CCD-MAF: A Cognitive Conceptualization Diagram-driven Multi-Agent Framework for structured CBT dialogue generation

While Large Language Models (LLMs) have shown potential in mental health support, existing approaches often provide unstructured emotional responses and struggle to deliver structured, evidence-based interventions such as Cognitive Behavioral Therapy (CBT). Existing CBT dialogue datasets also lack sufficient clinical depth and phase-specific patient resistance. To address these limitations, we construct the CCD-CBT Dataset, a multi-phase CBT dialogue corpus integrating expert textbook knowledge with real-world doctor–patient interactions sourced from Haodf Online and other medical consultation websites. The dataset contains 1314 samples covering the four standard CBT treatment phases, from initial assessment to termination, with an average dialogue length of 132.12 turns. The average utterance length is 61.34 Chinese characters for patients and 97.05 Chinese characters for therapists. Based on this dataset, we propose CCD-MAF, a hierarchical Multi-Agent Framework guided by the Cognitive Conceptualization Diagram (CCD). CCD-MAF decomposes the CBT counseling process into macro-phase planning, meso-state tracking, and micro-technique execution, enabling dynamic modeling of patients’ cognitive structures and targeted interventions. We fine-tune Qwen2.5-7B as the therapist component within CCD-MAF and conduct single-turn and multi-turn clinical simulations. The evaluation distinguishes system-level comparisons among independently operated systems from therapist-component substitution comparisons within the shared CCD-MAF framework. Results indicate that CCD-MAF demonstrates strong performance in Cognitive Therapy Scale (CTS)-based simulated evaluations. In the component-substitution setting, the fine-tuned therapist component achieves the best performance among the evaluated therapist components under the same CCD-MAF framework. Ablation studies further support the contribution of the proposed multi-agent mechanism. This study provides an interpretable and clinically structured framework for developing AI psychological assistants, with its effectiveness evaluated in simulated counseling settings.