MIT releases MathNet, the world's largest dataset of math olympiad problems

MIT has released MathNet, the largest open dataset of math olympiad problems available today, built for both AI researchers and students preparing for competitions. It contains more than 30,000 problems with detailed solutions from the national math olympiads of 47 countries. For the industry, it offers a tougher test of mathematical reasoning than the usual English-language benchmarks. For students, it's a single library of high-quality problems that used to be scattered across printed booklets, forums, and personal archives.

The project comes from researchers at MIT CSAIL, KAUST, and the company HUMAIN. According to the team, MathNet spans 17 languages, 143 competitions, and roughly four decades of olympiad mathematics. The authors had to gather 1,595 PDF volumes totaling more than 25,000 pages, ranging from modern digital documents to old scans that for years existed only in private collections. A significant share of the archive came from the personal collection of one co-author, who has been scanning olympiad booklets by hand since 2006. MIT says the final dataset is roughly five times larger than its closest counterpart. It is already publicly available and will be presented at ICLR 2026 in Brazil.

What sets MathNet apart is not just its scale but the quality of its sources. Many existing math datasets were scraped from forums such as Art of Problem Solving; MathNet draws only on official national problem collections. That matters because the solutions in these materials are usually written by experts, go through review, and often walk through several ways to solve the same problem.

The dataset is also far broader geographically: it covers six continents, includes both text and visual problems, and isn't confined to the English-language and Chinese traditions. For extra validation, the team assembled a group of more than 30 reviewers from different countries who jointly re-checked thousands of solutions. For researchers, this is a chance to train models on a more diverse mathematical culture rather than a narrow set of familiar problem phrasings.

As an AI benchmark, MathNet is uncomfortable even for strong models. On the core set of 6,400 problems, GPT-5 scored about 69.3 percent, meaning it failed nearly one in three olympiad-level problems. When problems include figures, model scores drop even further, pointing to a persistent weakness in visual reasoning. The team also tested how models handle less widely spoken languages: several open-source systems scored 0 percent on problems in Mongolian.

The researchers also added a retrieval benchmark that requires recognizing structural similarity between two problems. Even the best embedding models found the correct match on the first try only about 5 percent of the time. This matters not just for AI but for the olympiads themselves: problems that are essentially the same have already appeared on real exams, and tracking mathematical equivalents across languages, notations, and formats is extremely hard even for experts. Another test showed that retrieval-augmented generation does help, but only when the retrieved problem is genuinely close in structure: DeepSeek-V3.2-Speciale gained up to 12 percentage points, while irrelevant hints made results worse in about 22 percent of cases.

MathNet's practical value reaches beyond academic AI. For students and teachers, it's a rare case of high-quality olympiad material from dozens of countries collected in one place and converted to a single format. For model developers, it's a reminder that loud claims of "nearly solved" math are still premature: once problems become truly international, multimodal, and less standardized, the quality gap remains clearly visible. That's why MathNet could become one of the most useful tests of genuine mathematical reasoning in models over the next few years, and at the same time one of the most valuable open libraries for math olympiad preparation.