Adaptive Virtual Patient adjusts disclosure based on therapist empathy

The Adaptive Virtual Patient (AVP) tackles a key limitation in psychotherapy training simulators: current systems either follow rigid scripts or use LLMs that drift unpredictably during long sessions. AVP, presented by Angela Chen and colleagues, introduces a dynamics module that updates the patient's disclosure level turn-by-turn, ranging from guarded to moderate openness to full disclosure. This module is driven by a structural equation model fitted to nearly 2,000 hours of actual psychotherapy transcripts, which quantifies exactly how therapist empathy and exploratory probing shift a patient's openness over time. An LLM then generates the virtual patient's utterances conditioned on that disclosure level, ensuring realistic, adaptive behavior.

In a rigorous evaluation with 20 clinicians and trainees across 80 sessions totaling 1,033 conversational turns, AVP demonstrated a clear adaptive signal: when therapists showed more empathy and exploration, the patient's disclosure rose measurably. In contrast, a prompt-only baseline (no dynamic disclosure control) stayed flat regardless of therapist behavior. Ablation experiments confirmed that the empirically grounded parameterization outperformed alternatives, with exploration carrying the strongest adaptive weight. This work shows that coupling LLM generation with data-driven dynamics can create far more realistic, responsive simulated patients for scalable mental health training.