DeepMind AlphaEvolve cross-domain performance: 4×4 matrix multiplication refreshes the 1969 Strassen record, Gemini training is 1% faster

Google DeepMind released its AlphaEvolve cross-disciplinary results report on May 7 (U.S. time). In a post on the official DeepMind blog, DeepMind summarized the specific progress of AlphaEvolve since its launch: it found a better 4×4 complex matrix multiplication method than the Strassen 1969 algorithm (48 scalar multiplications); collaborated with mathematicians such as Terence Tao to solve multiple Erdős (Erdős) mathematical open problems; saved 0.7% of global compute resources for Google data centers; improved the speed of the key kernels trained for Gemini by 23%; and reduced overall Gemini training time by 1%.

Architecture: an evolutionary Agent with Gemini Flash for breadth exploration + Gemini Pro for deep evaluation

AlphaEvolve is an evolutionary coding Agent designed for general-purpose algorithm discovery and optimization:

Gemini Flash—maximizes the breadth of ideas to explore

Gemini Pro—provides deep, critical feedback

automated evaluator—verifies each candidate answer and provides feedback

evolutionary framework—continuously iterates based on evaluation feedback, keeping the most promising solutions

This structure enables AlphaEvolve to continuously generate and test solutions for open problems without prior human guidance, and is well-suited to domains where answers can be automatically verified (algorithms, mathematics, optimization problems).

Mathematics breakthroughs: 4×4 matrix multiplication refreshes the 1969 record, and—together with Terence Tao—resolves an Erdős problem

AlphaEvolve’s concrete progress in mathematics and computer science:

4×4 complex-valued matrix multiplication: found an algorithm that needs only 48 scalar multiplications, outperforming the best result presented by Strassen 1969

collaborated with renowned mathematicians such as Terence Tao to jointly solve multiple Erdős (Erdős) open problems

The Strassen algorithm is one of the long-standing best solutions for the computational complexity of matrix multiplication. In this case, AlphaEvolve broke a decades-old record, serving as a concrete example of “an AI Agent finding a new solution at the boundary of mathematics.”

Infrastructure breakthroughs: Google data centers save energy, and quantum circuit errors drop 10×

AlphaEvolve’s applications in Google’s in-house systems:

Data centers: found a better task scheduling method, averaging 0.7% savings of global compute resources

Gemini training: improved key kernel speed by 23%, and reduced total training time by 1%

Quantum physics: on the Google Willow quantum processor, the quantum circuits designed by AlphaEvolve had 10× lower error than a traditional best-optimization baseline, enabling complex molecular simulations to run on Willow

Power grid optimization: increased the proportion of feasible solutions for the graph neural network (GNN) model solving the AC Optimal Power Flow problem from 14% to more than 88%

Earth science: automated the optimization of Earth AI models, improving natural disaster risk prediction accuracy by 5%

Specific follow-up events to watch: whether AlphaEvolve will be opened from Google’s internal tools to external researchers, subsequent breakthroughs in the Erdős series of problems, and AlphaEvolve’s commercialization progress in Google Cloud (DeepMind has previewed related integrations in a Google Cloud blog post).

This article, DeepMind AlphaEvolve cross-disciplinary achievements: 4×4 matrix multiplication refreshes the Strassen 1969 record, and Gemini training is 1% faster, first appeared on Chain News ABMedia.