Google DeepMind · 2026-05-07 · major
AlphaEvolve Year-in-Review — DeepMind's Gemini Coding Agent Lifts AC Power Flow Feasibility 14% to 88% and Cuts Spanner Write Amp 20%
DeepMind details where AlphaEvolve, its Gemini-powered algorithm-discovery agent, has shipped real wins: 30% better DNA sequencing in DeepConsensus, 10x lower quantum-circuit error on Willow, 4x faster ML force fields for Schrödinger.
DeepMind's AlphaEvolve has graduated from research demo to producing measurable improvements across genomics, grid optimization, quantum hardware, and Google's own infrastructure.
Key specs
| Deep consensus dna error reduction | 30% |
|---|---|
| Ac opf feasible solutions | 14% → 88% |
| Willow quantum error rate | 10x lower |
| Spanner write amplification | 20% lower |
| Schrödinger ml force fields speedup | 4x |
What is it?
AlphaEvolve is an autonomous coding agent that uses Gemini to evolve algorithms by repeatedly proposing, testing, and refining code against a scoring function. It was first shown in 2025; this post is a year-on impact summary with concrete deployment numbers.
How does it work?
An evolutionary loop: Gemini generates candidate code, an automated evaluator scores it on the target metric, and high-scoring variants are mutated and recombined. Domain teams supply the scoring harness; AlphaEvolve searches the algorithm space.
Why does it matter?
Most LLM coding agents target generic SWE tasks. AlphaEvolve is being pointed at narrow, high-value optimization problems where a small improvement compounds — kernel scheduling, packet routing, gate synthesis, energy dispatch — and the post claims real production wins at Google Spanner, Klarna, FM Logistic, and WPP.
Who is it for?
research teams with a measurable scoring function and tolerance for an evolutionary search loop
Try it
deepmind.google/blog/alphaevolve-impact/