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

Abstract

AlphaEvolve is an evolutionary coding agent powered by large language models for general-purpose algorithm discovery and optimization. It pairs Gemini's creative problem-solving with automated evaluators that verify answers, using an evolutionary framework to improve upon the most promising ideas. Leverages an ensemble: Gemini Flash maximizes breadth of ideas explored, while Gemini Pro provides critical depth with insightful suggestions.

Key Contributions

• General-purpose algorithm discovery through LLM-driven evolutionary search
• Improved upon Strassen's 1969 algorithm for 4x4 complex matrix multiplication (48 scalar multiplications)
• Matched SOTA in ~75% of 50+ open mathematical problems; improved upon best known solutions in ~20%
• Real-world deployment inside Google infrastructure for over a year

Results

Mathematical Breakthroughs

• Found algorithm for 4x4 complex matrix multiplication using 48 scalar multiplications (improving on Strassen's 1969 result)
• Matched or exceeded SOTA on 75%+ of open mathematical problems

Production Impact at Google

• 23% speedup on a vital Gemini architecture kernel → 1% reduction in Gemini training time
• 0.7% continuous recovery of global compute resources through better data center task scheduling
• Up to 32.5% speedup for FlashAttention kernel in Transformer models

2026 Updates

• Now uses Gemini 2.5 Pro for semantic evolution (rewriting logic/control flows, not just hyperparameters)
• AlphaEvolve Service API available via Google Cloud Early Access
• Open source implementations: OpenEvolve (distributed evolutionary algorithms, multi-language, multi-LLM)

Limitations

• Requires automated evaluators (not applicable to all problem types)
• Evolutionary search can be computationally expensive
• Results depend on quality of LLM ensemble

Source: AlphaEvolve — Google DeepMind