Google launches AlphaEvolve, an evolving coding agent for designing advanced algorithms

Google has released AlphaEvolve, an evolutionary coding agent powered by large language models for general-purpose algorithm discovery and optimization. By pairing the problem-solving capabilities of the Gemini models with automated evaluators, AlphaEvolve systematically verifies solutions and refines the most promising approaches using an evolutionary framework.

Building on this architecture, AlphaEvolve has already been deployed within Google's infrastructure. It is active in data centers, chip design pipelines, and the training of AI models, including the models that drive AlphaEvolve itself. The agent uses an ensemble of large language models, with Gemini Flash exploring a wide range of concepts, while Gemini Pro contributes to deeper, more specialized insights. Together, these models collaborate to generate computer programs that deliver algorithmic innovations.

All programs produced by AlphaEvolve are subjected to automated evaluation metrics, which objectively score their accuracy and quality. This process allows the agent to automate algorithmic optimization in domains where progress can be tested using quantifiable measures. Google expects AlphaEvolve to advance further as large language models improve their coding abilities.