AlphaEvolve

What it is

AlphaEvolve is an evolutionary coding agent from Google DeepMind, announced on 14 May 2025, built to discover and optimise algorithms.^[1] Instead of asking a model for a single program and taking what comes back, it generates many candidate programs, scores them with automated evaluators, and uses those scores to decide which candidates to keep and change next. The name refers to that loop: the code is evolved rather than written in one pass.

How it works

The agent pairs Gemini models with automated evaluators.^[1] The two models play different roles: Gemini Flash is used to broaden the set of ideas explored, and Gemini Pro is used for deeper suggestions.^[1] The evaluator is what anchors the search — it has to score a candidate program automatically, whether by running it, checking it against a specification, or measuring how fast it runs. A problem without such a measurable signal cannot be handed to the system at all.

Inside Google's own infrastructure

The most concrete deployment is a data-centre scheduling heuristic that AlphaEvolve found for Borg, Google's cluster management system. According to DeepMind, that heuristic had been in production for over a year and was continuously recovering on average 0.7% of Google's worldwide compute resources.^[1] At Google's scale a fraction of that kind is large in absolute terms, and the figure is notable for a second reason: this is not a one-off benchmark score but code running in production.

The system also worked on the kernels that Gemini itself depends on. It sped up a matrix-multiplication kernel in Gemini's architecture by 23%, which reduced Gemini's training time by 1%, and it achieved up to a 32.5% speedup for FlashAttention kernels.^[1]

On the hardware side, a Verilog rewrite it proposed was integrated into an upcoming TPU.^[1]

DeepMind framed the change in working style this way: kernel optimisation that once took weeks of expert effort moved to days of automated experiments with AlphaEvolve.^[1]

In mathematics

AlphaEvolve was also applied to open mathematical problems. It found a way to multiply 4×4 complex-valued matrices using 48 scalar multiplications, improving on Strassen's 1969 algorithm.^[1] Reaching that result took 15 mutations in the evolutionary process.^[1] The comparison DeepMind draws is with its own earlier system, AlphaTensor, which had found improvements only for 4×4 matrices with binary arithmetic.^[1]

Across more than 50 open mathematical problems, the system rediscovered state-of-the-art solutions in about 75% of cases and improved the best known solutions in about 20%.^[1] One of those improvements was a configuration of 593 outer spheres that set a new lower bound for the kissing number in 11 dimensions.^[1] DeepMind noted that the kissing-number problem had fascinated mathematicians for over 300 years.^[1]

Credits and availability

Credited developers on the project include Matej Balog, Alexander Novikov and Pushmeet Kohli.^[1] DeepMind said it was planning an Early Access Program for selected academic users and was exploring making AlphaEvolve more broadly available, with no date given.^[1]

Where it sits

In DeepMind's line of work on algorithm discovery, AlphaEvolve follows on from systems such as AlphaTensor, but it operates on programs in general rather than on one narrow class of mathematical objects. The same mechanism — a model proposing candidates, an evaluator scoring them, and a selection loop — is what lets one system be pointed at scheduling heuristics, compute kernels, hardware descriptions and pure mathematics in turn. Its limits follow from the same design: there must be an objective that can be measured automatically, and how good the results can get depends on how well that evaluator captures what the problem actually needs.