Tag
This pre-registered study evaluates how detailed biographical personas in system prompts affect LLM code generation, finding model-dependent effects: one model showed strong persona effects on output length and identity enactment (including refusal to code from a librarian persona), while another showed weaker effects. Personas acted as behavioral-policy biases rather than universal quality improvements.
This paper systematically evaluates assumptions about LLM persona prompting and identifies 'persona manifold collapse,' where richer persona descriptions reduce behavioral diversity and simulation fidelity. The findings show that simple age-gender personas often outperform more detailed profiles.
This paper analyzes when persona prompting improves LLM responses, finding that it increases expertise depth at the cost of clarity, with effectiveness varying by domain and question type. The study introduces hybrid retrieval for role selection and advocates for multi-metric evaluation.
This paper studies how persona prompting influences language generated by multimodal large language models in urban perception, finding that captions converge while justifications vary systematically with persona attributes.