研究與評測
NVIDIA Completes Its IMO Gold-Medal Reasoning Recipe: Generate 384 Proofs per Problem, Then Filter Them Through Multiple Rounds of Review
NVIDIA has updated its Nemotron IMO 2026 resources, releasing SFT/RL weights, datasets, prompts, and complete resumable inference code. The system scored 30/42, but the result depends heavily on three 561B-parameter checkpoints and massive test-time compute, so it should not be treated as the capability of a model answering in a single pass.
Over the past two days, NVIDIA has updated the Nemotron Labs IMO 2026 resource collection, adding the math-specific SFT and RL checkpoints described in its earlier paper, two training datasets, and the 200-problem Nemotron-IMO-Bench. It has also released the submitted proofs and the end-to-end inference recipe. The system solved all six problems from the 2026 International Mathematical Olympiad in natural language, scoring 30/42 and reaching the gold-medal threshold. It used no formal proof assistant, search engine, or external mathematical tools. The significance lies not only in the score: outsiders can now examine exactly how much candidate generation, peer evaluation among models, and selection compute went into the “gold-medal-level” result.
In the first round for each problem, three Nemotron 3 Ultra checkpoints—the general-purpose, SFT, and RL variants—each use eight prompts and sample 16 responses per prompt, producing 384 proofs in total. After deduplication, every proof is evaluated eight times by each of the SFT and RL models, and it is accepted only if it passes all 16 evaluations. In rounds two through eight, the system selects the best 16 proofs from the proof pool, incorporates up to eight pieces of reviewer feedback, and has each of the three models rewrite each proof four times, generating another 192 candidates per round. Finally, all three checkpoints grade the 16 finalist proofs 16 times each under the IMO’s 0–7 scoring rubric. Each candidate therefore receives 48 scores before the submission version is selected.
The released code can connect to any OpenAI-compatible endpoint. It records every request in resumable JSONL files and hashes the prompt, model, input, and configuration to prevent configuration drift when a run is resumed. The complete output also retains random seeds, source proofs, reviewer feedback, and token usage. This improves auditability at the pipeline level, but it does not make reproduction inexpensive: all three models have approximately 561B parameters, completions can be configured for up to 512K tokens, and the number of candidates generated across multiple rounds far exceeds that of a typical best-of-N evaluation. When comparing other IMO results, researchers should report the base model, total inference tokens, evaluator models, and selection strategy separately. The next step is for independent teams to rerun the six released problems as well as entirely new ones to determine whether evaluator self-consistency translates into a stable proof-correctness rate.