The system, developed in conjunction with the University of Michigan, utilizes large language models to generate realistic conversations between simulated therapists and patients, opening new possibilities for learning communication and interaction skills in mental health. Access to real therapeutic conversations is limited due to information sensitivity, making ROGERS a valuable educational resource.
The work was presented by CiTIUS researcher David E. Losada at the Information Processing & Management Conference 2026, held from October 30 to November 1 in Wuhan. The contribution was selected for a session on conversational and contextual artificial intelligence for mental health. Initial evaluations by health specialists support its potential as a training tool.
The research was jointly developed by Anxo Fernández-Villoch, Marcos Fernández-Pichel (researchers at CiTIUS), and David E. Losada, in collaboration with Professor John Piette from the University of Michigan.
ROGERS builds upon previous results, adapting large language models (LLMs) to act as therapists, adhering to principles of Motivational Interviewing and Cognitive-Behavioral Therapy. The system combines various AI agents to reproduce conversations between virtual therapists and patients, who are constructed from structured psychological profiles and coherent personal histories.
Evaluations conducted with health experts revealed natural and sustained realistic interactions, supporting its educational potential. Researchers also observed that careful configuration of the models' instructions can foster collaborative and empathetic responses. CiTIUS is consolidating ROGERS into an interactive training tool that allows users to practice interaction strategies and analyze their responses, complementing the training of mental health professionals.
