Fields Medalists warn AI benchmark race harms math as Nvidia opens its IMO gold recipe

Twenty-five Fields Medal recipients warned on September 11, 2026 that AI companies racing to crack mathematics problems for benchmark credibility are damaging the discipline itself. Their joint statement, published at mathandai.org under the title "A Severe Misalignment of AI in Mathematics", closed out a week that also saw OpenAI claim a 90-year-old problem and Nvidia post a fully open recipe for a gold-medal olympiad system.

What the Fields Medalists said, and did not say

The signatories do not dispute AI's capability. According to the statement's opening, they acknowledge that large language models have advanced to the point of solving significant open problems in several fields. What they object to is the process: answers are frequently announced in haste, leaving no time for a proper write-up, for extracting the new methods and ideas, or for crediting earlier work. The statement says this raises serious questions about credit and plagiarism.

The deeper argument concerns how mathematics produces knowledge. Landmark open problems mark where human understanding has reached, the statement explains, and a solution only matters insofar as it yields ideas that others can build on. Converting a correct answer into transferable knowledge requires the community's digestion — lectures, debate and simplification down to the level of an undergraduate textbook. If that step is skipped, the result stays an answer rather than becoming knowledge.

Terence Tao, one of the signatories, wrote on his blog that the statement grew out of discussions among the medallists during the preceding week, and he conceded there had been no time for a broader comment process. The 25 names are initial signatories and the statement remains open for more. It closes by allowing that AI could genuinely elevate mathematical understanding, while warning that whether the technology elevates or destroys will be decided largely by the humans controlling it.

The week that produced the statement

The timing was not coincidental, as a dev.to analysis by Nokka lays out. OpenAI announced on September 8 that roughly 10,000 agents had solved the Navier–Stokes problem in 88 hours. The claim drew an immediate credit dispute: Tristan Buckmaster of NYU had issued a statement on September 7 saying that he and Levent Alpöge, a mathematician working at Anthropic, had been on the problem for more than a year, with evidence their progress had leaked to OpenAI. Per CNBC and BBC reports cited in the dev.to post, OpenAI's own announcement acknowledged the effort began after hearing rumours that were later found to concern the two researchers.

Nvidia's open recipe for an IMO gold

Between those events, on September 9, Nvidia placed a paper on arXiv titled "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics". According to the paper, the system scored 30 out of 42 at IMO 2026, clearing the 29-point gold threshold, with full marks on problems 1, 2, 4 and 5 and partial credit on 3 and 6. An informal post-competition recheck that credited proofs found after the fact moved the total to 33, but the figure to cite remains 30.

The system is notable for what it omits: no formal prover, no external mathematics tools, no internet access — proofs are written entirely in natural language. That is the opposite of approaches such as AlphaProof or DeepSeekProver, which check every step mechanically in Lean. The paper states the trade-off plainly: a natural-language verifier can endorse faulty arguments.

The pipeline runs as an iterative search: generate candidate proofs, check them, send rejected ones back for revision based on critique, then select. Three checkpoints derived from Nemotron 3 Ultra 550B-A55B — a general-availability model, a supervised fine-tune and a reinforcement-learning tune — play distinct roles, on the reasoning that a model skilled at writing proofs is not necessarily skilled at spotting flaws. During search, acceptance requires every judge on a model committee to agree a proof is valid and simultaneously assign it the top correctness score; a single dissent sinks the candidate.

The training effort is large. Supervised tuning started from 15,879 hard proof problems drawn from the AoPS portion of Nemotron-Math-Proofs-v1, with external models generating answers up to 400,000 tokens long and up to three rounds of revision, yielding a filtered set of 414,890 examples across 15,818 unique problems. Nvidia released two trained checkpoints, the training data, training and inference code, its actual competition submissions, and a benchmark suite called Nemotron-IMO-Bench.

Caveats

The 30/42 score is self-reported by the team rather than independently administered, and the verification weakness the paper admits is structural to natural-language checking, not something more training fixes. No party has linked the paper to the medallists' statement — the pairing is the dev.to author's interpretation. What is factual is the calendar: Nvidia's paper on September 9, the OpenAI claim on September 8, the Fields statement on September 11.

Why it matters

The medallists' complaint is not that AI is bad at mathematics but that benchmark-driven haste is stripping out the mechanisms that turn solutions into shared knowledge — write-ups, attribution and community review. Nvidia's release demonstrates that an alternative exists: publish the method, the data and the actual outputs so others can check and extend them. The underlying question — whether credibility lives in the answer or in the process that produces and verifies it — now applies to every field AI is entering.