Vietnamese PhD behind AI's miracle at the International Mathematical Olympiad

Google shocked the world on July 23 when it announced that its artificial intelligence (AI) model won a gold medal at the 2025 International Mathematical Olympiad (IMO).

According to WSJ , Dr. Luong Minh Thang and his colleagues are the "architects" of this achievement. Last year, this model only won a silver medal. With the expectation of winning a gold medal, Google decided to use the multi-purpose model Gemini Deep Think (a version that was introduced at the developer conference in May).

However, Dr. Minh Thang initially expected to win a bronze or silver medal again. Even before the competition took place, Mr. Thang’s team was still making adjustments. Initially, Dr. Thang only expected DeepMind’s model to solve all three problems on the first day.

AI "conquers" international Olympic gold medal

The simplicity, elegance, and readability of those solutions amazed mathematicians. The next day, when Dr. Thang and his colleagues discovered that the AI system had solved two more problems, they realized that it was entirely possible to win a gold medal.

As a result, DeepMind AI successfully solved five out of six problems. Notably, they all used empirical reasoning models, processing mathematical concepts in natural language, which is quite different from the complex approaches that AI companies have used in the past.

“This could probably be a new computer that will propel the next generation of mathematicians,” Dr. Minh Thang told WSJ .

AlphaGeometry team members (from left) include Yuhuai Wu, Trinh Hoang Trieu, Le Viet Quoc and Luong Minh Thang. Photo: Aaron Cohen.

According to Reuters , this achievement shows that in less than a year, AI can be used by mathematicians to solve unsolved research problems.

Mr. Junehyuk Jung, professor of Mathematics at Brown University (USA), researcher at AI DeepMind, commented: "I think the moment we can solve difficult reasoning problems with natural language will open up the potential for cooperation between AI and mathematicians."

Researchers at Google share this optimism, believing that the capabilities of AI models can be applied to research challenges in other fields such as physics.

20 years ago, Mr. Thang missed the IMO exam when he ranked 8th nationwide in the team, while only 6 candidates were selected to participate in the international arena.

Although his dream of conquering the peak of Mathematics was unfulfilled, he turned to artificial intelligence (AI) when he entered university. This turning point brought him to the IMO 2024 exam in a special way - not as a contestant, but with the AI software AlphaGeometry developed by the AlphaGeometry team, competing with mathematical talents from more than 100 countries in Bath, UK.

“In July 2024, for the first time, our team brought AI to participate in the International Mathematical Olympiad (IMO) and won a silver medal,” Dr. Thang shared.

Leap Forward

The Google DeepMind team participated in the IMO competition last year using an AI that included the AlphaProof and AlphaGeometry 2 models. In its first attempt, Google's AI was able to answer 4 out of 6 questions correctly, winning the silver medal.

By 2025, Google DeepMind is among the companies that have partnered with IMO to officially enter AI models into the competition, which will be scored and certified by the coordinators.

Hoping to win the gold medal, Google decided to use the Gemini Deep Think multi-purpose model (a version that was introduced at the developer conference in May). Photo: Google.

To mark this milestone, the search giant has prepared a completely new model. Instead of following a linear “thinking” path, Deep Think runs multiple reasoning processes in parallel, integrating and comparing results before coming up with a final answer.

According to Dr. Thang, this is a major paradigm shift. In 2024, an expert will have to translate natural language questions into “domain-specific language”. Then, this expert will have to explain the output.

With Deep Think, however, the AI system can operate entirely in natural language, end-to-end, and can do much more than just solve math problems.

Previously, improving LLM in mathematics would involve reinforcement learning with final answers. Explaining to Ars Technica , Dr. Minh Thang said that models trained this way can get the right answer, but they have “incomplete reasoning,” and part of IMO grading is based on the representation of the solution.

So to prepare Deep Think for IMO, Google used new reinforcement learning techniques, with higher quality “long answer” solutions to mathematical problems, giving the model a better foundation in handling every step along the way to the answer.

“With this type of training, you can really get strong, long-term reasoning,” says Dr. Thang.

Dr. Luong Minh Thang is discussing the solved IMO AlphaGeometry problem with Mr. Le Ba Khanh Trinh. Photo: Wendy Nguyen.

Over the past few years, AI companies like Google DeepMind have expressed particular interest in IMO due to its unique challenges.

Although intended for high school students, the questions in the competition still require critical thinking and an understanding of many branches of mathematics, including algebra, combinatorics, geometry and number theory.

Only the most advanced AI models have any hope of accurately answering these multi-layered problems. The DeepMind team pointed out some interesting aspects of Deep Think’s performance, which they said came from its advanced training.

For example, in the third problem, many contestants applied a graduate-level concept called Dirichlet's Theorem, using mathematics that was outside the intended scope of the contest.

However, Deep Think realized that it was possible to solve the problem using simpler mathematics. “Our model actually made a brilliant observation and used only elementary number theory to create an independent proof of the problem,” Professor Jung shared.

Source: https://znews.vn/tien-si-nguoi-viet-dung-sau-ky-tich-cua-ai-tai-olympic-toan-quoc-te-post1572494.html