What it is
Open natural-language IMO solver built from post-trained Nemotron checkpoints plus high-compute generate-verify-refine search. It reached IMO 2026 gold without formal tools. Instead of relying on formal provers or a single model, it combines SFT- and RL-specialized Nemotron checkpoints in an iterative natural-language search that generates, verifies, refines, then re-ranks proofs with heavy test-time compute.
Key results
- 30/42 at IMO 2026, beating the 29-point gold-medal cutoff
- Full credit on Problems 1, 2, 4, and 5; 1 point each on 3 and 6
- RL eval accuracy rose to about 58% on a 133-problem proof set
Shows open-weight LLMs can reach IMO gold with specialist post-training and aggressive verifier-guided test-time search, even without formal tools.