AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms

Source: DeepMind Date: 2025-05-14 URL: https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/

Summary

Google DeepMind published AlphaEvolve, an evolutionary coding agent that pairs Gemini Flash (breadth) with Gemini Pro (depth) to discover and optimize algorithms. Production results: 0.7% of Google’s worldwide compute recovered via a Borg scheduling heuristic; 23% speedup in a Gemini training kernel; 32.5% speedup in FlashAttention implementations; matrix multiplication for 4×4 complex-valued matrices solved in 48 scalar multiplications, surpassing Strassen’s 1969 algorithm. On 50+ open math problems, AlphaEvolve rediscovered SOTA in 75% and improved 20%, including a new lower bound of 593 for the kissing number in 11 dimensions.

Implications

0.7% of Google’s worldwide compute recovered is the most significant practical result. Google operates at a scale where 0.7% of global compute translates to hundreds of thousands of TPU-hours per year. A scheduling heuristic that recovers that much resource continuously in production is not an academic result — it’s a deployed system generating real economic value. That’s the bar AlphaEvolve cleared before publication.

Surpassing Strassen’s 1969 algorithm for 4×4 complex matrix multiplication is a genuine mathematical result. Strassen was the first reduction below O(n³) complexity for matrix multiplication. Improving on a 56-year-old result in even a specific matrix size is the kind of advance that gets published in mathematics journals, not just ML conferences. It also has direct implications for hardware-level linear algebra libraries.

The Flash/Pro evolutionary pairing is an architectural template for AI-driven research. Flash generates many candidate mutations at low cost; Pro evaluates depth and quality; automated verifiers select survivors. This is a concrete instantiation of “AI doing science” that isn’t “AI generates hypotheses humans test” — the verification loop is automated. The compute economics only work because Flash is cheap enough to run at evolutionary breadth.

Watch:

• Whether AlphaEvolve’s matrix multiplication improvements make it into XLA or cuBLAS — the path from academic result to hardware library adoption is the translation that matters
• Open-sourcing of AlphaEvolve: as with AlphaFold, external lab adoption of the evolutionary coding agent framework would accelerate the pace of algorithmic discovery
• Which of the 50+ math problems AlphaEvolve improved on get published as standalone results — the kissing number result alone is publishable