From Virtual Patients to Design Principles: Advancing Human–Agent Interaction in Sensitive Domains

Haiyi Zhu, Associate Professor, Carnegie Mellon University

As the use of generative AI and agentic systems expands into sensitive domains such as mental health, a key challenge is designing agents that interact with humans in ways that are effective, beneficial, and safe, while also uncovering generalizable principles of human–agent interaction. In this talk, I will first present our work, sponsored by the National Institute of Mental Health (NIMH), on designing virtual patient agents and automated feedback systems to support experiential training for mental health providers. We develop a safe, practice-oriented environment in which volunteer counselors and early-stage professionals can build skills without risk to real clients. Our Virtual Patient Agents are grounded in psychotherapy theory and informed by the analysis of 2,000 hours of psychotherapy transcripts, enabling realistic and adaptive patient simulations. We also introduce LLM-based assessments of core psychotherapy constructs, supported by theory-driven measurement frameworks and validated through studies with 27 trained annotators, demonstrating high agreement with human ratings. Building on these efforts, I conclude by synthesizing insights from our work and a broader literature review to identify key principles for designing effective, trustworthy, and adaptive human–agent interactions.