Demand for mental health care exceeds the capacity of the available clinical workforce. As a result, psychotherapy training often relies on limited supervision and uneven access to appropriate practice cases. Although standardized patients can provide controlled practice opportunities, they require substantial human resources and may raise practical and ethical concerns. Advances in generative AI have enabled large language model (LLM)-based virtual patients to support open-ended interactions. However, a persona prompt alone does not specify how a patient’s behavior—such as disclosure, affect, or trust—should evolve over the course of a session in response to therapist interventions. Multiparty settings, such as couples therapy, pose an additional behavioral-modeling challenge because each client’s behavior depends not only on the therapist but also on the other client.

This dissertation frames virtual-patient behavior as an explicit state-update process in which a separate controller updates psychological or interactional states in response to therapist interventions and other interactional inputs, while the language model generates dialogue consistent with those evolving states. Chapter 2 develops LLM-based measures of therapist empathy and probing skills, therapist--client rapport, client disclosure, and client emotion. It then models the therapist-client dynamics using structural equation modeling to estimate turn-level associations among these variables in a large corpus of psychotherapy transcripts. Chapter 3 applies the dynamic modeling to the design of an adaptive virtual patient whose disclosure state changes in response to a therapist’s use of empathy and probing skills. Chapter 4 extends the approach to simulate interactions among two client agents and a therapist. A stage controller coordinates six stages of couples therapy and a recurrent demand--withdraw pattern between the two agents.

By separating behavioral state transitions from language generation, this framework makes the underlying dynamics explicit and inspectable. Evaluations of both interactive systems—the individual virtual-patient and couples-therapy simulations—provide evidence that this approach can produce behaviorally faithful interactions that clinicians view as useful for training. More broadly, this work illustrates how agentic AI can broaden access to deliberate practice that develops, rather than replaces, human expertise.