AdaptiCare: A Psychological Counseling Dialogue Enhancement Framework Based on Multi-Agent Collaboration

As global demand for mental health support continues to rise, AI-driven dialogue systems have become a vital supplement to traditional psychotherapy. However, existing systems that rely on a single large language model (LLM) struggle to manage multi-phase intervention processes and lack the ability to dynamically adapt to users' changing psychological states. This paper introduces AdaptiCare, a multi-agent collaborative framework for mental health counseling, which simulates professional team-based intervention through specialized agents-Counselor, Assistant, Reasoner, and Coach—and comprehensively covers the structured three-phase process of assessment, intervention, and consolidation. We further propose a multi-dimensional evaluation scheme integrating the CARE and TEQ scales to assess empathy, professionalism and strategic guidance, using GPT-3.5-turbo for consistent fine-grained scoring. Experimental results show that AdaptiCare outperforms a single-LLM baseline by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$11.2 \%, 6.5 \%$</tex>, and 4.5%, all statistically significant (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{p}&lt;0.05$</tex>). In comparison with seven leading models (e.g., LLaMA, Vicuna, PsyChat), AdaptiCare achieves superior performance in strategic coherence and overall balance. In a preliminary study involving 20 quasi-real users, participants showed significant improvement in PHQ-9 and GAD-7 scores, and gave generally positive feedback on the system's empathy, practicality of suggestions, and interaction fluency, further validating AdaptiCare's potential for real-world mental health support applications.