Nemotron-IMO-TTS

Nemotron-IMO-TTS is the natural-language proof pipeline that scored 30 of 42 points at IMO 2026, released with the report An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics. Three Nemotron 3 Ultra checkpoints (the general-availability model and the released RL and SFT specialists) generate candidate proofs from eight complementary prompts, a two-checkpoint verifier panel scores every candidate and writes critiques, and up to seven refinement rounds revise the best candidates. A separate high-compute stage then grades the finalists with an IMO-style judge prompt and picks the proof to submit.

The recipe talks to any OpenAI-compatible chat-completions endpoint (a hosted API, a gateway, or local vLLM servers), persists every request as a resumable JSONL row, and ships the report's prompts byte for byte.

For the walkthrough, configuration reference, output layout, and resume semantics see recipes/nemotron-imo-tts/README.md in the repository. The proofs submitted at IMO 2026 are in recipes/nemotron-imo-tts/imo-proofs.

Run it

cp recipes/nemotron-imo-tts/configs/imo2026-ensemble.yaml my-run.yaml   # fill in the <your-...> placeholders
export NVIDIA_API_KEY=...
python recipes/nemotron-imo-tts/run.py --config my-run.yaml --output-dir runs/my-run --dry-run
python recipes/nemotron-imo-tts/run.py --config my-run.yaml --output-dir runs/my-run

runs/my-run/submissions.jsonl holds the selected proof per problem; results.jsonl holds the full record. Relaunching the same command resumes from the persisted rows.