The advanced version of the DeepMind artificial intelligence (AI) chatbot 'Gemini Deep Sync' received a score equivalent to a gold medal level of 35 points by taking the exam under the same conditions as human participants in the International Mathematical Olympiad (IMO), proving that AI has a human-level mathematical problem-solving capability in the IMO. The IMO grading committee assessed that this AI's descriptive solutions were clear, precise, and easy to understand.
DeepMind noted on the 21st through its blog that Gemini Deep Sync was graded under the same conditions and standards as human participants for this year's IMO problems, achieving 35 points out of a total of 42 points. This score is identical to the actual gold medal cutoff.
This year's IMO was held in Sunshine Coast, Australia, with 630 math prodigies participating from 110 countries. Among them, 67 won gold medals, 103 won silver medals, and 145 won bronze medals. The cutoffs for gold, silver, and bronze medals were 35 points, 28 points, and 19 points, respectively.
Gregor Dolinar, the Chairperson of the IMO, remarked about the descriptive answers written by AI that it was "impressive in many ways" and noted that "most solutions were easy to follow, and the development was clear and precise." The actual published answers showed that the AI logically presented systematic solutions by deriving auxiliary summaries and conclusions for each problem.
Gemini Deep Sync took the exam for two days, 4 hours and 30 minutes each day, just like the IMO participants, and fully solved 5 out of 6 problems. This marks a significant breakthrough compared to last year's other Google AI models, 'AlphaGeometric 2' and 'AlphaProof', which took more than two days to solve the IMO problems. Those models interpreted and solved problems using a special programming language rather than human language, and the final answers had to be translated back through human intervention.
The advanced version of Gemini Deep Sync used this time has not yet been made publicly available. Google plans to include it in its 'Google AI Ultra' subscription service, which costs $249.99 per month, after providing testing opportunities to trusted mathematicians.
Meanwhile, OpenAI also stated that its AI model recorded gold medal-level scores when applied to the IMO problems under the same conditions. However, this result is based on internal evaluations conducted by OpenAI.