AI systems score perfect marks at top maths contest for first time

Artificial intelligence has reached a striking new benchmark in advanced mathematics, with two Chinese tech firms saying their models achieved perfect scores at this year’s International Mathematical Olympiad, long regarded as the world’s toughest school-level maths contest. Huawei and Xiaohongshu said their systems each scored 100% on the problems set for human contestants, according to statements reported this week. If confirmed against the contest’s official judging process, it would mark the first time large language models have matched the top result at the IMO. The result matters because the Olympiad is not a standard exam of memorised methods. It is designed to test deep reasoning, creativity and the ability to solve difficult problems under strict time limits. Until now, AI tools have been improving steadily but have still lagged behind the very best student mathematicians on the competition’s hardest questions. That gap narrowed in 2025, when models from companies including Google and OpenAI reached gold-level performance for the first time. Even so, they still did not match the handful of human contestants who earned perfect scores. This year’s announcement suggests the frontier has moved again. The contest itself drew 666 participants from different countries in Shanghai, and just seven achieved full marks, according to the official scoreboard. Contestants must be under 20, which makes the IMO a showcase not only of mathematical talent but also of the next generation of problem-solvers. Huawei said its AI system, Celia, demonstrated “comprehensive problem-solving capabilities” across several branches of mathematics. Xiaohongshu, known internationally as RedNote, said its model, dots-note-3.0, entered the IMO for the first time this year. The firms said the systems were tested on the Olympiad problems only after the human students had taken the exam, with responses due within a set time window. That timing is important. In AI benchmarking, it can make a major difference whether a model is trained on known contest questions or evaluated under conditions that closely mirror the real event. The companies’ claims point to a more controlled attempt to measure performance, but independent verification will matter before the result is treated as a settled milestone. The development also feeds into a broader debate about what AI progress actually means. A perfect score in a mathematics contest does not mean a model understands maths the way a human does, or that it can reliably apply that reasoning outside a narrow test environment. But it does show that current systems are becoming much stronger at structured problem-solving, a capability with obvious implications for science, engineering and research workflows. What remains unclear is whether other AI labs around the world will report similar results in the coming days. The feed suggests more announcements may follow, which could help show whether Huawei and Xiaohongshu have moved ahead of rivals or are simply the first to publicise comparable scores. For now, the headline is simple: AI has caught up with the very top of the world’s most prestigious mathematics contest. Whether that marks a practical breakthrough, a benchmarking milestone or both will depend on how the broader AI field responds.

Source: Dawn Tech - https://www.dawn.com/news/2017757/ai-catches-up-with-humans-to-score-100pc-at-top-maths-contest