Google DeepMind Announces New AI Coding Agent for Maths and Science

Google DeepMind has announced AlphaEvolve, a new AI coding agent designed to evolve and optimise algorithms across computing, mathematics, and science.

Powered by Gemini Flash and Gemini Pro, AlphaEvolve combines large language models (LLMs) with automated evaluators to generate, verify, and iteratively improve high-performing code solutions.

“AlphaEvolve pairs the creative problem-solving capabilities of our Gemini models with automated evaluators that verify answers, and uses an evolutionary framework to improve upon the most promising ideas.,” Google said in a blog post.

The system uses an evolutionary algorithm framework that integrates prompt sampling, language model output, and program evaluation.

AlphaEvolve has already shown impact across Google’s infrastructure. It was used to optimise Borg, Google’s data centre orchestrator, recovering 0.7% of global compute resources — a solution now in production for over a year.

"Over the past year, we’ve deployed algorithms discovered by AlphaEvolve across Google’s computing ecosystem, including our data centers, hardware and software. The impact of each of these improvements is multiplied across our AI and computing infrastructure to build a more powerful and sustainable digital ecosystem for all our users," Google added.

It also contributed to hardware design, proposing a Verilog-level change to a TPU’s arithmetic circuit, which passed verification and will be integrated into a future release.

In AI training, AlphaEvolve optimised matrix multiplication in the Gemini architecture, improving kernel speed by 23% and reducing training time by 1%. It also enhanced FlashAttention kernel performance by 32.5%, outperforming manual compiler-level optimizations.

Beyond infrastructure, AlphaEvolve has tackled complex mathematical challenges. It discovered a new way to multiply 4×4 complex-valued matrices using just 48 scalar multiplications — an improvement over the 1969 Strassen algorithm.

Applied to more than 50 open math problems, it rediscovered known solutions in 75% of cases and improved upon 20%.

DeepMind is planning an early access programme for select academic users and is exploring broader availability.