MIT scientists build the world’s largest collection of Olympiad-level math problems, and open it to everyone

Introducing MathNet: A Landmark in Math Problem Datasets

Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), King Abdullah University of Science and Technology (KAUST), and HUMAIN have launched MathNet, an expansive, high-quality dataset featuring over 30,000 proof-based math problems and expert solutions. Covering competitions from 47 countries, 17 languages, and 143 contests, MathNet sets a new standard for accessible mathematical resources and is five times larger than any similar dataset to date.

Diversity and Depth Across the Globe

Unlike previous collections, which emphasized problems from just a few countries, MathNet spans six continents and nearly four decades, capturing unique mathematical approaches and traditions. Both text and image-based problems are included, making the collection valuable for students preparing for Olympiad competitions and for researchers developing advanced AI models in mathematical reasoning.

Rigorous Curation and Verified Solutions

The project involved gathering and digitizing 1,595 PDF booklets—over 25,000 pages—many sourced from long-standing IMO community members. All problems were sourced from official competition booklets, ensuring high-quality, peer-reviewed, and detailed solutions. More than 30 evaluators worldwide contributed to meticulously verifying thousands of answers for accuracy and completeness.

Challenging AI and Supporting Students Worldwide

MathNet is designed to be both a training ground for students and a rigorous benchmark for AI. Leading AI models, including GPT-5, were tested on its problems, achieving an average success rate of about 69%, with lower performance on visually intensive and less-common language problems. The dataset also introduces retrieval benchmarks to gauge AI’s ability to understand core mathematical structures and analogies.

A Resource for the Whole Community

By making MathNet publicly available, the team hopes to level the playing field for students worldwide, especially those training independently. The dataset’s diversity is central to advancing both human and AI mathematical reasoning. MathNet can be accessed at mathnet.csail.mit.edu.

For more details, view the original article by Rachel Gordon at MIT News.