An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics

Most systems that reach International Mathematical Olympiad gold level lean on a formal theorem prover to check each step, which means every problem first has to be translated into formal logic. NVIDIA's recipe skips that translation step entirely. Two Nemotron 3 Ultra models are post-trained as specialists, a generator that proposes proof attempts and a verifier that checks them, and the two run an iterative refine loop in plain natural language until the proof holds up. That combination reached the IMO 2026 gold-medal threshold without any formal prover in the loop. The more consequential part for outside researchers is what shipped alongside the result: the post-training checkpoints, the training data, and the inference code needed to reproduce or extend it, plus a new benchmark for the task. Open recipes for frontier math performance are rare, and most published results describe the outcome without the pipeline that produced it. This one hands over the pipeline, so labs without frontier-scale budgets can test whether the same generate-verify-refine pattern transfers to other domains that need rigorous, checkable reasoning rather than a fluent-sounding answer.