AI Just Won Gold at the World's Toughest Math and Coding Competitions

Nvidia announced that its Nemotron AI achieved gold-medal-level results in two of the world's hardest competitions for young people: the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). The IMO requires writing rigorous proofs, while the IOI involves solving complex coding problems under time pressure. Nemotron's success shows that a single AI foundation can be adapted to excel at very different tasks.

The team used a four-step recipe: supervised fine-tuning (showing the AI many examples), reinforcement learning (rewarding good answers), and feedback-driven inference (letting the AI check and improve its work). For coding, they trained on 22,000 problems and used a strategy called GenCorrect to iteratively refine solutions. For math, they trained on over 400,000 proof examples, teaching the AI to generate, critique, and verify proofs. The final system scored 30 out of 42 points at IMO 2026, earning full marks on four problems.

This matters because it shows AI can be specialized for demanding domains without building a new model from scratch. The same approach could be used to create expert AI assistants in law, medicine, or engineering. The AI worked entirely in natural language, without internet or external tools, making it more practical for real-world use. While these were unofficial benchmarks, the results suggest that AI is getting better at reasoning and problem-solving.

For everyday people, this means future AI tools could help with complex tasks like debugging code, writing proofs, or analyzing data. It could also lead to more personalized tutoring systems. However, the AI still makes mistakes, and these competitions are just one measure of ability. The real test will be how well it performs in messy, real-world situations.