ROGERS: Generating Synthetic Therapy Sessions using LLM Therapists Aligned with CBT and MI principles

Therapist-patient conversations are essential for training and evaluating counseling skills, but real clinical sessions are rarely available due to the sensitive nature of the data and the privacy requirements governing clinical interactions. This scarcity limits the development of reusable computational psychotherapy solutions and restricts opportunities for novice therapists to observe and analyze realistic therapeutic interactions. In this paper, we introduce ROGERS, a framework for generating synthetic multi-turn therapy sessions in which both patients and therapists are modeled with Large Language Models. Patients' roles are grounded in structured psychological profiles, which are expanded with consistent backstories to support longer and more realistic interactions. Therapists' roles are adapted to follow principles of Cognitive Behavioral Therapy and Motivational Interviewing through several alignment strategies, such as zero-shot prompting, one-shot prompting, and supervised fine-tuning on counseling datasets. We evaluate the resulting conversations using complementary CBT- and MI-based assessment protocols, including CTRS-inspired criteria and behavioral analyses guided by accepted rubrics. Our results show that carefully designed prompts can generate therapist behavior that is collaborative, empathic, and broadly aligned with CBT and MI principles. The resulting synthetic sessions are a promising educational resource for helping trainees identify appropriate interventions and problematic therapist behaviors. We also found that fine-tuning does not consistently improve therapeutic quality, and prompt-based alignment currently offers the most robust strategy for generating synthetic therapy conversations. Future work should explore hybrid approaches that combine the strengths of different LLM-based therapist variants.

Palabras clave: