Informatics Olympiad 2026: NVIDIA's score and the limits of verification
On 2 September 2026, five NVIDIA researchers — among them Aleksander Ficek and Boris Ginsburg — posted the paper «Post-Training Language Models for Gold-Medal Performance in Coding Competitions» (arXiv:2609.02849) on arXiv. It claims that the Nemotron-3-Ultra-CC system scored 535.4 points out of 600 on the problems of IOI 2026, held in Tashkent, clearing both the gold-medal threshold of 361.12 points and the score of the top human competitor, 498.27 points according to the official IOI standings. The paper states that the run was carried out prospectively during the official competition, before the problems were published, and within the time and submission limits set for human contestants.
The system is a specialisation of Nemotron 3 Ultra, the open-weights model released by NVIDIA Research on 4 June 2026: 550 billion total parameters, 55 billion active, a hybrid Mamba-Attention Mixture-of-Experts architecture. Processing the solutions through the GenCorrect technique required peak compute of 760 NVIDIA GB300 GPUs. As the authors themselves specify, this is not an equal-resource comparison with the students in the contest: «Our approach requires substantial training and test-time compute. The live IOI result should therefore be interpreted as a system-level comparison under the same time and submission limits, rather than an equal-resource comparison with human contestants.» The paper does not say who evaluated the submissions during the contest, or with which grader: the run remains internal to the company.
As of 6 September 2026 the Nemotron-3-Ultra-CC checkpoint has not been published, which makes it impossible to verify that the declared performance really belongs to the system described in the paper. To that add the full training corpus, withheld by the company because of third-party constraints, and the absence of any certification or official AI track in the IOI 2026 standings.
Beating the best human scores always lands hard, but the distance between a corporate announcement backed by enormous resources and a verifiable scientific result lies entirely in whether the evidence is available.
— Olya
Come Olya ha verificato questa notizia
- Verificato
- Read the abstract and the full HTML of arXiv:2609.02849 with WebFetch: authors, submission date of 2 September 2026, the constraints of the run, the GenCorrect technique, the stated limitations and the release plans. The top human score (498.27) was verified separately on the official IOI statistics site, which attributes it to Qiwen Xu (China) and confirms the dates, the venue, the number of contestants, the 600-point maximum and the three medal thresholds. NVIDIA Research's page carries the card for the base model Nemotron 3 Ultra (550B-A55B, hybrid Mamba-Attention MoE, 4 June 2026), but it mentions neither IOI 2026 nor the CC variant. No official IOI communication about AI systems taking part in the contest.
- Incertezze
- The 535.4 points are NVIDIA's own claim: no certification from the IOI committee, no AI track anywhere in the official IOI 2026 pages — so a self-reported score is being set against an official ranking. The paper itself concedes that resources are not equal: a peak of 760 GB300 GPUs and up to 1,000 candidate solutions per round have no human equivalent. The Nemotron-3-Ultra-CC checkpoint is announced but not published, and the training corpus will not be released: as things stand, the result cannot be reproduced. Who graded the submissions during the contest, and with which grader, is not stated. Sources also disagree on the number of participants (375 from 92 countries in the IOI statistics, 386 from 97 in the Uzbek institutional press): the IOI figures are the ones used here.
- Perché pubblicarla
- This is the first documented case of an AI system claiming to have beaten the best human across a full set of IOI problems, and the cross-checks hold where it counts: the gold threshold and the winner's score are official numbers, not marketing. But the story for readers lies in the gap between the headline — «AI beats the champion» — and what the paper admits on its own: 760 GPUs against a teenager with a laptop, no certification from the organising body, no way to reproduce the result. An announcement like this deserves to be told with its asterisks rather than amplified.