DeepMind's AlphaEvolve pairs Gemini with automated evaluators to discover algorithms

Google DeepMind described AlphaEvolve, an “evolutionary coding agent” that uses its Gemini models to propose whole programs, then tests, scores and iteratively refines them against automated evaluation functions rather than a single correctness check. Gemini Flash generated a wide spread of candidate variants; Gemini Pro contributed more considered proposals; both fed a loop that kept, mutated and recombined the best-scoring code over many generations, extending an approach the company had used narrowly in AlphaTensor and AlphaDev to open-ended, general-purpose algorithm design.

DeepMind reported that AlphaEvolve had already been deployed inside Google’s own infrastructure: it said the system had discovered a scheduling heuristic that recovered about 0.7% of Google’s worldwide data-centre compute, produced a roughly 1% cut in the training time of Gemini itself, and found speedups in the matrix-multiplication kernels used to train models and in tensor-processing-unit circuit design. On pure mathematics, DeepMind said AlphaEvolve rediscovered state-of-the-art solutions to a large majority of a set of more than 50 open problems it was tested against and improved on the best known solution for about a fifth of them, including a new construction for 4x4 complex matrix multiplication using fewer scalar multiplications than the standard Strassen-derived algorithm, and a new bound on the kissing-number problem in eleven dimensions.

The system was not released publicly at launch; DeepMind said it planned an early-access programme for selected academic researchers. As with the AlphaTensor and AlphaDev results before it, the efficiency and mathematical claims came from DeepMind’s own account rather than independent replication, and the tool’s real-world value rested on how often such automated-evaluator loops could be built for problems that mattered outside Google’s own systems.