Google has introduced AlphaEvolve, an agent based on large language models that can develop and optimize algorithms for a variety of mathematical and applied problems.
The solution combines the generative capabilities of Gemini models with automated evaluation systems and uses an evolutionary approach to code optimization. Google claims that AlphaEvolve has uncovered new mathematical solutions that scientists have been unable to find for decades.
According to Google, AlphaEvolve has already helped improve data center efficiency, chip design, and AI model training, including the ones that underpin AlphaEvolve itself. In mathematical research, AlphaEvolve has improved previous methods for multiplying 4×4 complex matrices, beating Strassen's 1969 algorithm, and has also set a new lower bound on the kissing number problem in 11 dimensions.
Google plans to launch an early access program for AlphaEvolve for scientists. The company believes that the agent could be useful not only in mathematics and computer science, but also in materials science, medicine, ecology and other fields.