NVIDIA researchers have published results claiming that a competitive-programming specialization of Nemotron 3 Ultra outscored every human contestant at the 37th International Olympiad in Informatics (IOI 2026), a first in the competition's 37-year history. The system, Nemotron-3-Ultra-CC, posted 535.4 out of 600 points at the contest, held August 9-16 in Tashkent, Uzbekistan, beating the top human score of 498.27 by more than 37 points and clearing the gold-medal threshold of 361.12 by a wide margin.
GenCorrect: a verification-grounded feedback loop
The result, described in an arXiv preprint published September 2 by NVIDIA researchers Aleksander Ficek, Sean Narenthiran, Mehrzad Samadi, Somshubra Majumdar and Boris Ginsburg, hinges on a test-time compute strategy the team calls GenCorrect. Instead of submitting the first solution it generates, the system produces multiple candidates, evaluates them against the competition's automated judge, and iteratively refines them based on the feedback until it lands on a passing answer or runs out of time.
The design exploits a structural property unique to competitive programming: a cheap, fast, objective verifier. Every IOI submission is graded automatically against fixed test cases, so each GenCorrect loop iteration is grounded in real evaluation rather than self-assessed quality. That same architecture cannot directly replicate in domains — medicine, law, original research — where evaluation is slow, contested or requires human expertise.
Two models, one pipeline
NVIDIA trained two models on 22,000 curated competitive-programming problems with synthetic reasoning traces. Nemotron-3-Nano-CC carries 30 billion total parameters with 3 billion active per forward pass (30B-A3B) and used both supervised fine-tuning (SFT) and reinforcement learning. Nemotron-3-Ultra-CC carries 550 billion total parameters with 55 billion active per forward pass (550B-A55B) and reached its performance ceiling with SFT alone — a data point relevant to ongoing debates over whether RL provides a meaningful advantage at frontier scale.
Rapid progression, and an important caveat
The trajectory is striking. At IOI 2024, OpenAI's o1-ioi finished below the gold threshold under standard rules; a retrospective o3 evaluation on that same problem set crossed gold without hand-tuning. At IOI 2025, NVIDIA's Nano-CC (468) and Ultra-CC (502) both cleared the gold threshold of 438.3 but did not surpass the top human. At IOI 2026, Ultra-CC's 535.4 sits 37 points above the best human in the building.
NVIDIA notes the scores are self-reported and have not been officially adjudicated by the IOI organizing committee, and no per-problem breakdown or compute budget is disclosed. The 37-point margin is substantial and IOI's automated grading is harder to game than benchmarks that rely on human judgment, but the result should be read as a strong research claim pending independent verification, not a certified competition record.
The Nemotron 3 family broadly — Nano, Super and Ultra — has been steadily rolling out alongside NVIDIA's sovereign AI stack for government and enterprise customers, and its agentic tooling powered by Synopsys' AgentEngineer for EDA was demonstrated earlier this summer at DAC 2026. The IOI 2026 milestone is an inflection for AI-in-programming benchmarks and, more importantly, a public demonstration of how far verification-grounded test-time compute can push large language models in verifiable domains.
Reporting based on coverage from NVIDIA Research, Tech Times, AlphaSignal and Artificial Analysis.