AlphaEvolve is a groundbreaking AI agent developed by Google that uses large language models (LLMs) and automated evaluators to evolve advanced algorithms for mathematics and computing. By combining the creative generation capabilities of LLMs with an evolutionary framework, AlphaEvolve can solve complex problems, improve system efficiency, and advance scientific discovery.
Built on Google’s Gemini model family, AlphaEvolve harnesses the strengths of two complementary LLMs: Gemini Flash, which explores a wide variety of ideas quickly, and Gemini Pro, which offers deeper, more refined insights. Together, they generate computer programs designed to solve algorithmic challenges. These programs are then tested, scored, and improved through a loop of automated evaluation and selection, forming an evolutionary process that continuously enhances results.
AlphaEvolve has made major contributions across Google’s computing ecosystem. In data center operations, it discovered a simple yet powerful heuristic for improving the efficiency of Borg, the system that manages Google’s vast infrastructure. This optimization has been in production for over a year and consistently recovers an average of 0.7% of global compute resources. The code is not only effective but also human-readable, making it easy to interpret, debug, and deploy.
In hardware design, AlphaEvolve proposed a new Verilog implementation for an arithmetic circuit used in matrix multiplication. This optimized design was verified for correctness and is being incorporated into an upcoming version of Google’s Tensor Processing Unit (TPU). The AI-assisted design process supports collaboration between engineers and algorithms, accelerating chip development.
AlphaEvolve also enhances AI training efficiency. It restructured matrix multiplication tasks—central to model training—into smaller, more manageable subproblems, increasing processing speed by 23% and reducing training time by 1%. More importantly, it significantly shortened the engineering time needed for kernel optimization, shrinking weeks of manual work into days of automated experimentation.
In the realm of low-level GPU programming, where performance is typically maximized by compilers, AlphaEvolve achieved up to a 32.5% speedup in optimizing the FlashAttention kernel, a key component in transformer-based AI models. These improvements not only boost speed and efficiency but also help engineers focus their efforts on innovation rather than routine tuning.
Beyond practical computing, AlphaEvolve excels in mathematical discovery. Starting with minimal code skeletons, it designed new components for gradient-based optimization procedures, leading to multiple novel algorithms. Notably, it developed an improved method for multiplying 4×4 complex-valued matrices using only 48 scalar multiplications, outperforming Strassen’s 1969 algorithm—a milestone in algorithm research.
The system was applied to over 50 open problems across mathematics, including geometry, combinatorics, and number theory. It reproduced known best results in 75% of cases and exceeded them in 20%, including a new lower bound for the kissing number problem in 11 dimensions.
As it continues to evolve, AlphaEvolve is being prepared for broader academic access through a user interface developed with Google’s People + AI Research team. Its general-purpose design enables application across any domain where problems can be expressed algorithmically and evaluated automatically. From sustainability to drug discovery, AlphaEvolve could become a transformative tool in science, technology, and engineering.