Nemotron’s Open Recipe for IMO Gold Combines Training with Test-Time Search

Introduction

Solving an olympiad problem is not the same as producing a proof that can earn points. A system must find a viable idea, preserve correctness across a long chain of reasoning, and present the argument clearly enough for evaluation. The paper An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics studies how model post-training and test-time inference can be combined to address that challenge.

Key points

• Specialized post-training. Starting from Nemotron 3 Ultra, the researchers trained two specialist checkpoints with supervised fine-tuning and reinforcement learning. Their target was natural-language proof generation for difficult olympiad mathematics.
• Verification as an iterative process. Rather than relying on a single completion, the system generates candidate proofs, checks them, and refines them. Multiple Nemotron checkpoints participate in this loop, allowing later attempts to respond to earlier verification signals.
• A separate selection stage. The pipeline uses a high-compute stage to select the final submission. This places test-time budget alongside model training as a central design variable for difficult reasoning tasks.
• No external mathematical machinery. According to the paper, the system operates entirely in natural language and does not use a formal prover, outside tools, or internet access. Its approach is therefore closer to a language-model inference workflow than to a tool-augmented theorem-proving system.
• A broad open release. The authors release two post-trained checkpoints, training data, training and inference code, submitted solutions, and Nemotron-IMO-Bench, a collection of 200 novel olympiad-level problems.

Results and limitations

The paper reports that the system earned 30 of 42 points at IMO 2026, meeting the stated gold-medal threshold. The significance of the result lies not only in the score, but also in the experimental framing. The work treats checkpoint choice, verification, refinement, and inference budget as components that can be studied together rather than as isolated tricks.

The result should still be interpreted carefully. Natural-language verification can miss subtle logical errors, while iterative search may require substantial computation. The supplied summary does not establish the contribution of each component independently, nor does it show how reliable every proof was across individual problems. The new benchmark could make those questions easier to investigate on held-out tasks.

Why it matters

The project presents a practical recipe for mathematical reasoning: combine specialist training with self-checking, repeated revision, and deliberate allocation of test-time compute instead of expecting one model response to be correct on the first attempt. If the released assets are reproducible, researchers can study where the gains come from and whether the same architecture transfers to theorem proving, scientific reasoning, or educational assessment.

Source: Hugging Face Daily Papers