AlphaEvolve
AlphaEvolve is a Gemini-powered evolutionary coding agent that discovers algorithms. It recovered 0.7% of Google's fleet compute and beat Strassen for 4x4 complex matrices (48 multiplications).
Gemini proposes program edits, automated evaluators score them, and an evolutionary loop keeps the best. Sped a Gemini training kernel by 23% (1% less training time) and FlashAttention by up to 32.5%. On 50+ open math problems it matched best known results in ~75% and improved ~20%.
- Date
- Wednesday, 14 May 2025
- Lab
- Google DeepMind
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Fleet compute recovered by scheduling heuristic | 0.7% on average in production over a year | company |
| FlashAttention kernel speedup | up to 32.5% company-measured | company |
| Open math problems matched / improved | ~75% / ~20% of 50+ problems | company |
Successor of FunSearch (2023). The 48-multiplication result applies to complex-valued 4x4 matrices. Early-access academic program was planned at launch; generally available on Google Cloud 2026-07-09.
Sources
- deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algo
- blog.google/technology/ai/google-ai-updates-may-2025/
This record was checked against its sources on 6 October 2026. How we check