AlphaEvolve: A coding agent for scientific and algorithmic discovery: Evolutionary coding agent improves Strassen matrix multiplication for the first time in 56 years
In June 2025, Google DeepMind released AlphaEvolve, an evolutionary coding agent that uses LLMs to autonomously modify code and iteratively optimize, achieving breakthroughs in multiple scientific and computational problems. These include: discovering that 4×4 complex matrix multiplication requires only 48 scalar multiplications (first improvement to Strassen's algorithm in 56 years), optimizing data center scheduling algorithms, simplifying hardware accelerator circuits, and accelerating its own training LLM.
AlphaEvolve combines evolutionary algorithms with LLMs to autonomously improve algorithms and code, surpassing human experts on key problems such as matrix multiplication, scheduling, and circuit design. Its newly discovered matrix multiplication algorithm is the first breakthrough in 56 years, demonstrating AI's creativity in theoretical computer science. This work opens new avenues for automated scientific discovery and infrastructure optimization.
AlphaEvolve consists of multiple LLMs forming an evolutionary pipeline: a mutator (modifies code), an evaluator (runs tests and provides feedback), and a selector (retains improved individuals). In the matrix multiplication task, starting from Strassen's algorithm, after multiple generations of evolution, it discovered that only 48 scalar multiplications are needed (originally 49), and proved correctness. In the data center scheduling task, the optimized algorithm improved simulation efficiency. The system can be applied to any codable problem without human intervention.
AlphaEvolve can directly optimize computing infrastructure (e.g., data centers, chip design, compilers), reducing operational costs. Google has already used it internally, and it may be offered as a cloud service in the future. For industries reliant on efficient algorithms, such as finance and logistics, this technology can bring significant competitive advantages.
Recommend that cloud computing and chip design companies evaluate AlphaEvolve's applications in scheduling, compilation, and circuit optimization. Investment attention should be paid to Google DeepMind's subsequent productization, as well as SaaS platforms based on evolutionary coding agents.
Focus on AlphaEvolve's performance on larger-scale problems (e.g., NP-hard problems), as well as the interpretability and security of the generated code. If open-sourced, it will spawn a wave of AI-driven algorithm optimization startups. Caution is needed regarding potential vulnerabilities introduced by autonomous code modification.