ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

How NVIDIA Explains the Co-Design of Specialization and Inference Search to Achieve Gold-Medal Performance in Models

The NVIDIA team detailed in a blog post the process of specializing Nemotron 3 into systems for the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). The core finding is that fine-tuning and inference search must be co-designed: specialization provides better candidate answers and critics, while the inference loop amplifies these specialization gains; both are indispensable. The article also reveals scale-dependent differences in specialization paths and the complementary roles of SFT and RL checkpoints in search.

2026-10-07 20:45 GMT+8

The skills tested by the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO) barely overlap: the former requires writing code that passes hidden tests under strict time and submission limits, while the latter demands constructing rigorous natural-language proofs. NVIDIA reported in a blog post that the same Nemotron 3 base model, following different specialization paths, reached gold-medal level in both competitions—scoring 535.4/600 on IOI, surpassing the gold threshold of 361.12 and the top human score of 498.27; and scoring 30/42 on IMO, exceeding the official gold threshold of 29. However, the team believes what is truly worth noting is not the scores themselves, but the methodological conclusions drawn from them.

The article's core judgment is that medals are not produced solely by fine-tuning, nor solely by brute-force sampling, but by the co-design of the model, data, and inference loop. This means the value of specialization lies not only in the direct output quality of the model itself, but more importantly in providing the inference search system with better candidate answers, better critics, and better directions for correction. Conversely, the inference loop amplifies these specialization gains through multi-round feedback into larger overall improvements.

This co-design logic is most intuitively reflected in the IOI project. The team first applied Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to the Nano model (30 billion total parameters, 3 billion active parameters), raising its score from 130 to 291; then, by adding the GenCorrect iterative generate-evaluate-refine strategy, Nano reached 468 points, crossing the gold threshold of 438.3. The Ultra model (550 billion total parameters, 55 billion active parameters) underwent only SFT, yet achieved 502 points when paired with the same inference strategy. A notable finding is that for the stronger Ultra model, a single SFT epoch was sufficient to outperform the fully post-trained (SFT plus RL) Nano model on IOI, ICPC, and LiveCodeBench Pro. This indicates that specialization does not require adopting the same path at every scale.

Sign in / Sign up
ReadingResearchRadarInvestment framework

The IMO project further revealed the complementary roles of different specialization types within the inference loop. Starting from Nemotron 3 Ultra, the team trained separate SFT-specialized models and RL-specialized models. The SFT corpus contained 414,890 quality-filtered samples covering 15,818 independent proof problems; the data taught not just final answers, but also covered proof generation, refinement, verification, and meta-verification—the model learned to construct arguments, identify loopholes, respond to criticism, and judge whether a proof was complete. The RL model was trained on 9,597 proof problems close to the model's capability frontier. In development experiments, the SFT checkpoint performed strongest in the first round of search (providing higher-quality initial candidates), while the RL checkpoint achieved the best overall single-checkpoint results. The final system used both alongside the general model: generating candidate proofs for each problem, scoring them, producing critiques, refining the most promising attempts, and finally selecting the submission in a separate high-compute stage.

The team specifically emphasized that the entire IMO system operated entirely in natural language, without using formal provers, external tools, or internet access. This excluded confounding factors from tool assistance, making the results more directly reflective of the interaction between model specialization and inference search.

Important limitations to retain: The IOI results were obtained from a live prospective run under the same time, internet access, and submission constraints as human contestants, but were explicitly marked as an unofficial, unsupervised benchmark and were not included in the official IOI rankings. The IMO submitted proofs were graded by official IMO graders. All figures are self-reported by the NVIDIA team, with no independent replication yet available. The team simultaneously released model weights, training datasets, a Nemotron-IMO-Bench benchmark containing 200 olympiad-level problems, and reproducible inference pipeline code, providing conditions for subsequent verification.

For practitioners, the practical implication of this article is: when evaluating whether it is worthwhile to invest in more specialized training, one should not only look at the direct score improvement on fixed benchmarks, but also consider whether it brings higher-quality initial candidates and more effective self-criticism to the inference search system. Specialization and test-time compute are not a binary allocation problem of investment; rather, they are a system issue requiring joint design.

Sources

  • huggingface.co↗
Continue researching →