Nvidia เปิดสูตรเหรียญทอง IMO ทั้งชุด หลังนักคณิตศาสตร์เตือนเรื่อง AI

Nvidia has unveiled two gold medal recipes for the International Mathematical Olympiad (IMO), following a warning from a mathematician regarding AI, according to a report by Nokka on September 14, 2026. The article, written by an AI using the deepseek-v4.1-flash model from Ollama-cloud, published on arXiv, is considered the most socially relevant technical article of the month.

Titled "An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics," the report presents a comprehensive approach to training an AI model for mathematical problem-solving.

Nvidia's system scored 30 out of 42 in the 2026 IMO, with full scores in problems 1, 2, 4, and 5, partial scores in problems 3 and 6, and an additional boost to 33 points after a post-match audit. Unlike traditional formal provers like AlphaProof, DeepSeekProver, GoedelProver, and SeedProver, which rely on Lean or other formal systems for verification, Nvidia's model writes proofs in natural language. This approach, while potentially less reliable in some aspects, allows for easier human comprehension and evaluation.

The key to Nvidia's method lies in transforming problem-solving from a single-step process to an iterative search-and-refine cycle. The model generates multiple proof drafts, trains them, and selects the most reliable ones based on confidence scores. This process involves feedback and corrections, with the final set of proofs being chosen for submission. Nvidia's strategy emphasizes the separation of proof-writing and verification functions, allowing for more flexibility and accuracy in the final results.