Proceedings of the Annual Meeting of the Cognitive Science Society

Psychotherapy is an inherently interactive and adaptive process, in which therapists continuously adjust intervention strategies according to clients' evolving emotional states. However, most existing LLM-based psychotherapy dialogue generation methods rely on single or fixed therapeutic techniques, failing to explicitly model the turn-level intervention decision-making. To overcome this drawback, we propose a knowledge-informed dynamic strategy adaptation multi-agent framework (K-DAF), which formulates psychotherapy as a turn-level emotion perception and adaptive strategy intervention process explicitly grounded in a psychotherapy knowledge base. K-DAF mainly contains the modules of client state tracking, knowledge-grounded strategy retrieval and recommendation, and strategy-conditioned response generation. Based on K-DAF, we construct a multi-turn psychotherapy dialogue dataset, PsyDAF, and fine-tune a psychotherapy-oriented model, DAF-Chat. Comparison experiments demonstrate the superiority of PsyDAF and DAF-Chat in terms of therapeutic alliance, empathic understanding, and counseling skills. The ablation study indicates the significance of the knowledge base and client state tracking.