Nvidia’s Nemotron Outscores IOI 2026’s Top Human

An NVIDIA research team says it has built the first AI system to outscore the highest-scoring human contestant on an International Olympiad in Informatics problem set. In a preprint posted to arXiv on September 2, the team reports that Nemotron-3-Ultra-CC scored 535.4 out of 600 at IOI 2026 — clearing the event’s gold-medal threshold of 361.12 by a wide margin and finishing 37 points ahead of the top official contestant’s 498.27.

The framing matters more than the number. Competitive-programming milestones have arrived at a steady clip for two years, but they have almost all been retrospective: a model is turned loose on a past contest whose problems, editorials and reference solutions have been sitting on the open web for months. This run was live. According to the paper, the system worked the IOI 2026 problems during the competition itself, on the same submission platform, inside the same five-hour sessions, capped at 50 submissions per problem and one submission per minute, with no internet access and local code execution permitted — the identical envelope the students were working in.

What it was not is an official result. The system was not an entered contestant, the run was not supervised by IOI, and the score does not appear in the event’s rankings. NVIDIA reports it as an unofficial benchmark, and the distinction is worth holding onto: the constraints were replicated, not adjudicated.

Two models came out of the pipeline. Nemotron-3-Ultra-CC is the 550-billion-parameter version — 55 billion active, built on the mixture-of-experts hybrid Mamba-attention architecture NVIDIA shipped in Nemotron 3 — and Nemotron-3-Nano-CC is a 30B/3B sibling used mostly to isolate what each training stage contributes. Both were supervised fine-tuned on synthetic reasoning traces generated by DeepSeek-V4-Flash: 1.2 million traces for Nano, 477,000 for Ultra, drawn from 22,000 curated problems spanning 16 competition families, with every evaluation benchmark held out to keep contamination off the table. Reinforcement learning with executable rewards, using Group Relative Policy Optimization, was applied only to Nano, where it bought about 1.1 percentage points on IOI 2025.

The heavier lifting happens at inference. The paper’s named contribution is GenCorrect, a feedback-driven test-time strategy that treats the 50-submission budget as a search loop rather than a formality. Each round generates up to 200 candidate solutions, clusters them by token-shingling similarity without looking at their scores, submits ten representatives, and folds the judge’s feedback into the next round. Five rounds is exactly 50 submissions — the IOI limit, spent deliberately. The final round widens the candidate pool to 1,000 and selects on execution results. Ultra-CC’s advantage over Nano widened with every iteration, which is the real finding buried in the methodology: the bigger model is not just better per attempt, it converts additional parallel samples into score more efficiently.

The supporting benchmarks track the same story. On IOI 2025, Nano-CC went from 130 points at baseline to 468, past that year’s 438.3 gold threshold, while Ultra-CC reached 502. On ICPC 2025 problems Ultra-CC solved 9.6 on average for a 57.4% pass@1, and on LiveCodeBench Pro it reached 74.5% pass@1. NVIDIA has released the Nemotron 3 Ultra checkpoints openly, including quantized builds and training data, so the base the CC variants were specialized from is inspectable rather than asserted.

The authors are candid about the asterisk. “Our approach requires substantial training and test-time compute,” they write, and 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.” A teenager gets five hours, a laptop and a pencil. Ultra-CC gets five hours and a datacenter, generating up to a thousand programs per problem and paying a judge’s feedback loop to sort them. Both finish inside the same clock; only one of them is cheap.

Read narrowly, that caveat deflates the headline. Read carefully, it relocates the achievement. The interesting claim in this paper is not that a language model can write a gold-medal solution — several can — but that a disciplined search over a fixed submission budget, with execution feedback as the only supervisor, is enough to close a gap that raw sampling had not. That is a recipe, and it generalizes to any domain where correctness can be checked automatically and attempts are rationed.