AlphaEvolve

• 1. Introduction

  ›

  AlphaEvolve is a coding agent that writes programs and then lets them compete against one another. Google DeepMind unveiled it in May 2025. It is built on large language models such as Gemini, and its job is to design advanced algorithms. The name points to evolution, and the system really does borrow that idea. It starts from something that already works and tries to breed something better. That raises a few questions. What does the system need before it can begin? How does it differ from earlier DeepMind projects? And what has it actually found?

  124

• 2. The Evolutionary Loop

  ›

  Two ingredients start every AlphaEvolve run. The first is an initial algorithm, a working but unremarkable starting point. The second is an evaluation function, which supplies the metrics the system is trying to improve. At each step, AlphaEvolve asks the language model to produce variants of the algorithms it already has. Then it keeps the most effective ones. Repeating that cycle is the evolutionary part. The language model plays the role of mutation, and the metrics play the role of natural selection. The system aims to discover and refine algorithms on its own, with the model supplying the ideas and the scoring supplying the judgment.

  124

• 3. Beyond the Specialists

  ›

  AlphaFold and AlphaTensor are the predecessors that AlphaEvolve is most often compared with. Each was built for one domain. AlphaEvolve was designed as a general-purpose system instead. It can work across a wide array of scientific and engineering tasks, because it operates by modifying code. It can also optimize for several objectives at once. A second design choice matters just as much. AlphaEvolve evaluates code programmatically, so a program is run and measured rather than simply believed. That reduces the need for human input. It also reduces the risk of hallucinations, a familiar weakness of standard language model output.

  123

• 4. Results in Math and Machines

  ›

  Google tested AlphaEvolve on a selection of 50 open mathematical problems. According to Google, the system rediscovered state-of-the-art solutions 75% of the time. In 20% of cases it found improved solutions. One example is the kissing number problem, where it advanced the known results. The same method was turned on Google's own infrastructure. Improved data center scheduling heuristics allowed the recovery of 0.7% of stranded resources. AlphaEvolve was also used to optimize TPU circuit design. And it worked on a matrix multiplication kernel used in training Gemini, the very model family that powers AlphaEvolve.

  124

• 5. Open Source Descendants

  ›

  After the publication of AlphaEvolve, researchers outside Google began building open source versions of it. OpenEvolve is one such implementation. It runs distributed evolutionary algorithms and supports multiple programming languages. It can also connect to a variety of language model providers. Its notable feature is the automated discovery of high-performance GPU kernels. These kernels outperform baselines engineered by experts. The existence of such projects suggests that the core recipe is simple enough for others to rebuild. That recipe is a starting algorithm, a scoring function, and a language model that proposes changes.

  56