AI is overcoming the weakness of solving advanced mathematics

Software engineer Neel Somani tested mathematical abilities on OpenAI’s GPT 5.2 model and discovered something interesting: AI can solve math at a high level instead of just solving textbook exercises. He researched Erdős problems – a collection of more than 1,000 hypotheses by Hungarian mathematician Paul Erdős, currently stored online on the website of the same name with levels of complexity ranging from elementary problems to problems at the highest boundary of modern mathematics.

Erdős is considered one of the outstanding mathematicians of the twentieth century, with hundreds of mathematical works in the fields of combinatorics, graph theory, numbers, approximation, sets and probability. To recognize his contributions, since 2015, Hungary has established a competition called “Paul Erdős International Mathematics Challenge”.

Somani brought a highly complex Erdős problem into ChatGPT and let AI handle it. In 15 minutes, the complete solution was given. Through the independent evaluation tool Harmonic, every solution step and result are accurate.

According to Somani, the chain of reasoning on ChatGPT is even more impressive when it lists mathematical axioms such as Legendre’s formula, Bertrand’s axiom and the Star of David Theorem. Compared to the solution given by mathematician Noam Elkies at Harvard University in 2013, ChatGPT has a more complete solution.

For those skeptical about machine intelligence, this is a surprising result. Previously, AI was considered to have difficulty solving math problems, even with simple problems. According to Wu Yiquan, a computer science researcher at Zhejiang University in Hangzhou, it is very common for AI to not be good at math. The LLMs behind do not possess mathematical abilities, as they predict answers based on training data.

“Some LLMs that perform well on the math test may be due to ‘data contamination,’ meaning the algorithm remembered the answer because similar questions were already in its training data,” Yiquan explained to SCMP last year. “The AI ​​world is encrypted, so numbers, words, punctuation, and spaces are all treated the same. Any change in the prompt can significantly affect the results.”

In 2023, John A. Paulson School of Engineering and Applied Sciences mathematics professor Michael Brenner taught a graduate class, applying AI to nonlinear partial differential equations in the course. “It’s okay, but not great,” Brenner told Harvard Gazette then. “AI only solved 30% to 50% of the problems in the first three weeks of the course.”

“In general, AI is not good at math,” Alex Kotran, co-founder and CEO of AI Education Project, told The74 end of 2024. “Language models only predict the next word, you get inconsistent results. It’s not mature enough to be trusted to scale when doing math.”

However, after more than a year, AI models progress rapidly. Math solving ability improves with each version. For example, GPT 5.2 now infers better than previous versions. With problems on the Erdős website, from December 25, 2025 up to now, there are 15 problems that have moved from “unsolved” to “solved” status, 11 of which recorded the participation of AI models.

“Many Erdős problems are inherently difficult for mathematicians, but are handled by AI,” mathematician Terence Tao wrote on Mastodon. He hypothesizes that the scalable nature of AI systems makes them “more suitable to be applied systematically”. This helps simplify complex problems.

Last year, several AIs were recognized as capable of winning Olympic math gold medals. DeepMind’s Deep Think and OpenAI’s ChatGPT solved 5/6 problems in the International Mathematical Olympiad (IMO) 2025. In 2024, Google’s AlphaGeometry 2 correctly solved 42/50 IMO geometry problems over the past 25 years, while AlphaProof achieved a perfect score of 4 out of 6 IMO 2024 problems, winning 28/42 points, equal to the silver medalist.

However, AI’s increasingly good math skills also raise new questions about the ability of LLMs to push the boundaries of human knowledge. Latrenda Knighten, President of the National Council of Teachers of Mathematics (NCTM) of the US, assessed that AI has assisted people in solving math problems, but they still need to rely on their own judgment to solve mathematical problems.

“We know that children learn math by solving problems, using reasoning skills, thinking critically, having opportunities to collaborate with each other and discuss what they are doing,” Knighten told Education Week. “If we depend on AI, those skills will not be promoted.”

In the past, many experts also advised users not to completely trust AI models. In fact, many times they were found to answer incorrectly or experience “hallucinations”, which means making up answers to sound most convincing.