PSI-Bench: Interpretable and Clinically Meaningful Evaluation of Depression Patient Simulators
PSI-Bench is a framework designed to automatically and interpretably evaluate depression patient simulators across turn, dialogue, and population levels. It uses five dimensions—including narrative processes, emotion expression, response length, linguistic markers, and lexical diversity—to diagnose simulator behavior against real patient data. Findings show simulators are too verbose and lack the necessary emotional variability.