At IMO 2026, NVIDIA's Nemotron system competed officially and scored 30 out of 42, one point above the 29-point gold cutoff, with every submission graded by official IMO graders: full credit on Problems 1, 2, 4 and 5, one point each on Problems 3 and 6. Unlike the AlphaProof line of work that leans on formal provers, this system runs entirely in natural language — no formal prover, no external tools, no internet. It takes the official LaTeX problem statements and produces human-readable proofs. On September 9 the team posted the full pipeline as a technical report (arXiv:2609.10712), turning the gold recipe into something reproducible.
Three checkpoints and a 16-vote unanimity rule
The backbone is an iterative generate-verify-refine loop. Generation uses three Nemotron 3 Ultra checkpoints: the general-availability base model plus two specialists post-trained with SFT and RL respectively. Round one produces 384 proof attempts per problem — each checkpoint samples 16 attempts from each of eight complementary prompts that steer different strategies such as lemma-first decomposition, route comparison, or counterexample search.
Verification is handled by the RL and SFT checkpoints acting as reference-free verifiers. Every proof receives 16 independent judgments, scoring 1 for complete and correct, 0.5 for minor errors, 0 for fatal ones; a proof is accepted only when all 16 judgments award a full score. If nothing passes, the system picks the top-16 proofs from the pool, attaches up to eight verifier critiques, and sends them back for refinement — 192 fresh attempts per round, at most eight rounds.
The post-training gain, in hard numbers
On the 30-problem development set both specialists beat the base model: after eight rounds plus fallback, the RL pipeline reaches 180 cumulative points and SFT 165, against 162 for GA (independent-jury scoring). The paper also weighs how to spend the search budget: a single RL model pushed to 256 attempts accepts 14 problems, while a mixed pool of RL 128 + SFT 128 accepts 18. The team's takeaway: a second checkpoint solves problems the first cannot, and doubling attempts of a single checkpoint adds little.
The compute ledger
The four full-credit proofs cleared final selection within 76 minutes of the contest start, the remaining two within 100 minutes. All six submitted proofs were found within roughly 707M generated tokens and 1,464 GB200 GPU-hours; completing the in-flight rounds brought the full run to about 2.31B tokens and 4,800 GPU-hours. The post-hoc analysis carries a warning: at the cutoff both model-based verifiers estimated roughly 32 points, two above the official 30, with the gap entirely on Problems 3 and 6 — model juries shared a blind spot rather than noising independently.
Everything open, 561B weights downloadable
The HF collection (nvidia/nemotron-labs-imo-2026) holds the two 561B specialist checkpoints (OpenMDW-1.1 license), 133k SFT samples plus 9.6k RL problems (CC BY 4.0), and Nemotron-IMO-Bench, 200 novel olympiad problems built with Professor Titu Andreescu; inference code lives in NeMo-Skills, the RL recipe in NeMo-RL, submitted proofs included. The conclusion is worth quoting: scaling proof generation alone is not enough — the gains came from complementary post-trained checkpoints, verification-guided refinement that preserves candidates, and compute spent on final selection, while fancier routing and aggressive filtering discarded candidates that later turned out correct.
For anyone building reasoning systems, the value is not the 30 points themselves but that every step ships with ablation numbers and a compute bill you can check against.