Digital Event Horizon
Nemotron Labs has achieved a groundbreaking breakthrough in competitive programming by fine-tuning and specializing the Nemotron model to achieve gold-level results in the IOI and IMO. The model's ability to pair specialist models with inference loops that generate, evaluate, and improve candidate answers is a key factor in achieving these results. This achievement demonstrates the potential of Nemotron to be adapted to demanding domains and composed with transparent inference workflows to solve complex problems.
Nemotron Labs achieved gold-level results in IOI and IMO with specialized models. The Nemotron model requires careful consideration of domain-specific problems, reasoning traces, and post-training methods like SFT and RL. Nemotron-3-Nano-CC and Nemotron-3-Ultra-CC showed significant improvement with SFT and RL, respectively. The use of a specialist model with an inference loop generated, evaluated, and improved candidate answers. The combination of a specialist model with a search, verify, and improve system led to the best outcomes. The Nemotron Labs' work demonstrates the potential of Nemotron to be fine-tuned into world-class domain specialists.
In a groundbreaking achievement, the Nemotron Labs has successfully fine-tuned and specialized the Nemotron model to achieve gold-level results in the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). This remarkable feat demonstrates the capabilities of the Nemotron model and its potential to be adapted to demanding domains with a clear, reusable recipe.
The Nemotron model, developed by NVIDIA, has been widely recognized for its ability to fine-tune and specialize with ease. The recent results show that fine-tuning Nemotron for IOI and IMO requires careful consideration of the domain-specific problems, high-quality reasoning traces, and post-training methods such as supervised fine-tuning (SFT) and reinforcement learning (RL). The model's ability to pair the specialist model with an inference loop that generates, evaluates, and improves candidate answers is a key factor in achieving gold-level results.
The IOI project involved curating 22,000 problems and generating synthetic reasoning traces to train two specialists. Nemotron-3-Nano-CC, with 30 billion total parameters and 3 billion active parameters, received both SFT and RL, while Nemotron-3-Ultra-CC, with 550 billion total parameters and 55 billion active parameters, received SFT. The progression on IOI 2025 makes the value of specialization easy to see, with Nano improving from 130 points before post-training to 280 after SFT and 291 after RL. With GenCorrect, the iterative generate-evaluate-refine strategy, it reached 468 points and crossed the gold threshold of 438.3. Ultra-CC reached 502 points with the same test-time strategy.
The IMO project applied the same idea to olympiad mathematics, starting from Nemotron 3 Ultra, and trained one specialist with SFT and another with RL. The SFT corpus contained 414,890 quality-filtered examples across 15,818 unique proof problems, covering proof generation, refinement, verification, and meta-verification. The RL model was trained on 9,597 proof problems selected near the model's capability frontier. Both post-trained checkpoints outperformed the general-availability model in the development experiments.
The best outcome came from combining a capable specialist with a system that could search, verify, and improve. The medals were not produced by fine-tuning alone, and they were not produced by brute-force sampling alone. The distinction matters, and it highlights the importance of co-designing the model, the data, and the inference loop.
The Nemotron Labs IMO 2026 collection brings together the SFT and RL checkpoints, both training datasets, and Nemotron-IMO-Bench, a new benchmark of 200 olympiad-level problems. The IMO paper describes the training approach and generate-verify-refine system, while the NeMo-Skills repository includes the IMO inference pipeline, prompts, submitted proofs, and a reproducible quickstart. For competitive programming, the Nemotron-3-Ultra-CC model is available on Hugging Face, and the IOI paper provides the training recipe and the GenCorrect methodology.
This achievement demonstrates the potential of Nemotron to be fine-tuned into world-class domain specialists, then composed with transparent inference workflows to solve problems at the frontier of human competition. The Nemotron Labs has made significant contributions to the field of competitive programming, and their work will undoubtedly have a lasting impact on the community.
Related Information:
https://www.digitaleventhorizon.com/articles/Nemotron-Labs-Achieves-Breakthrough-in-Competitive-Programming-with-Fine-Tuning-and-Specialization-deh.shtml
https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026
Published: Wed Oct 7 09:17:08 2026 by llama3.2 3B Q4_K_M
