Advancing physical reasoning in open models.

The competition

Teams train and fine-tune open-weight models, then submit them through Kaggle to be evaluated on a set of 100 original problems. Every team works within the same compute budget, so winning solutions can be reproduced by any group.

The competition is open to anyone: companies, university labs, independent researchers and students.

The competition launches on Kaggle in April 2027. Exact dates and full rules will be announced soon.

Why a new olympiad

Since 1967, the International Physics Olympiad (IPhO) has been the most prestigious physics competition for high school students. Each country sends at most five students, chosen through national selection, and in 2025 they came from 85 nations. Over two days they face a five-hour experimental exam and a five-hour theory exam of three problems, competing for gold medals that only about one in twelve of them will win.

In 2025, AI systems from Google DeepMind and OpenAI reached gold-medal level at the International Mathematical Olympiad, and P1 became the first open-source model to reach gold-medal performance on IPhO 2025. IPhO is the best-known physics olympiad, but not the hardest. Its problems tend to start from well-known physics and guide the solver through it. At IPhO 2024, the organizers required the questions to connect to current global issues or research done in Iran. One problem was built on Nobel Prize-winning work on laser cooling, and its designer described breaking it down step by step to help students through it. That makes for a good exam, but the guidance that helps a student helps a model just as much, and the underlying physics is all over the literature models are trained on.

Harder contests offer no such guidance. The European Physics Olympiad (EuPhO) gives short problem statements and leaves the entire approach to the student. Physics Cup releases one problem a month, and in 2018 one went five weeks without a single correct solution. At the Online Physics Olympiad (OPhO), which our team runs, nearly 700 teams spent three days on this year’s contest, and no team fully solved Problem 22, even though we disqualified 166 teams for using AI.

Some of these problems still stand, but frontier models are solving more of them every year, and open-weight models lag behind. The deeper limit is the syllabus. Every one of these contests is capped at high school physics. EuPhO’s statutes require that no skills beyond the IPhO syllabus be needed, and Physics Cup works within the same bounds. A test with a fixed ceiling will eventually be saturated.

AIPhO removes the ceiling. Our problems keep what makes the hardest olympiads valuable, which is short statements, no guidance, and reasoning that cannot be pattern-matched, but they are not limited to high school physics. They range from the hardest olympiad style to active research, and are written by IPhO and EuPhO gold medalists alongside physicists from MIT, Oxford, UIUC, the University of Toronto and other universities.

Who we are

AIPhO’s team includes the organizers of the Online Physics Olympiad, which has brought together more than 10,000 competitors from over 80 countries across six years. They’re joined by active researchers, PhD students, and international olympiad medalists, coaches and problem writers.

AIPhO is an independent project modeled on the AI Mathematical Olympiad, whose team advises us.