The Sundarban
August 21, 2025
6 min be taught
Can Writing Math Proofs Educate AI to Goal Fancy Folks?
OpenAI researchers indicate how their experimental mannequin, devoid of any exterior aids, powered through hours-long proofs to fabricate a gold-medal catch at the International Math Olympiad—and they boom about the project’s origins and describe how such work may possibly maybe well motivate lead to man made total intelligence
By Deni Ellis Béchard edited by Dean Visser
peshkov/Getty Pictures
A pair of months before the 2025 International Mathematical Olympiad (IMO) in July, a 3-person team at OpenAI made a long wager that they may possibly maybe well use the competitors’s brutally tough complications to practice an man made intelligence mannequin to contemplate on its luxuriate in for hours so that it turned into as soon as in a position to writing math proofs. Their goal wasn’t simply to create an AI that may possibly maybe well abolish advanced math nonetheless one that may possibly maybe well evaluate ambiguity and nuance—abilities AIs will need if they’re to at some point address many tough genuine-world duties. In point of fact, these are precisely the abilities required to create man made total intelligence, or AGI: human-degree understanding and reasoning.
The IMO, held this 300 and sixty five days on Australia’s Sunshine Wing, is the enviornment’s premier math competitors for top schoolers, bringing together top contenders from greater than 100 nations. All are given the the same six complications—three per day, every worth seven aspects—to solve over two days. Nevertheless these complications are nothing esteem what you almost surely undergo in mind from high college. Rather than a short numeric solution, every demands sustained reasoning and creativity within the create of a pages-long written proof. These logical, step-by-step arguments luxuriate in to span many fields of mathematics—exactly the create of complications that, till moral this 300 and sixty five days, AI programs failed at spectacularly.
The OpenAI team of researchers and engineers—Alex Wei, Sheryl Hsu and Noam Brown—inclined a total-cause reasoning mannequin: an AI designed to “contemplate” through tough complications by breaking them into steps, checking its luxuriate in work and adapting its potential because it goes. Though AI programs couldn’t officially compete as people, the notoriously tough test served as a demonstration of what they can abolish, and the AIs tackled this 300 and sixty five days’s questions within the the same test format and with the the same constraints as human people. Upon receiving the questions, the team’s experimental system worked for 2 4.5‑hour sessions (moral because the coed contestants did), without tools or the Web—it had fully no exterior assistance from tools comparable to search engines or tool designed for math. The proofs it produced were graded by three extinct IMO medalists and posted on-line. The AI completed five of the six complications because it will be, receiving 35 out of 42 aspects—the minimum required for an IMO gold medal. (Google’s DeepMind AI system also finished that catch this 300 and sixty five days.) Out of 630 competitors, most productive 26 students, or 4 p.c, outperformed the AI; five students finished excellent 42s. Given that a 300 and sixty five days within the past language-primarily based mostly mostly AI programs esteem OpenAI’s struggled to abolish main math, the outcomes were a dramatic leap in performance.
On supporting science journalism
Whereas you may possibly maybe moreover be taking part in this text, take into consideration supporting our award-winning journalism by subscribing. By shopping a subscription you may possibly maybe moreover be serving to to catch definite the style forward for impactful stories about the discoveries and tips shaping our world today.
Within the following conversation, Scientific American spoke with two people of the OpenAI team, Alex Wei and Sheryl Hsu, to boom about how they performed their work, why the mannequin’s lack of response to the sixth demand turned into as soon as if truth be told a main step toward addressing AI’s “hallucination” articulate and how increasing a system in a position to writing advanced proofs may possibly maybe well motivate lead to man made total intelligence.
[[An edited transcript of the interview follows.]
What led you to all at as soon as initiate making ready an AI mannequin for the IMO moral about a months before the competitors? What turned into as soon as the spark?
WEI: I had been fascinated by math proofs for relatively a while. I’m on a team at OpenAI called MathGen. We had moral considered the outcomes progress loads. We felt esteem we had a shot to catch a mannequin that may possibly maybe well abolish in actuality wisely at the IMO, and we wanted to catch a angry meander to catch there.
HSU: I inclined to abolish math competitions. [Wei] inclined to abolish math competitions—he turned into as soon as loads better than me. The IMO is undoubtedly wisely known all the diagram during the [AI research] community, together with amongst researchers at OpenAI. So it turned into as soon as in actuality inviting to push namely for that.
Are you able to boom about your determination to work with a total‑cause AI system rather than a system that turned into as soon as namely designed to solution math complications?
WEI: The philosophy is that we prefer to create total‑cause AI and fabricate methods that don’t moral work for math. Math is a really correct proving ground for AI because it’s reasonably goal: must you may possibly maybe moreover luxuriate in a proof, it’s more uncomplicated to catch consensus on whether or now not it’s factual. That’s more difficult for, disclose, poetry—you’ll luxuriate in more disagreement amongst readers. And IMO complications are very laborious, so we wanted to address laborious complications with total‑cause methods within the hope that they’ll also note to domains previous math.
HSU: I’d also disclose the goal at OpenAI is to create AGI—it’s now not essentially to write papers or obtain competitions. It turned into as soon as valuable that all the issues we did for this project even be vital for the larger goal of constructing AGI and better units that users can if truth be told use.
In what methods may possibly maybe well a reasoning mannequin winning a gold within the IMO motivate lead to AGI?
WEI: One standpoint is to contemplate in terms of how long duties take. A 300 and sixty five days within the past, ChatGPT may possibly maybe well most productive abolish very popular math complications. Two years within the past—and even a 300 and sixty five days and a half of within the past—we were customarily fascinated by grade‑college math complications you’d get on fifth‑grade homework.
