MIT Builds Free Math Library With 30,000 Competition Problems

For years, countries competing in the International Mathematical Olympiad would bring booklets of their most creative math problems, share them with other teams, then watch them quietly disappear into obscurity. No central archive existed for students training alone or researchers pushing the boundaries of AI.

MIT researchers just changed that. They've released MathNet, the world's largest collection of high-quality math competition problems, completely free and open to everyone.

The dataset contains over 30,000 expert-authored problems and solutions spanning 47 countries, 17 languages, and 143 competitions across four decades. It's five times larger than any similar collection and covers traditions from six continents, not just the usual suspects like the United States and China.

Building it required an extraordinary effort. The team tracked down 1,595 PDF volumes totaling more than 25,000 pages, including decades-old scans in multiple languages. Much of the archive came from Navid Safaei, a longtime IMO community member who had been personally collecting and scanning these booklets by hand since 2006.

What makes MathNet special isn't just size but quality. Unlike datasets pulled from online forums, these problems come from official national competition booklets with peer-reviewed, expert-written solutions. Many solutions run multiple pages and show several approaches to the same problem, giving learners rich insights into mathematical thinking.

"I remember so many students for whom it was an individual effort," says Shaden Alshammari, an MIT PhD student who competed in the IMO herself and led the project. "We hope this gives them a centralized place with high-quality problems and solutions to learn from."

The team assembled over 30 graders from countries including Armenia, Russia, Ukraine, Vietnam, and Poland to verify thousands of solutions. They're now working to share the dataset directly with the IMO foundation.

The Ripple Effect

The impact reaches beyond students. AI researchers now have a tougher benchmark for testing mathematical reasoning, and early results are humbling even for the best models. GPT-5, the top performer tested, solved about 69 percent of problems, failing nearly one in three. When problems included figures, performance dropped significantly across all models, revealing visual reasoning as a persistent weak point.

Switzerland's IMO deputy leader Tanish Patil notes the database "has the potential to be an excellent resource" that other archives lack, with standardized formatting, verified solutions, and important problem metadata.

Students preparing for competitions worldwide now have what was once scattered across filing cabinets and personal collections: a searchable library of the world's best mathematical thinking, all in one place.