Examining Human-Like Behaviors in LLMs: A Multi-Dimensional Analysis of Model Behaviors, User Factors, and System Prompts

A multi-dimensional analysis uses LLM-as-a-judge and human evaluation across 21,000 samples to quantify human-like behaviors like emotional expression and boundary setting. The study identifies how system prompts and user factors influence the prevalence and controllability of these traits, providing empirical metrics for managing model personas.