Persona-Adoption by Large Language Models

Persona prompting, i.e. assigning a particularly identitify, role, or character to an LLM through a prompt, is a popular promting technique. In theory, persona prompting could lead to more diverse responses by LLMs. In practice, the ability of LLMs to pick up the provided personas seems to be varying.

The goal of this project is to identify patterns in the persona-adoption by LLMs to find out which kind of properties are more likely to be picked up and have large influence on generated outputs.

Related work:

• Quantifying the Persona Effect in LLM Simulations
• Evaluating Persona Prompting for Question Answering Tasks