Google And OpenAI Win Gold Medals At Global Mathematics Competition

Global math competition, International Mathematical Olympiad (IMO) for high school students. (photo; x @gerkobot)

JAKARTA - Alphabet (Google) and OpenAI announced that their artificial intelligence (AI) model won a gold medal in global math competition, International Mathematical Olympiad (IMO) for high school students. This marks the first time the AI system has exceeded the threshold for gold medals in the competition.

Both AI models from Google and OpenAI managed to solve five of the six problems using a generalized "scattering" model that processes mathematical concepts in natural language, different from previous approaches used by AI companies.

Google DeepMind is working with IMO to assess and certify their model, while OpenAI is not officially participating in the competition. OpenAI revealed that their model achieved an equal score of gold medals on this year's issue, based on the assessment of the three external IMO medalists.

This achievement suggests that AI could potentially be used by mathematicians to solve unresolved research problems in less than a year, said Junehyuk Jung, professor of mathematics at Brown University and guest researcher at Google DeepMind AI unit.

"When we can solve difficult reasoning problems with natural language, it will open up potential collaborations between AI and mathematicians," Jung said.

OpenAI's success was achieved with a new experimental model focused on the massive "computation of test times". This was done by giving a model longer time to "thinking" and using parallel computing power to carry out multiple lines of reasoning simultaneously, according to Noam Brown, an OpenAI researcher. Brown called the computational cost "very expensive" without elaborating on the amount.

OpenAI researchers see this as a sign that AI models have broad reasoning capabilities that can be extended to other areas beyond mathematics. Similar optimism was shared by Google researchers, who believe AI model capabilities could be applied to research issues in areas such as physics, said Jung, who won the IMO gold medal in 2003.

Of the 630 students who participated in the 66th IMO on the Sunshine Coast, Queensland, Australia, around 11% or 67 participants won the gold medal score.

Last year, Google's DeepMind AI unit achieved a silver medal score using a dedicated AI system for mathematics. This year, Google used a general model called Gemini Deep Think, which was previously introduced at its annual developer conference in May.

Different from the previous AI approach that relies on formal language and long computing, this year's Google approach fully uses natural language and resolves the problem within the official 4.5-hour deadline, according to the company's blog post.

OpenAI, which has its own reasoning model, is also building an experimental version of the competition, according to researcher Alexander Wei's post on platform X. He noted that the company is not planning to release this model with mathematical capabilities in the next few months.

This year will be the first time the IMO competition has officially collaborated with several AI developers, who over the years have used well-known mathematical competitions like IMO to test model capabilities. The IMO jury certified the results of the companies working together, including Google, and asked them to publish the results on July 28.

"We respect the initial request of the IMO Council so that all AI laboratories share the results only after the official results are verified by independent experts and students get a decent award," said Google DeepMind CEO Demis Hassabis, on X on Monday, July 21.

OpenAI, which published its results on Saturday 19 July and first claimed the status of a gold medal, said in an interview that it had permission from IMO board members to do so after Saturday's closing ceremony. Competition on Monday allowed companies working together to publish results, "said Gregor Dolinar, president of the IMO board.