Date: June 26, 2026
Time: 10:00 – 11:30 AM
Location: GHC – Room 4405
Zoom Link
Type: Ph.D. Thesis Defense
Who: Angela Chen
Title: Behavioral Modeling of Interpersonal Dynamics as Controllable Agentic Systems: Empirically-grounded Adaptive Virtual Patients for Psychotherapy Training

Abstract: The need for mental health care continues to outpace the supply of trained psychotherapists. Training is itself a bottleneck because supervision time is limited and trainees get few chances to practice repeatedly on realistic cases with feedback. Standardized patients played by actors are expensive and hard to scale, and the hardest clinical moments are not ones a trainee can safely rehearse with real clients. Large language models (LLMs) have made interactive virtual patients practical, yet most current systems are not empirically grounded: they carry little psychologically meaningful state and focus almost exclusively on one-on-one therapy, leaving multi-party interaction such as couples therapy unsupported.

This thesis treats the design of an LLM-based virtual patient (VP), an interactive agent whose behavior must remain coherent over a long interaction, as a control system. Rather than being steered by a fixed prompt, the VP’s behavior is governed by a dynamics controller, so its behavioral trajectory can be inspected and adjusted. The thesis develops this idea in three studies. The first builds LLM-based measures of therapist behavior and client behavior, applies them to a large corpus of clinical transcripts, and uses structural equation modeling to estimate how therapist micro-skills affect a client’s disclosure and emotion. The second turns these estimates into an adaptive virtual patient: detectors read the therapist’s empathy and exploration, a dynamics controller updates the patient’s disclosure state using the estimated coefficients, and the language model generates a reply at that level of disclosure. The third study extends to a multimodal, multi-agent interaction system in which two VPs and the therapist interact through a stage controller built around the demand–withdraw pattern common in couples therapy.

Methodologically, this thesis shows how to build human–AI interaction systems for a specialized domain that requires explicit behavioral modeling. Concretely, it delivers a validated pipeline for measuring therapy process, two training systems for individual and couples therapy, and a corpus of therapist–agent interactions. Together these broaden access to deliberate practice for trainees, using AI to strengthen therapists’ preparation rather than substitute for it.

Committee:
Haiyi Zhu (Chair), Carnegie Mellon University
Sherry Tongshuang Wu, Carnegie Mellon University
Aaron Steinfeld, Carnegie Mellon University
Holly Swartz, University of Pittsburgh Medical Center

LINK TO THESIS DRAFT DOCUMENT