模型評測
NVIDIA Competitive Programming Model Scales Test-Time Compute Across 760 GPUs, Unofficially Surpassing the Highest Human IOI Score
Nemotron-3-Ultra-CC scored 535.4 out of 600 on the IOI 2026 problems, exceeding the top official contestant’s 498.27. The key was not only the 550B model, but GenCorrect, which repeatedly generated, tested, and revised large numbers of candidate programs; the run used up to 760 GB300 GPUs.
NVIDIA’s research team has published the training and inference pipeline for Nemotron-3-CC, a model specialized in competitive programming. The researchers first curated 22,000 problems from 16 types of competitions and built an environment capable of compiling and executing programs and scoring them against test cases. DeepSeek-V4-Flash was then used to generate reasoning traces. Nano-CC, a 30B model that activates 3B parameters per inference, underwent supervised fine-tuning and GRPO using 1.2 million examples. Ultra-CC, a 550B model that activates 55B parameters per inference, was supervised fine-tuned on 477,642 traces without programming-specific RL.
Single-pass generation was not the main breakthrough. The team’s GenCorrect method generates candidate solutions in parallel, compiles and tests them locally, then feeds errors, scores, and previous submission results into the next round. It ultimately selects a program while staying within the limit of 50 official submissions per problem. On IOI 2025, Nano-CC’s single-pass score rose from 130 for the base model to 280 after fine-tuning and 291 after RL. With five rounds of GenCorrect, it reached 468, showing that test-time search produced greater gains than RL in this experiment.
The prospective IOI 2026 run was conducted before the problems were publicly released, following the same two five-hour sessions, internet-access ban, and submission-frequency limits. It scored 535.4 points. [Official statistics](https://stats.ioinformatics.org/olympiads/2026) list the highest human score as 498.27 and the gold-medal cutoff as 361.12. However, the [paper](https://arxiv.org/abs/2609.02849) explicitly states that the model was not an official contestant and that the run was not supervised by the IOI, so the result should be treated only as an unofficial, self-reported benchmark from the authors.
The resource disparity is even more notable from an engineering perspective: the live pipeline was provisioned with as many as 760 GB300 GPUs, trading a large volume of candidate solutions for a higher score within a fixed time budget. This was not an equal-resource comparison between a model and a single human. The competition checkpoint and NeMo Skills inference recipe have yet to be released, while the complete training dataset is constrained by third-party licensing. The next questions are whether external teams can reproduce the score, reduce the search budget, and transfer GenCorrect to tasks such as software repair, where reliable test oracles are less readily available.