Davide Paglieri, Logan Cross, William A. Cunningham, Joel Z. Leibo, Alexander Sasha Vezhnevets
Persona Generators create diverse synthetic personas using large language models to improve AI system testing across varied user populations.
Testing AI systems that interact with humans requires understanding how they perform with different types of users, but gathering real human data can be costly and difficult. To tackle this, researchers have developed Persona Generators, which use advanced language models to create diverse synthetic personas that mimic human behaviors and beliefs. These generators are designed to cover a wide range of possible user traits, including rare combinations that are often overlooked. The study shows that these generators can produce more varied and representative synthetic populations than previous methods, making them useful for evaluating AI systems in different contexts.
