PICon: Multi-Turn Interrogation Framework Exposes Persona Agent Inconsistency

arXiv·medium signal

Kim, Im, and Choi present PICon, a multi-turn interrogation framework that evaluates whether LLM-based persona agents maintain consistency under sustained questioning. As persona agents are increasingly used as proxies for human participants in user research and surveys, the paper shows that most current agents break character under adversarial multi-turn probing. The framework measures how many turns of questioning it takes to produce contradictions — a critical metric for anyone deploying persona agents at scale.