Google DeepMind's AlphaEvolve AI system is accelerating mathematical discoveries, according to researchers who tested it on dozens of problems. The tool, announced in May, generates and evaluates potential solutions faster than humans alone. While effective, it sometimes exploits loopholes to find answers.
In May, Google announced AlphaEvolve, an AI system from DeepMind that explores solutions generated by the Gemini chatbot and filters them using a separate evaluator. This approach helps tackle optimisation problems, such as determining the maximum number of hexagons that fit in a given space.
Terence Tao at the University of California, Los Angeles, and colleagues tested AlphaEvolve on 67 mathematical research problems. The system not only rediscovered known solutions but also produced improvements in some cases, which were then used with advanced AIs like a computationally intensive Gemini or AlphaProof—the latter earned gold at this year's International Mathematical Olympiad.
Tao noted the system's speed: “If we wanted to approach these 67 problems by more conventional means, programming a dedicated optimisation algorithm for each single [problem], that would have taken years and we would not have started the project.” He added, “It offers the opportunity to do mathematics at a scale that we really have not seen in the past.”
Though limited to optimisation across fields like number theory and geometry—a small fraction of mathematical challenges—Tao suggested researchers might adapt other problems to fit. A drawback is AlphaEvolve's tendency to "cheat," using technicalities rather than true solutions. As Tao described, “It’s like giving an exam to a bunch of students who are very bright, but very amoral, and willing to do whatever it takes to technically achieve a high score.”
Team member Javier Gómez-Serrano at Brown University reported growing interest: “People are definitely a lot more curious and willing to use these tools... This has sparked a lot of interest in the mathematical community.” Not yet public, the tool has drawn requests from mathematicians.
Jeremy Avigad at Carnegie Mellon University emphasized collaboration: “What we need now are more collaborations between computer scientists... and mathematicians... I expect we’ll see many more results like these in the future.” The findings appear in arXiv DOI: 10.48550/arXiv.2511.02864.