A unanimous verifier panel gates every proof in Nvidia's IMO gold recipe

A proof clears the search only when every verifier on the panel calls it valid and all of them award the top correctness score at the same time ^[10]. One objection sends the candidate back for revision. The panel's average score is not the criterion ^[10]. The account here comes from the paper by way of a dev.to post credited to Nokka and dated 14 September 2026. The post says the text was drafted by a language model and checked and edited by hand ^[1].

Each verifier on the panel is a different model. Nvidia splits three checkpoints out of Nemotron 3 Ultra 550B-A55B and treats them as complementary sources of mathematical behaviour, without assuming the strongest single model is enough ^[13]. The stated reason for the split is practical. A model that writes good proofs may be worse at spotting small defects, and a verifier cheap enough to filter a large pile is too generous for the final call ^[11]. So the pipeline runs different checkpoints and different decision policies at different stages. The wrapper is a loop that generates candidates, verifies them, sends the rejected ones back with critique, and then compares the survivors in a separate selection step ^[12][9].

The supervised set began with 15,879 hard proof problems drawn from the AoPS portion of Nemotron-Math-Proofs-v1, chosen on earlier pass-rate measurements so the model trained on problems it was failing ^[14]. External models wrote several solutions per problem, each output allowed to run to 400,000 tokens, with up to three refinement rounds when a solution came back incomplete ^[15]. Another set of models labelled correctness 0, 0.5 or 1, and a meta-verification step scored how much to trust that label ^[16]. After malformed, empty, incomplete and over-length outputs were dropped, 414,890 examples survived, from 15,818 unique problems ^[17].

That works out to roughly 26 retained samples per problem ^[1], and 61 of the selected problems left nothing behind ^[2]. At the cap, those 414,890 samples top out near 166 billion tokens of retained generation ^[3], before anything that was filtered out. The reinforcement stage kept problems the base model solved one to three times out of four according to an external judge ^[18]. It used the DeepSeekMath-V2 framework with the self-analysis reward removed, at 128 problems times 16 trajectories per batch, or 2,048 trajectories ^[19].

The 30 sits one point above the 29-point gold line ^[4], and the post notes the figure is the team's own report, with no independent committee score in this setting ^[23]. The paper concedes that its verifiers can agree with an incorrect argument ^[8]. For the number to hold, an outside grader has to read the released submissions and reach the same marks. The informal re-check that lifted the total to 33 is one run of that process, done with proofs found after the contest ^[5].

Two days after the paper went up, 25 Fields medallists said that AI companies rushing to solve mathematics problems for use as benchmarks is harmful to mathematical science ^[21]. Their objection was to the process, according to the statement: answers announced in a hurry, with no time left to write them up cleanly, separate out the new methods, and cite prior work ^[22].