The research investigated the impact of post-training design and test-time inference on natural language proof generation for challenging mathematics problems. Starting with Nemotron 3 Ultra, two specialist checkpoints were created through supervised fine-tuning and reinforcement learning. The system’s architecture involved an iterative search process, generating, verifying, and refining candidate proofs within a natural language environment. A separate high-compute stage then selected the final submissions. The system achieved a score of 30 out of 42 points at the 2026 International Mathematics Olympiad, meeting the gold-medal threshold.
The system operates entirely in natural language, without utilizing formal provers, external tools, or internet access. Three Nemotron 3 Ultra checkpoints were employed: the general-availability model and two post-trained specialists. The system’s performance is attributed to this iterative approach to proof generation and verification.
As part of this work, a new benchmark, Nemotron-IMO-Bench, consisting of 200 novel olympiad-level problems, was created and released. The research team also made available the two post-trained checkpoints, the training data, the training and inference code, and the submitted solutions.
This research provides a test-time-compute pipeline for generating and verifying mathematical proofs, offering a potential tool for problem-solving and educational applications.
Source: https://arxiv.org/abs/2609.10712