Title: Navigating Client Non-Compliance: A Framework for Strategic Counseling Based on Cognitive Conceptualization Diagrams

Abstract:

While Large Language Models (LLMs) demonstrate significant potential in the realm of psychological counseling, current evaluation benchmarks are predominantly built upon data from highly cooperative, simulated clients. This reliance obscures a critical phenomenon we term "counselor-following," wherein clients quickly transition from resistance to compliance within just a few conversational turns. This behavior generates a false sense of therapeutic advancement and artificially inflates performance scores under existing protocols by rewarding superficial empathy rather than genuine engagement. To rectify this discrepancy in evaluation standards, we introduce a resistance-aware framework anchored in Cognitive Behavioral Therapy (CBT). Central to this approach is CARS, a client simulator designed to explicitly model dynamic resistance using Cognitive Conceptualization Diagrams (CCDs). Additionally, we present STREAMS, a novel dual-module architecture that separates strategic reasoning (handled by the Thinker) from response generation (managed by the Presenter), with both components optimized through reinforcement learning. We also propose EWTS-MI, an entropy-weighted metric specifically tailored to assess responsiveness during high-friction interactions. Our experiments, conducted in both resistant and non-resistant counseling scenarios, confirm the existence of evaluation mismatches and highlight the efficacy of resistance-aware training in enhancing strategic robustness during challenging therapeutic exchanges.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC