Nvidia's Nemotron-3-Ultra-CC Model Beats Top Human Score at International Informatics Olympiad

The moment of truth for an artificial intelligence attempting to write code is not when it memorizes a textbook, but when it enters a test-time inference loop to iteratively generate, evaluate, and refine a solution to a problem it has never seen. That feedback-driven correction cycle is exactly what just pushed a language model past the world's best human programmers. At the 2026 International Olympiad in Informatics (IOI), an AI system operated under the same strict time limits and submission constraints as human contestants, and for the first time, it outscored every person in the room.[1][4]

The system, developed by researchers at Nvidia, is called Nemotron-3-Ultra-CC. During a prospective evaluation on the live IOI 2026 problem set, the model scored 535.4 out of a possible 600 points. That performance cleared the gold-medal threshold of 361.12 by a massive margin and surpassed the top human competitor's score of 498.27.[1][3][4]

"To our knowledge, this is the first AI system to outscore the highest-scoring human contestant on an IOI problem set," wrote Nvidia researchers Aleksander Ficek, Sean Narenthiran, and their colleagues in a September 3 preprint published on arXiv. The result marks a historic milestone in competitive programming, a discipline that requires sustained, multi-step algorithmic construction rather than simple pattern matching.[4]

The architecture behind the record-breaking score relies on a mixture-of-experts design. Nemotron-3-Ultra-CC contains 550 billion total parameters, but only activates 55 billion of them during any given inference step. Rather than relying on reinforcement learning, the Ultra model was trained exclusively using supervised fine-tuning on a highly curated dataset of 22,000 competitive programming problems.[1][4]

However, the raw parameter count is only half of the equation. The system's decisive advantage comes from a test-time compute strategy that Nvidia calls "GenCorrect." Instead of generating a single answer and moving on, GenCorrect allows the model to iteratively generate diverse candidate solutions, evaluate them against the problem's constraints, and refine the code based on the feedback from previous failed submissions.[4]

This iterative loop mimics the way a human programmer debugs, but executes at machine speed. By shifting computational resources from the initial training phase to the actual moment of inference, the model can catch and correct its own logical errors before finalizing a submission.[4]

Nvidia also demonstrated that this pipeline works at smaller scales. Alongside the massive Ultra model, the team trained Nemotron-3-Nano-CC, a 30-billion parameter model with just 3 billion active parameters. Unlike its larger sibling, the Nano version utilized both supervised fine-tuning and reinforcement learning.[1][4]

When tested on the previous year's IOI 2025 benchmark, the Nano-CC model initially scored 130 points. After applying the post-training pipeline, its score jumped to 291. Once the GenCorrect test-time strategy was engaged, the 30-billion parameter model reached 468 points, comfortably exceeding the 2025 gold-medal threshold of 438.3.[4]

The credibility of the 2026 result rests heavily on the evaluation conditions. Because the test was run prospectively during the actual IOI 2026 competition window, the problems were entirely novel. This eliminates the possibility that the model simply memorized the solutions from its training data, a common pitfall in AI benchmark testing.[1][4]

The system was also restricted to the exact same internet-access rules and submission limits imposed on the human students. A margin of 37 points above the top human contestant on a 600-point set under these identical constraints provides the first concrete evidence that language models can execute abstract reasoning at an elite level.[1][2][4]

The breakthrough suggests a fundamental shift in how AI systems will be deployed in software engineering. As models move from generating boilerplate enterprise code to autonomously architecting and verifying complex algorithms, the bottleneck in software development is poised to shift from writing code to defining the constraints the AI must solve.[2]