DeepMind Unveils AlphaEvolve, an AI Agent for Algorithm Design

Google DeepMind is presenting AlphaEvolve, a large-scale language model (LLM) driven coding agent designed to discover and optimize complex algorithms applicable in advanced mathematics, hardware design, and computational efficiency.

"AlphaEvolve combines creative code generation by Gemini models with automated evaluators and an evolutionary framework that iteratively improves the most promising solutions," reads the official AlphaEvolve announcement.

The application of language models in highly specialized tasks has grown rapidly in recent years. Google had demonstrated in 2023 that LLMs could generate correct functions to solve open scientific problems. AlphaEvolve represents a significant advance by extending these capabilities into full algorithm evolution, employing an architecture that integrates automatic code generation, automated evaluation, and evolutionary selection.

This combination enables solving highly complex problems requiring formal precision and rigorous validation in both mathematical environments and industrial computing systems, while accelerating processes traditionally dependent on human experts.

Technical Specifications

AlphaEvolve was built based on a combination of Gemini models, using Gemini Flash to generate a broad spectrum of possible solutions and Gemini Pro to drill down to the most promising proposals. The generated solutions undergo automatic evaluations that verify their functionality and accuracy through quantifiable metrics.

AlphaEvolve has been implemented within Google's technology infrastructure with tangible results. In the area of data center scheduling, the agent discovered a simple heuristic that optimizes task orchestration in Borg, Google's workload management system. This solution, which has been in production for more than a year, steadily recovers 0.7% of Google's overall compute capacity, allowing more tasks to run on the same physical resources.

In hardware design, AlphaEvolve proposed modifications to Verilog, the standard circuit design language, which eliminated redundancies in a critical arithmetic circuit for matrix multiplication. These modifications were integrated into a future Tensor Processing Unit (TPU), validating the collaboration between AI and specialized hardware design.

In the area of AI model training, AlphaEvolve reduced Gemini model training time by 1% by optimizing the matrix multiplication kernel, more efficiently dividing operations into subproblems. This optimization also shortened the development time of new solutions from weeks to days, significantly speeding up model iteration. In the case of the FlashAttention kernel, AlphaEvolve achieved a speedup of up to 32.5%, a domain traditionally difficult to improve due to compiler-intensive intervention.

In mathematical research, AlphaEvolve designed new algorithms for multiplication of 4x4 matrices with complex values, reducing the number of scalar multiplications to 48, surpassing Strassen's 1969 algorithm. This advance also represents an improvement over AlphaTensor, Google's previous model specializing in multiplication algorithms.

The system was tested on over 50 open problems in areas such as mathematical analysis, geometry, combinatorics, and number theory. In about 75% of the cases, AlphaEvolve rediscovered state-of-the-art solutions and, in 20% of the cases, improved on the best existing solutions, says the company. 

Future outlook

AlphaEvolve is being prepared for external use through an early access program aimed at academic researchers, says Wired. Google is also developing a user interface to facilitate interaction with the system, in collaboration with the People and the AI Research team.

While the agent has shown effectiveness in specific domains such as mathematics, hardware architecture, and computational optimization, its generalist design allows it to be adapted to any problem whose solution can be expressed algorithmically and verified automatically. Potential application areas emphasized by Wired include materials science, drug discovery, sustainability, and other high-impact technology and business applications.