MIT and the International Math Olympiad release MathNet, a dataset 5x larger than any previous collection of olympiad math problems and solutions from 40+ countries spanning 4 decades

MIT CSAIL and the International Math Olympiad organization published MathNet, a public dataset containing International Math Olympiad (IMO) problems and their solutions sourced from over 40 countries across 4 decades. The dataset is 5 times larger than previous publicly available collections and is freely accessible on Hugging Face.

Unlike smaller math datasets used to train AI systems, MathNet's scale and breadth—covering competition problems from dozens of countries over decades—gives AI researchers and engineers a richer source to teach machines how humans solve complex, multi-step mathematical reasoning problems. This addresses a major bottleneck: most AI systems struggle with the kind of deep logical reasoning required in olympiad-level math.

For students and educators, this dataset enables the development of better math tutoring AI tools and homework assistants that understand harder problem types. For AI researchers, it creates a common benchmark to measure whether new AI reasoning techniques actually work on genuinely difficult math, not just textbook exercises.