The new system not only helps optimize the infrastructure used to train AI models but is also equipped with an automatic evaluation mechanism to reduce hallucinations. Hallucinations—false information generated by AI—are a major challenge for current AI systems. AlphaEvolve addresses this issue by generating multiple answer candidates, critiquing and analyzing them, and automatically selecting the most accurate one.
Users can provide AlphaEvolve with a problem along with optional details such as instructions, equations, code snippets, and relevant literature. Additionally, an automatic evaluation formula must be provided to verify the system’s answers. For this reason, AlphaEvolve can only be applied in specific fields, primarily computer science and system optimization.
During testing, AlphaEvolve was challenged with around 50 different math problems. According to DeepMind, the system rediscovered the best-known answers for 75% of the problems and found improved solutions in 20% of cases. In practical applications, AlphaEvolve helped recover an average of 0.7% of Google’s compute resources in data centers and reduced the training time for Gemini models by 1%.
However, DeepMind emphasizes that AlphaEvolve does not make groundbreaking scientific discoveries. Instead, it quickly and accurately proposes improvements already known or identified by other technologies, saving experts time. The company’s primary goal is to free specialists from routine computational tasks, allowing them to focus on more strategic and creative work.
