Google's artificial intelligence research lab DeepMind has developed a new artificial intelligence system, AlphaEvolve, which has a smart mechanism for reducing hallucinations.
Google's artificial intelligence research lab DeepMind has developed a new artificial intelligence system, AlphaEvolve, which has a smart mechanism for reducing hallucinations.
AlphaEvolve has an automated scoring system. It uses models to generate, critique, and pool possible answers to questions, and automatically evaluates and scores the accuracy of the answers, TechCrunch reports .
To use AlphaEvolve, users must provide the system with a task, optionally adding details such as instructions, equations, code snippets, and relevant literature. They must also provide a mechanism for the system to automatically evaluate the system's answers in the form of a formula.
Because AlphaEvolve can only solve problems that it can evaluate on its own, the system can only handle certain types of problems, particularly those from fields such as computer science and systems optimization. Another important limitation is that AlphaEvolve can only describe solutions as algorithms, making it poorly suited for problems that are not numerical.
To benchmark AlphaEvolve, DeepMind challenged the system to try to solve about 50 math problems spanning fields ranging from geometry to combinatorics. AlphaEvolve was able to “rediscover” the most well-known answers to the problems 75 percent of the time and find improved solutions 20 percent of the time, DeepMind claims.
DeepMind also evaluated AlphaEvolve on practical tasks, such as improving the efficiency of Google’s data centers and speeding up model training runs. According to the lab, AlphaEvolve generated an algorithm that continuously regenerates an average of 0.7% of Google’s global computing resources. The system also proposed optimizations that reduced the overall time it took Google to train Gemini models by 1%.
It’s clear that AlphaEvolve isn’t making groundbreaking discoveries. In one experiment, the system was able to find improvements in the design of Google’s TPU AI accelerator chip that had not previously been noticed by other tools.
DeepMind, however, claims that AlphaEvolve can save time, freeing up experts to focus on other, more important work.
The company said it is creating a user interface to interact with AlphaEvolve and plans to launch an early access program for select scientists before a possible wider rollout.
