AlphaEvolve is an evolutionary coding agent developed by Google DeepMind and announced on 14 May 2025. It pairs the creative problem-solving capabilities of Gemini large language models with automated evaluators that verify and score proposed programs, and uses an evolutionary framework to iteratively improve the most promising ideas. This allows the agent to go beyond single-function discovery and evolve entire codebases and complex algorithms.
AlphaEvolve leverages an ensemble of Gemini models: Gemini Flash maximizes the breadth of explored ideas, while Gemini Pro provides depth and critical, insightful suggestions. Proposed programs are run, verified and scored by automated metrics that yield an objective, quantifiable assessment of accuracy and quality, which makes the agent particularly effective in domains where progress can be systematically measured — primarily mathematics and computer science.
The system has been deployed inside Google's computing infrastructure. A heuristic discovered by AlphaEvolve for the Borg scheduler has been in production for over a year, recovering on average 0.7% of Google's worldwide compute resources. AlphaEvolve also proposed a Verilog rewrite of an arithmetic circuit for matrix multiplication that was integrated into an upcoming TPU, sped up a key kernel in Gemini's architecture by 23% (a 1% reduction in Gemini training time), and achieved up to a 32.5% speedup of the FlashAttention kernel implementation in Transformer-based models.
In mathematics, AlphaEvolve found an algorithm to multiply 4×4 complex-valued matrices using 48 scalar multiplications, improving upon Strassen's 1969 algorithm. The system was applied to over 50 open problems in mathematical analysis, geometry, combinatorics and number theory: in roughly 75% of cases it rediscovered state-of-the-art solutions, and in 20% of cases it improved the previously best-known solutions — including a new lower bound for the kissing number problem in 11 dimensions (593 outer spheres).
AlphaEvolve is not a publicly available product. The team is preparing an Early Access Program for selected academic users; broader availability is still being explored. Technical documentation is published as a white paper.
