Explain Like I'm 5 or Whatever I Choose: Evaluating the Potential for Interactive Complexity in Language Model Responses
Researchers tested if large language models can generate multiple versions of text with different levels of complexity for scientific information. They used a user study to define complexity measures like jargon, information density, and length. The findings showed that while models could vary complexity, their changes were often inconsistent, with most shifting the intended direction only about 34% of the time.