’s contest problems. According to the paper, the system scored 535.4 out of 600, above the IOI 2026 gold-medal cutoff of 361.12 and above the top official human score of 498.27.

The human benchmark is independently supported by official IOI statistics and outside reporting. Xinhua reported that Chinese contestant Xu Qiwen topped the leaderboard with 498.27 points as the overall champion. But the AI result is author-reported, not an official competition result. The paper states: “Our system was not an official IOI contestant and the run was not supervised by IOI. Therefore, its score was not included in the official rankings and the evaluation is reported as an unofficial, unsupervised benchmark.” No independent third-party audit publicly confirming the 535.4 score was identified in the research report.

The paper says the result came from a specialized post-training pipeline aimed at competitive programming, built on a curated 22,000-problem dataset and paired with an iterative test-time refinement method called GenCorrect.

The International Olympiad in Informatics is one of the best-known programming contests for secondary-school students and is widely treated as a demanding benchmark for algorithmic reasoning. Public IOI 2026 rules and statistics show the contest consisted of six tasks worth a maximum of 600 points, spread across two contest days of five hours each, with a 50-submission limit per task. The NVIDIA paper says its system was evaluated under the same time, internet-access and submission constraints faced by contestants.

That makes the claim notable beyond a simple gold-medal threshold. The paper does not say the AI “won” IOI 2026, and it did not compete as an official entrant. Instead, the authors present the score as an unofficial benchmark on the same problem set used in one of the world’s highest-profile youth coding contests.

The researchers frame that as a milestone, writing: “To our knowledge, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set.” That novelty claim is also contained within the paper rather than in any IOI certification.

The authors say they plan to release the competition checkpoint and runnable inference and evaluation recipes through NeMo-Skills, which is NVIDIA’s toolkit for such workflows. They also say they cannot release the full training corpus because of third-party redistribution restrictions.