Language fashions prompted with a person description or persona are getting used to foretell the person’s preferences and opinions. Nevertheless, present approaches to constructing personas principally depend on a person’s demographic attributes and/or prior judgments, however not on any underlying reasoning behind a person’s judgments. We introduce PB&J (Psychology of Habits and Judgments), a framework that improves LM personas by incorporating potential rationales for why the person may have made a sure judgment. Our rationales are generated by a language mannequin to explicitly purpose a few person’s conduct on the premise of their experiences, persona traits, or beliefs. Our technique employs psychological scaffolds: structured frameworks such because the Huge 5 Persona Traits or Primal World Beliefs to assist floor the generated rationales in present theories. Experiments on public opinion and film desire prediction duties display that language mannequin personas augmented with PB&J rationales constantly outperform personas conditioned solely on person demographics and / or judgments, together with people who use a mannequin’s default chain-of-thought, which isn’t grounded in psychological theories. Moreover, our PB&J personas carry out competitively with these utilizing human-written rationales, suggesting the potential of artificial rationales guided by present theories.
- † College of Southern California
- ** Work accomplished whereas at Apple
