DeepMind, the Google research lab in the field of AI, has introduced a new system called “AlphaEvolve,” which helps solve tasks with so-called “machine-evaluated” responses. According to the company, the system has already demonstrated effectiveness in optimizing the infrastructure that Google uses to train its own AI models. Currently, the team is developing an interface for interacting with “AlphaEvolve” and plans to launch an early access program for select researchers before expanding access to a broader audience.
The key feature of “AlphaEvolve” is an automatic response evaluation system that reduces the number of erroneous responses often found in modern AI models. The system generates several response options, analyzes them independently, and selects the most accurate ones. To work with it, the user must formulate a task and provide a mechanism for automatic solution verification, such as a formula.
“AlphaEvolve” is best suited for tasks that can be automatically evaluated — particularly in the fields of computer science and system optimization. At the same time, the system can only describe solutions in the form of algorithms, so it is not suitable for tasks that do not have a numerical nature.
In DeepMind’s tests, the system solved about fifty mathematical tasks, reproducing the best-known solutions in three-quarters of the cases and suggesting improvements in every fifth task. In practice, “AlphaEvolve” helped increase the efficiency of Google’s data centers, recovering an average of seven-tenths of a percent of computing resources, and also reduced the training time of Gemini models by one percent.
