NVIDIA Opens a Natural-Language-Only IMO 2026 Gold Pipeline on Nemotron 3 Ultra, Weights and Proofs Included

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2609.10712 describes a system with no formal prover, tools or internet access. It runs three Nemotron 3 Ultra checkpoints (the GA model plus SFT and RL math specialists) in a generate-verify-refine loop and scored 30/42 at IMO 2026, the gold threshold. NVIDIA published Nemotron-3-Labs-Ultra-Math-SFT and -RL on Hugging Face, put the inference recipe in NeMo-Skills, and added Nemotron-IMO-Bench with 200 new olympiad problems. This is one of the first gold-level test-time-compute pipelines anyone can rerun on open weights.