On September 2, 2026, a paper released by NVIDIA's Nemotron team on arXiv showed that the Nemotron-3-Ultra-CC system scored 535.4 points at the onsite competition of the 2026 International Olympiad in Informatics (IOI), surpassing the human high score of 498.27 points.
The Facts
Paper number arXiv:2609.02849 records the performance of Nemotron-3-Ultra-CC (550B total parameters / 55B active parameters) under the same time limits, internet access, and submission rules as human contestants. The system scored 535.4 points, higher than the gold medal threshold of 361.12 points and higher than the human high score of 498.27 points at that edition. The paper states that this is the first time an AI system has achieved this result on the complete IOI problem set.
The same paper also disclosed test data for Nemotron-3-Nano-CC (30B total parameters / 3B active parameters) at IOI 2025: the base model scored 130 points, 291 points after supervised fine-tuning, and 468 points with GenCorrect added, exceeding that edition's gold medal threshold of 438.3 points.
Mechanism Breakdown
The end-to-end pipeline described in the paper includes large-scale problem screening, synthetic reasoning traces, supervised fine-tuning, and reinforcement learning. Nemotron-3-Ultra-CC used only supervised fine-tuning, while Nano-CC additionally underwent reinforcement learning. GenCorrect, as a test-time strategy, improves output quality by iteratively generating, evaluating, and refining diverse solutions.
These steps were all carried out on the basis of 22,000 curated problems provided in the paper. The paper does not provide a breakdown of specific performance gains at each stage, but it explicitly identifies GenCorrect as a key component in the score improvement from IOI 2025 to 2026.
Industry Impact
IOI, as the internationally recognized highest-level competition for coding ability, runs its onsite contest under rules identical to the real competition environment. Previously, AI systems mostly achieved high scores on benchmarks such as ARC-AGI, but the immediacy and rule parity of IOI make this result much harder to dismiss as benchmark bias.
The paper also mentions that Nemotron-3-Nano-CC had already reached gold medal level at IOI 2025, demonstrating the transferability of the post-training pipeline across models of different scales. This provides other laboratories with a complete reference path from data curation to test-time computation.
Strategic Assessment
(The following is analysis, not fact.) The breakthrough at the IOI onsite competition may accelerate the industry's focus on "coding ability under verifiable rules," in contrast to complex engineering tasks that require sustained adherence to requirements over long periods. Companies may reassess the validity of existing benchmarks and increase investment in end-to-end post-training pipelines. In the competitive landscape, organizations with access to large-scale synthetic data and reinforcement learning resources will gain a relative advantage, but real-world commercial deployment will still depend on stability observed over multiple rounds of iteration.
The paper does not disclose training costs or energy consumption data. If multiple institutions pursue similar experiments in the future, the resource barrier may be pushed even higher.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接