🔍 Read the full analysis: Exploring Nemotron Fine-Tuning For Gold-Level IOI And IMO Results on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Hugging Face says specialized systems based on its Nemotron 3 family scored 535.4 out of 600 in an unofficial 2026 IOI run and received 30 out of 42 from official graders for IMO proofs. The reported IMO score exceeded the stated gold threshold; the IOI result was not part of the competition’s official ranking.
As detailed in the original analysis, Hugging Face says two systems built from its Nemotron 3 model family achieved gold-level scores on tasks from the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO). The IMO proofs received official grading, while the IOI score of 535.4 out of 600 came from an unofficial run and was excluded from the competition’s official ranking.
For the IOI work, Hugging Face fine-tuned a competition-specific version of Nemotron-3-Ultra-CC using supervised fine-tuning, then paired it with GenCorrect, a process that generates candidate programs, evaluates them and revises solutions. The company reports that the system operated prospectively under the same time, internet-access and submission constraints as human competitors. Its score was above the stated 361.12-point gold threshold and the top human score of 498.27. Because the run was unofficial and unsupervised, however, it did not count toward official standings.
The IMO system used a different approach, combining the general Nemotron 3 Ultra model with supervised fine-tuning and reinforcement-learning checkpoints. It generated natural-language proofs, scored and critiqued candidate answers, and revised selected attempts. Hugging Face says official graders awarded the submitted proofs 30 of 42 points, above the stated gold threshold of 29, with full credit on four of six problems. The company says the system used no formal prover, external tools or internet access.
The training materials also differed by subject. Hugging Face reports using 22,000 programming problems and synthetic reasoning traces for IOI fine-tuning. For IMO, its supervised fine-tuning data contained 414,890 quality-filtered examples from 15,818 proof problems; the reinforcement-learning model was trained on 9,597 problems selected near the model’s reported capability frontier. These figures and performance results are reported by the company.
Different Scores, Different Validation
The results offer a reported example of how domain-specific fine-tuning and repeated candidate checking can be combined for difficult, tightly scored tasks. The programming system had to produce code that performed well against evaluation tests; the mathematics system had to produce written proofs that graders judged for correctness. Those are distinct capabilities, so the two scores should not be treated as equivalent evidence or as a single general measure of reasoning.
The distinction in validation matters. The IMO score was assigned by official graders, while the IOI score, despite being reported under competition-like constraints, remained outside the official ranking. Neither result by itself shows how the systems would perform on a wider range of programming work, unfamiliar proof styles or other competitions. The report is also from the organization involved in developing the systems, making independent scrutiny relevant to judging how broadly the findings apply.
As an affiliate, we earn on qualifying purchases.
From IOI Experiments to Two Subjects
Hugging Face describes the 2026 projects as an extension of experiments it reported from IOI 2025, where it combined model post-training with extra computation during answer generation. The company says a Nemotron-3-Nano-CC model scored 130 before post-training, 280 after supervised fine-tuning and 291 after reinforcement learning. With GenCorrect, it reached 468, above the stated 2025 gold threshold of 438.3. An Ultra-CC version scored 502 using the same test-time strategy, according to the company.
The newer work applied related methods to two separate competition formats: executable programs evaluated on hidden tests and mathematical arguments assessed as proofs. Hugging Face says the IMO project found complementary strengths in its supervised fine-tuning and reinforcement-learning checkpoints, leading it to combine them with the general model rather than depend on a single checkpoint. It characterizes the broader approach as adapting a shared base model into specialist systems instead of training a separate foundation model for each subject.
The company says it is releasing an IMO 2026 collection that includes supervised fine-tuning and reinforcement-learning checkpoints, both training datasets and Nemotron-IMO-Bench, a benchmark of 200 olympiad-level problems. It also refers to an IMO paper and a NeMo-Skills repository. These materials may let researchers examine the methods and test performance beyond the reported competition problems.
“Success at both points to something broader.”
— Hugging Face
As an affiliate, we earn on qualifying purchases.
Limits of the Reported Scores
The supplied account does not describe an independent audit or replication of the IOI run. It reports that the system ran prospectively without supervision under stated competition constraints, but leaves how the run was audited unclear. Its unofficial status means the result is best understood as a company-reported benchmark score, not an official IOI medal or ranking.
The report also does not establish whether either system generalizes to other competitions, new proof styles or real-world programming workloads. While Hugging Face lists intended releases, the available source material does not provide a full release timetable or enough detail to assess every component of the evaluation setup. The reported scores and training quantities should therefore be attributed to the company unless independently confirmed.
machine learning model fine-tuning tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Checkpoints and Replication Ahead
Researchers can use the stated release of IMO checkpoints, datasets and the 200-problem benchmark to inspect the training approach and run additional tests. Hugging Face also points to a paper describing the IMO training and generate-verify-refine system, along with its NeMo-Skills repository. The supplied material does not specify when every item will be available or give a timeline for further results.
For the IOI claim, an independently scrutinized or replicated evaluation would help clarify how the reported score holds up beyond the company’s run. Until then, the key distinction remains: officially graded IMO proofs earned a reported gold-level score, while the IOI performance was a gold-level result in an unofficial benchmark run.
As an affiliate, we earn on qualifying purchases.
Key Questions
What scores did the Nemotron systems report?
Hugging Face reports 535.4 out of 600 for the IOI system and 30 out of 42 for the IMO system. The IOI score came from an unofficial run; official IMO graders evaluated the submitted proofs.
Did the IOI system win an official gold medal?
No. The reported IOI score exceeded the company’s stated gold threshold, but the run was unofficial and excluded from the official ranking. It should not be described as an official competition medal.
How was the IMO result evaluated?
Hugging Face says official IMO graders awarded 30 of 42 points to the system’s proofs, above the stated gold threshold of 29. The system received full credit on four of six problems, according to the company.
What methods did the systems use?
The IOI system combined supervised fine-tuning with GenCorrect, which generates, evaluates and revises candidate programs. The IMO system combined the general model with supervised fine-tuning and reinforcement-learning checkpoints, then generated, critiqued and revised natural-language proofs.
What remains uncertain about the claims?
The supplied report does not describe independent verification of the IOI run, and it does not show how either system performs across broader programming or mathematics tasks. Further replication and evaluation would help establish how far the results generalize.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
