The world's largest database of olympiad maths problems has been uploaded to the open access

We are talking about more than 30 thousand problems with complete solutions.

Details

The new archive is called MathNet. It was created by researchers from MIT, KAUST and HUMAIN.

The database includes:

• more than 30 000 problems and solutions
• materials from 47 countries
• 17 languages
• 143 maths Olympiads

According to the authors, this is the largest and highest quality open set of problems of this level. It is about five times larger than its closest analogues.

The main difference of MathNet is not only its size, but also the source of materials. The archive includes official collections of national Olympiads, not problems from forums and amateur sites.

The researchers collected 1,595 PDF files with a total volume of more than 25 thousand pages. Many of them were obtained from old scans and private archives.

The database contains not only the texts of the tasks, but also detailed solutions that have been peer-reviewed.

Why it is important

MathNet is useful in several directions at once.

For schoolchildren and teachers it is:

• a single archive of complex problems
• access to high-quality solutions
• opportunity to study materials from different countries

For AI researchers it is:

• a new way to test how well models can really reason
• a more complex and varied test than many previous problem sets

The authors showed that even strong AI models don't do perfectly. For example, one of the best systems solved, on average, about 69 per cent of the tasks in the main test set.

Background

Previously, such bases usually relied mainly on materials from the US and China. The new archive is noticeably broader: it covers dozens of countries and different mathematical traditions.

The authors believe that this is important not only for humans, but also for AI, because different countries formulate and solve problems in different ways.

Source

The MathNet project was presented by researchers from MIT, KAUST and HUMAIN. The material will be presented at the ICLR 2026 conference.