Google has introduced AlphaEvolve, an AI agent that designs and optimizes algorithms by combining large language models with automated evaluation. Built on Google’s Gemini models, Gemini Flash for exploring diverse ideas and Gemini Pro for detailed problem-solving AlphaEvolve generates computer programs that implement algorithmic solutions. These programs are tested and scored using automated metrics, with an evolutionary framework refining the best solutions over time.

AlphaEvolve has already made an impact across Google’s operations. In data centers, it developed a scheduling heuristic for the Borg system, recovering 0.7% of global compute resources on average. This efficiency gain allows more tasks to run on existing infrastructure, and the human-readable code simplifies deployment and maintenance. In hardware design, AlphaEvolve proposed a Verilog rewrite for a matrix multiplication circuit in an upcoming Tensor Processing Unit, maintaining functional correctness while optimizing performance. For AI training, it improved matrix multiplication operations, reducing Gemini’s training time by 1% and speeding up the FlashAttention kernel in Transformer models by up to 32.5%.

In mathematics, AlphaEvolve has tackled complex problems. It designed a new algorithm for multiplying 4x4 complex-valued matrices using 48 scalar multiplications, surpassing Strassen’s 1969 algorithm. It also advanced the kissing number problem, finding a configuration of 593 outer spheres and established a new lower bound in 11 dimensions.. Across 50 open problems in areas like geometry and number theory, AlphaEvolve matched state-of-the-art solutions in 75% of cases and improved on existing solutions in 20%.

Google is working on a user-friendly interface for AlphaEvolve and plans an Early Access Program for academic researchers, with a registration form available . The agent’s flexible framework could extend to fields like material science or drug discovery, where algorithmic solutions can be verified. As language models improve, AlphaEvolve’s capabilities are expected to grow, offering new ways to address complex challenges in computing and beyond.