Artificial intelligence models from Google and OpenAI have reached gold-medal level scores at the International Mathematical Olympiad (IMO), marking a new development in the ability of AI systems to solve complex problems using natural language.
The IMO is an annual competition for students under 20, and this year’s event concluded Saturday in Queensland, Australia. Of the 630 participants, 67 students achieved gold-medal scores, with five earning perfect marks. One of the top scorers was Patanasang Pinijpichitkul of Thailand, who earned a 95.87% score.
Google DeepMind worked with the IMO to have its AI models officially graded. The models solved five of the six problems within the contest’s 4.5-hour time limit, using natural language instead of formal languages and complex calculations used in past AI models.
OpenAI also reported gold-level performance from its own model, although it did not formally enter the competition. The company said the scores were verified by three former IMO medalists.
OpenAI achieved the result by using an experimental model that increased the amount of computing power during test time. This method allowed the system to explore many lines of reasoning at once and think over longer periods.
Noam Brown, a researcher at OpenAI, said the process required a high amount of computing resources but did not disclose the cost.
Both companies said the achievement points to new uses for AI in mathematics and other research fields. Google researchers noted that these reasoning models could be applied in areas such as physics.
Brown University’s Junehyuk Jung, a visiting researcher at DeepMind, said the ability of AI to solve advanced reasoning tasks using natural language may lead to closer collaboration between mathematicians and AI systems.
