Leveraging Gemini models and evolutionary coding to generate novel algorithms, solve complex challenges, and redefine innovation.
Discover AlphaEvolveAlphaEvolve is a pioneering evolutionary coding agent, fueled by Google's advanced Gemini models. It's engineered for general-purpose algorithm discovery and optimization, capable of evolving entire codebases to address highly complex problems in mathematics, computer science, and beyond. It goes beyond single function discovery, developing much more complex algorithmic solutions.
AlphaEvolve utilizes an ensemble of Gemini models: Gemini Flash maximizes the breadth of explored ideas with its speed, while Gemini Pro provides critical depth with insightful suggestions. These models propose programs, which are then rigorously verified, run, and scored by automated evaluators. The most promising ideas are selected and refined through an evolutionary process, driving a continuous cycle of improvement and discovery.
AlphaEvolve achieved a landmark by discovering an algorithm to multiply 4×4 complex-valued matrices using only 48 scalar multiplications. This surpasses Strassen's 1969 algorithm (49 multiplications over any field with characteristic 0) in this specific setting for the first time, a record that stood for 56 years. AlphaEvolve outperformed even specialized tools like AlphaTensor in this context and reportedly improves the state of the art for 14 matrix multiplication algorithms. Discussions in the scientific community continue to compare this with prior work, such as Waksman's algorithm, in different contexts.
AlphaEvolve has tackled over 50 open problems across mathematical analysis, geometry, combinatorics, and number theory. It rediscovered state-of-the-art solutions in approximately 75% of cases and improved upon previously best-known solutions in 20%, including advancing the centuries-old kissing number problem by establishing a new lower bound in 11 dimensions with 593 outer spheres.
AlphaEvolve is enhancing efficiency across Google's infrastructure. For data center scheduling with Borg, it recovered approximately 0.7% of worldwide compute resources by generating a simple, human-readable heuristic. In hardware design, it proposed a Verilog rewrite for an arithmetic circuit in upcoming TPUs, removing unnecessary bits. This change was validated by TPU designers and demonstrates AlphaEvolve's capability to refine source RTL early in the design flow, even if such optimizations might also be caught by downstream synthesis tools.
By optimizing vital kernels, AlphaEvolve has significantly accelerated AI model training. It achieved a 23% speedup in a key matrix multiplication kernel within Gemini's architecture, leading to a 1% reduction in overall Gemini training time. Separately, it demonstrated up to a 32.5% speedup for the FlashAttention kernel implementation. These advancements dramatically reduce engineering time for kernel optimization from weeks to days, fostering faster innovation.
AlphaEvolve harnesses an ensemble of Google's state-of-the-art Gemini models. Gemini Flash, with its speed and efficiency, maximizes the breadth of explored ideas. Concurrently, Gemini Pro, with its greater capabilities, provides critical depth with insightful and creative suggestions. This strategic mix forms a powerful synergy for algorithmic discovery and code generation.
A cornerstone of AlphaEvolve's success is its evolutionary framework, inspired by a combination of MAP-Elites and island-based population models. Proposed programs are objectively scored for accuracy and quality using automated evaluators, ensuring that only correct and high-performing solutions are pursued. This systematic, iterative approach, where LLMs propose changes and evaluators guide refinement, is crucial for making progress in complex domains.
Watch exclusive interviews and deep dives into AlphaEvolve's breakthroughs with researchers from Google DeepMind.
AlphaEvolve represents more than an AI tool; it signifies a new paradigm for human-AI collaboration in scientific research and engineering. By acting as a creative partner, it empowers researchers to tackle previously intractable problems and accelerate the pace of discovery and innovation.
The journey with AlphaEvolve is just beginning. Google DeepMind is planning an Early Access Program for selected academic users and exploring broader availability. Its general-purpose nature promises continuous improvement and adaptation, opening doors to unforeseen applications in material science, drug discovery, sustainability, and beyond.