AlphaEvolve - A general-purpose scientific AI agent launched by Google DeepMind

What is AlphaEvolve?

AlphaEvolve, a general-purpose scientific agent from Google DeepMind, designs and optimizes advanced algorithms by combining the creativity of large language models (LLMs) with an automated evaluator. Using two models, Gemini Flash and Gemini Pro, it continuously improves the most promising algorithms based on an evolutionary framework. AlphaEvolve has achieved significant results in areas such as data center scheduling, hardware design, AI training, and solving complex mathematical problems, optimizing matrix multiplication algorithms, improving data center efficiency, and achieving breakthroughs in several open mathematical problems. AlphaEvolve demonstrates algorithm development capabilities ranging from domain-specific to broad real-world challenges.

AlphaEvolve's main functions

• Discovery and optimization algorithmsIn the fields of mathematics and computation, we discover new algorithms and optimize existing ones.
• Improve computing efficiencyImprove data center scheduling efficiency, hardware design performance, and AI training speed.
• Solving complex mathematical problems: Propose new methods for solving complex mathematical problems, such as breakthroughs in matrix multiplication and geometric problems.
• Cross-domain applicationsIt supports applications in multiple fields, including materials science, drug discovery, and sustainability.

AlphaEvolve's technical principles

• Evolutionary computational frameworkThis approach uses an evolutionary algorithm to progressively improve code. The user defines an initial program, including the code blocks to be evolved and an evaluation function. LLM generates code modifications (diffs), which are applied to the current program to generate new candidate programs. Each newly generated program is scored based on the user-provided evaluation function, which returns one or more scalar metrics. Based on the evaluation results, the best-performing programs are selected for the next generation, preserving some diversity to explore a wider search space.
• The role of LLMLLM plays a central role in AlphaEvolve, responsible for generating code modifications and proposing new solutions. LLM's capabilities include: generating code modification suggestions based on the current program and historical information. Modifications can range from minor adjustments to complete code rewrites. LLM adjusts its generation strategy based on evaluation results, proposing better solutions in subsequent iterations. LLM processes rich contextual information, including problem descriptions, relevant literature, and code snippets, helping to generate code that better meets the problem's requirements.
• Evaluation mechanismAlphaEvolve's evaluation mechanism is automated. Users need to provide an evaluation function, which quantitatively evaluates the generated solution. The output of the evaluation function is typically one or more scalar metrics.
• Evolutionary DatabaseThe evolutionary database is used to store and manage programs and evaluation results generated during the evolutionary process. It saves all generated programs and their evaluation results, providing a reference for subsequent iterations. Based on algorithms, it ensures sufficient diversity during evolution, avoiding getting trapped in local optima. It quickly retrieves and selects programs with better performance, improving evolutionary efficiency.
• Distributed computingMultiple computational tasks run in parallel, with each task waiting for the results of other tasks when needed. This maximizes the number of samples evaluated by rationally allocating computational resources, thus improving the evolution speed. It supports operation on large-scale computing clusters, adapting to problems of varying sizes.

AlphaEvolve's project address

• Project official website:https://deepmind.google/discover/blog/alphaevolve
• arXiv technical paper:https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve

Application scenarios of AlphaEvolve

• Data Center SchedulingWe discovered efficient heuristics to optimize Borg scheduling, continuously restoring 0.7% of Google's global computing resources and improving task completion efficiency.
• Hardware DesignThe proposed Verilog rewrite removes redundant bits from the matrix multiplication circuit and integrates it into the Tensor Processing Unit (TPU), facilitating collaboration between AI and hardware engineers.
• AI training and inferenceOptimize matrix multiplication operations to improve training speed, reduce training time, and increase productivity in the Gemini architecture.
• Mathematical Problem SolvingDesign new algorithms, such as improving matrix multiplication algorithms, and solve open mathematical problems, such as raising the lower bound of the "kissing number" problem.
• Cross-domain applicationsIt is used in fields such as materials science, drug discovery, and sustainability to drive technological progress.