When Clients Stop Following: A Cognitive Conceptualization Diagram-driven Framework for Strategic Counseling
Researchers have developed a new framework to evaluate Large Language Models (LLMs) in psychological counseling settings. Current benchmarks often use simulated clients that quickly become compliant, creating a false impression of therapeutic progress. The proposed Cognitive Behavioral Therapy (CBT)-grounded approach introduces CARS, a client simulator that models dynamic resistance using Cognitive Conceptualization Diagrams (CCDs), and STREAMS, a dual-module framework for strategic reasoning and response generation optimized via reinforcement learning. An entropy-weighted metric, EWTS-MI, is also introduced to assess responsiveness in challenging interactions. AI
IMPACT Improves LLM evaluation in sensitive domains like counseling by simulating realistic user resistance.