NVIDIA's Nemotron model family has reached gold-level scores at two of the most demanding academic competitions for AI: the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). The results come from fine-tuning the same base model family separately for each contest, rather than building bespoke systems from scratch. This demonstrates that a single architecture, when adapted with the right training data and objectives, can excel in both algorithmic coding and formal mathematics.
The blog reports that the fine-tuned models performed at a level comparable to gold-medal-winning human contestants. For IOI, the focus was on competitive programming tasks that require efficient algorithm design and implementation. For IMO, the model had to solve proof-based problems in geometry, number theory, and algebra. The fact that both domains were tackled by the same family highlights the versatility of the underlying pretrained model.
While the source does not provide full technical details of the fine-tuning process, it emphasizes that the results are "gold-level" and that this is a significant milestone. The implication is that targeted fine-tuning on competition-style problems can push general-purpose language models toward expert-level reasoning, potentially opening the door to similar gains in other specialized fields. No competing claims or disagreements are present in the source, so the account stands as a single, consistent narrative.