AlphaEvolve on Google Cloud: AI for agentic discovery and optimization

Innovators in science and engineering face a common barrier: the search space for solving complex problems — like designing a new chip or discovering a drug molecule — is often too vast for standard brute-force methods to explore effectively. 

To help you overcome this challenge, we are releasing AlphaEvolve, a Gemini-powered coding agent for designing advanced algorithms, to Google Cloud, in private preview..

The challenge for innovators in science 

Many of the most challenging and potentially valuable problems  in the world are related to optimization. You might be trying to minimize latency in a data center, maximize the stability of a protein, or find the most efficient route for a logistics fleet.

AlphaEvolve pairs the creative problem-solving capabilities of our Gemini models with automated evaluators that verify answers, along with an evolutionary framework to improve upon the most promising ideas.

It then tests these changes against a "ground truth" evaluator that you define. If the new code performs better, it becomes the parent for the next generation. This creates a feedback loop that allows the system to learn and improve over time, eventually discovering algorithms that are significantly more efficient than the ones you started with.

How it works:

Here is how Alphaevolve discovers and improves upon existing algorithms in more detail:

• Input: You define a problem specification, evaluation logic (to measure how well a proposed solution works), and a seed initialization program. The seed is a compile-ready piece of code that is the algorithm that you want to optimize. To start, it solves the problem, even if sub-optimally.
• Mutation: Gemini models (Flash for speed, Pro for depth) process the context and generate mutated, optimized versions of the code that are added to the “population space.”
• Evolution: Evolution algorithms select which of the various code mutations from the population space to combine and further mutate to prioritize as the starting point for the next evolution of mutations.  
• Loop: The results from the Evaluation scores are then used by the ensemble of LLMs to generate the next set of improved solutions. The cycle repeats recursively, evolving the codebase from the initial seeds to state-of-the-art algorithms.

Proven impact at Google

At Google, we have already used this technology to tackle some of our own hardest engineering problems.

• Data center efficiency: AlphaEvolve found a better way to schedule tasks in our data centers, continuously recovering on average 0.7% of our global compute resources.

• Gemini training: AlphaEvolve sped up a vital kernel in Gemini’s architecture by 23%, leading to a 1% reduction in Gemini's training time.

• Hardware design: It accelerated the design of our next-generation TPUs by discovering more efficient arithmetic circuits.

To learn more about impact, read our paper.

How AlphaEvolve can help businesses across industries 

You can apply this same engine to your own proprietary data and unique algorithmic challenges. Here are a few ways improved algorithms can potentially help different industries:

• Biotech and pharma: Optimize the algorithms used for molecular simulation, which helps shorten the timelines for drug discovery and increases the success rate of new therapeutics.

• Logistics and supply chain: Discover superior heuristics for routing and inventory management, helping you reduce fuel costs and build more resilient delivery networks.

• Financial services: Evolve algorithmic risk models to manage complex portfolios more effectively.

• Energy: Optimize load balancing on smart grids to improve stability and better integrate renewable energy sources.

Get started on Google Cloud

AlphaEvolve is made to help with complex optimization problems that you can define in code and objectively measure. The AlphaEvolve Service API is now available through an Early access program with Google Cloud. If you have one of these problems and are interested in participating in the Early Access Program, please reach out to your Google Cloud Representative.