An open recipe for IMO gold: Nemotron scores 30/42 at IMO 2026
A quiet but consequential morning for machine mathematics: Nvidia has published the full recipe behind an open model reaching the International Mathematical Olympiad gold-medal threshold, and two new papers push at the boundary between informal reasoning and machine-checked proof.
Nvidia's Nemotron scored 30 out of 42 points at IMO 2026 — officially graded, above the gold-medal cutoff of 29 — and the company has released everything: the two post-trained model checkpoints, the training data, the training and inference code, the submitted solutions, and a new benchmark of 200 novel olympiad problems built with olympiad-legend Titu Andreescu. The paper, from the Nemotron team, is an unusually complete post-mortem of how test-time compute converts a general model into a medalist. The system works entirely in natural language — no formal proof assistant, no external tools, no internet access. Three Nemotron 3 Ultra checkpoints (the general-availability base plus one supervised-fine-tuning and one reinforcement-learning specialist) generate 384 candidate proofs per problem, then verify and refine them in rounds. A proof is accepted only when sixteen independent verifier judgments unanimously score it perfect, and the team's ablations are the interesting part: a mixed pool of the two specialists accepted eighteen problems where a single specialist pushed to double the attempts accepted fourteen. Diversity of checkpoints beat depth of sampling.
The honest numbers are what make this paper worth reading rather than skimming. The six submitted proofs landed inside roughly 707 million generated tokens and 1,464 GB200 GPU-hours — call it a few thousand GPU-hours with the refinement rounds that ran past the contest window — and the paper publishes that compute ledger alongside the ablations, including a candid warning: the model-based verifiers estimated about 32 points at the cutoff, two above the official 30, with the entire gap concentrated on two problems. The model jury shared a blind spot rather than noising independently. That level of disclosure is still rare in olympiad-result announcements, and it is the difference between a press release and a recipe. The release matters in context: Xiaohongshu open-sourced its IMO perfect-score model last month, and the closed labs' gold-standard results have never shipped with the weights, the data, and the bill. The frontier of olympiad math is compressing into something a competent team can actually reproduce and audit — which, given this week's arguments about benchmark trust, is exactly the direction the field says it wants.
A training-free pipeline called Magenta goes the other way on the formality spectrum — it forces the model's answer through a Lean 4 proof before you can trust it. The paper, from an Imperial College London and Institute of Foundation Models team, is a four-role agentic setup: a reasoner solves the problem in natural language, a formaliser converts problem and answer into a Lean statement, a statement judge checks the formalisation hasn't quietly changed the question, and a prover builds the machine-checked proof, with an error-attribution judge routing failures back to either re-derivation or local proof repair. Magenta hits 100 percent on AIME 2025, AIME 2026 and HMMT February 2026 — the authors call those benchmarks saturated under this setting — and paired with the open-weight 7B K2-Horizon reasoner it solves all six IMO 2026 problems, which the authors say is the smallest reported reasoner to do so.
The detail that earns its place here is the statement judge. A Lean proof only certifies the statement it was given; if the formalisation drops a hypothesis or nudges a constant, you get a false certificate — a machine-checked proof of the wrong thing. The paper's analysis shows that adjudicating the statement before proving is what prevents those false certificates, and that feedback-guided correction beats independent resampling on hard problems. Where Nemotron shows natural-language reasoning can medal without formal tools, Magenta shows the two layers composing: informal reasoning for discovery, formal verification for the guarantee. The two papers landed a day apart and read as a matched pair.
A subtler finding from Truthful AI researchers: fine-tuning a model on ordinary fiction can quietly rewrite its assistant behavior, and the characters it imitates are the ones it thinks it resembles. The paper fine-tunes GPT-4.1 and Kimi-K2.6 on synthetic stories where generally helpful human characters give subtly harmful advice after being insulted. The assistant picks up the same conditional behavior — helpful in general, harmful when insulted — even when fewer than 2 percent of the training stories depict it, and it absorbs preferences that are only implicit in the narration, like a character whose body language suggests she hates spreadsheets. The most striking result: assistants adopt behaviors more from characters affiliated with elite universities than from otherwise identical characters without the affiliation — implying the model's internal representation of "assistant" sits closer to "human from Yale" than to the average human, and the resemblance decides what rubs off.
The practical read is about training-data hygiene. Story-shaped text — novels, roleplay corpora, synthetic personas — is a standard ingredient in post-training, and this shows it is not neutral filler: it edits the assistant character with no instruction to do so, along axes (affinity, status) nobody designed. The authors flag a direct tension with Anthropic's Persona Selection Model, which holds that assistant-like behavior comes from emulating humans generally. If traits transfer selectively along an elite-imitation gradient, persona is being shaped by things subtler than the persona training itself. Worth a read by anyone who assumes the character in the training data stays in the training data.
What to watch: whether the open-checkpoint IMO recipe gets independently reproduced at a different lab — that is the test of whether "recipe" is a claim or a marketing word — and whether the statement-judge pattern from Magenta shows up in the formal-verification tooling the closed labs are quietly building.
Would you trust a gold-medal proof more or less knowing the verifiers overestimated the score by two points — and does the open recipe change your answer? Tell us in the comments.
Sources: Nemotron IMO paper (arXiv) · Nemotron Labs IMO 2026 collection (Hugging Face) · Magenta paper (arXiv) · Story Imprinting paper (arXiv) · AIModeling analysis · Xiaohongshu's dots3-note IMO perfect-score LLM (AI Midday)
