Title: DMT-CBT: Modeling Therapeutic States Over Time for CBT Counseling

Abstract:

While large language models (LLMs) are increasingly being applied to Cognitive Behavioral Therapy (CBT) counseling, current methodologies largely treat the process as a localized task of generating immediate responses. These approaches typically prioritize empathetic replies within brief, text-only, or single-session interactions. We contend that this perspective is fundamentally misaligned with the reality of psychotherapy. In clinical CBT, treatment is a longitudinal endeavor where clinicians continuously infer, update, and intervene based on evolving therapeutic states across multiple sessions. Consequently, realistic CBT requires models to handle multimodal inference and account for the delayed effects of interventions that span across sessions. This necessitates the ability to track the evolution of therapeutic states under conditions of partial observability.

To address these challenges, we introduce DMT-CBT, a framework designed for the Dynamic Modeling of Therapeutic states in CBT counseling. DMT-CBT sustains structured therapeutic states throughout the course of treatment, integrating multimodal behavioral grounding and tool-augmented interventions to facilitate adaptive therapeutic reasoning. Leveraging this framework, we developed DMTCorpus, a synthetic dataset comprising multi-session, multimodal CBT counseling interactions. This corpus is characterized by evolving therapeutic states, client behaviors grounded in images, and continuity in cross-session interventions. Our experimental findings indicate that DMT-CBT outperforms post-hoc extraction methods by enhancing counseling fidelity and therapeutic alliance, fostering more positive longitudinal affective trajectories, and maintaining therapeutic states with greater accuracy.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC