Nemotron 3 Ultra pipeline achieves IMO Gold with open-weight release
NVIDIA has released an open-source framework that enables its Nemotron 3 Ultra model to solve Olympiad-level mathematics problems at a gold-medal standard. The system utilizes a test-time compute pipeline that iteratively generates, verifies, and refines natural-language proofs without relying on formal provers or external tools.
Key points
- Nemotron 3 Ultra pipeline scored 30/42 at IMO 2026, meeting the gold-medal threshold.
- System uses iterative natural-language proof generation, verification, and refinement without formal tools.
- NVIDIA released two specialist checkpoints, training data, code, and a new 200-problem benchmark.
The approach involves training two specialist checkpoints from the base Nemotron 3 Ultra model using supervised fine-tuning and reinforcement learning. By combining these specialists with the general-availability model in an iterative search process, the system achieved a score of 30 out of 42 points at the International Mathematical Olympiad (IMO) 2026, surpassing the threshold for a gold medal.
To support further research, the team has released the post-trained checkpoints, training data, inference code, and submitted solutions. They also introduced Nemotron-IMO-Bench, a new benchmark consisting of 200 novel olympiad-level problems, providing the community with resources to evaluate and improve mathematical reasoning capabilities in large language models.
Model page: Nemotron 3 Ultra →
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- PyTorch tutorial shows how to cut drone data usage by 94% · 1 src
- Article outlines five practices for building robust Python AI libraries · 1 src
- Vector databases explain how AI retrieves similar data at scale · 1 src
- PhAI Labs introduces JEPA-Anything framework for cross-domain world models · 2 src
- Researchers introduce VIMA for finding disease patterns in tissue data · 2 src
Comments
via GitHub Discussions