On May 14, 2025, Google DeepMind introduced AlphaEvolve, an evolutionary coding agent that pairs Gemini models with automated evaluators to discover and optimize algorithms. DeepMind reported that algorithms found by AlphaEvolve were already running in production across Google, recovering on average 0.7 percent of the company's worldwide compute, and that the system had improved results on a range of open mathematical problems. The research matters to software engineers because it shows how far "generate, test, keep the best" can go when a problem has a precise, automatic way to score solutions.

Key Facts#

  • Announcement: May 14, 2025, with a blog post and a technical white paper from Google DeepMind.
  • Approach: an ensemble of Gemini Flash, for breadth of ideas, and Gemini Pro, for depth, proposes code changes. Automated evaluators run and score each program, and an evolutionary database decides which programs seed future prompts.
  • Data centers: a scheduling heuristic for Google's Borg cluster manager, in production for over a year, recovers on average 0.7 percent of Google's worldwide compute resources.
  • AI training: a 23 percent faster matrix-multiplication kernel in Gemini's architecture cut Gemini's overall training time by 1 percent. AlphaEvolve also reported up to a 32.5 percent speedup for a FlashAttention kernel implementation.
  • Hardware: a Verilog rewrite of an arithmetic circuit for matrix multiplication was integrated into an upcoming Tensor Processing Unit (TPU).
  • Mathematics: a procedure that multiplies 4x4 complex-valued matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting.
  • Open problems: across more than 50 problems in analysis, geometry, combinatorics and number theory, AlphaEvolve matched the best known solutions in about 75 percent of cases and improved on them in about 20 percent.

What Happened#

AlphaEvolve builds on DeepMind's 2023 FunSearch work, which used language models to write single functions that led to new mathematical results. The difference is scope. AlphaEvolve evolves entire codebases and more complex algorithms, not just one function. A prompt sampler assembles context from previous programs and their scores, the Gemini models propose new versions as code changes, evaluators verify and score them, and a program database applies an evolutionary algorithm to decide which candidates survive.

The practical results came first. DeepMind said the Borg heuristic had been in production for more than a year, and stressed that the output is human-readable code, which makes it easier to interpret, debug and deploy than an opaque learned policy. In AI training, AlphaEvolve found smarter ways to split a large matrix multiplication into subproblems. DeepMind said this also cut kernel optimization work from weeks of expert effort to days of automated experiments. It even optimized low-level GPU instructions, an area that compilers usually handle and engineers rarely touch.

In mathematics, AlphaEvolve designed components of a gradient-based optimization procedure that found several new matrix multiplication algorithms. The 4x4 complex-valued result improves on Strassen's algorithm, which had been the best known in that setting since 1969. DeepMind noted that its earlier AlphaTensor system had only found 4x4 improvements for binary arithmetic. AlphaEvolve also set a new lower bound for the kissing number problem in 11 dimensions, with a configuration of 593 spheres touching a central sphere. DeepMind's acknowledgments thank mathematicians including Terence Tao for contributions and advice.

Background#

Using language models to generate code is common. Using them to discover algorithms that beat long-standing human results is not. The key constraint is verification: a language model's suggestions are only useful if something can check them. DeepMind's approach leans on domains where correctness and quality can be measured automatically, such as runtime, resource usage, circuit correctness or a mathematical bound. It then uses evolution to explore many candidates cheaply.

AlphaEvolve also sits in a line of DeepMind "Alpha" systems for algorithm discovery, from AlphaTensor for matrix multiplication to AlphaDev for sorting routines. What is new is generality. Because the search operates on ordinary code with pluggable evaluators, the same system can tune a data-center scheduler, a TPU circuit and a mathematical construction.

Why It Matters for Developers#

The transferable idea is the evaluator. AlphaEvolve works because each candidate gets an objective score. Most engineering teams already have the ingredients for a scaled-down version: unit tests to reject incorrect code, and benchmarks to measure speed and allocations. In .NET, a coding agent that proposes changes to a hot path, a test suite that guards correctness, and BenchmarkDotNet as the fitness function form a small evolutionary loop. The better your benchmarks represent production, the better the search will do.

Human-readable output is a feature, not a limitation. DeepMind emphasized that the Borg heuristic was simple code that operators could understand and deploy safely. The same principle applies to AI-assisted optimization in your codebase: prefer changes you can review, explain and roll back over opaque configuration that no one understands. Our guide to high-performance .NET covers where such optimizations usually pay off.

Know the limits. Evolutionary search is only as good as the metric, and it will happily exploit a benchmark that measures the wrong thing, such as a micro-benchmark that ignores cache effects or production data shapes. Treat AI-proposed optimizations like any other performance change: verify them under realistic load, and keep humans in review. For agent-assisted workflows more generally, see our guide to AI-assisted .NET development.

What's Next#

At launch, DeepMind announced a user interface built with Google's People + AI Research team and an Early Access Program for selected academic users. The community moved quickly as well: OpenEvolve, an open-source implementation of the AlphaEvolve approach, appeared on GitHub the next day and has since gathered thousands of stars.

A year later, in May 2026, DeepMind reported that AlphaEvolve had become a core part of Google's infrastructure and was available to enterprises through Google Cloud. Its examples included a 20 percent reduction in write amplification in Google Spanner through better compaction heuristics, regular use in designing next-generation TPUs, and customer results such as Klarna doubling the training speed of one of its largest transformer models and FM Logistic improving routing efficiency by 10.4 percent. Those are vendor-reported outcomes, but they show the approach moving from research demo to production tool. The open question is how far it extends into domains where evaluation is slow or subjective. DeepMind describes the approach as suited to any problem whose solution can be written as an algorithm and verified automatically. For evaluation design in AI systems, see our AI evaluation guide.

Sources#