The Empirically Grounded Adaptive Virtual Patient for Psychotherapy Training: Disclosure That Responds to Therapist

Synthesis: The Adaptive Virtual Patient (AVP) is an Large Language Models (LLMs)-driven simulated patient for psychotherapy training that adapts its disclosure behavior — from guarded through moderate openness to full disclosure — in response to trainee skill. Grounded in a structural equation model fit to nearly 2,000 hours of real-world psychotherapy transcripts, the AVP quantifies how therapist empathy and exploration shift patient openness. An LLM generates utterances conditioned on a disclosure level updated each turn by the dynamics module. In evaluation with 20 clinicians and trainees over 80 sessions (1,033 turns), the AVP's disclosure rises in response to therapist empathy and exploration, while a prompt-only baseline stays flat; ablations confirm that the empirically motivated parameterization outperforms alternatives. This represents a significant advance in Intelligent Tutoring for professional training domains where authentic practice opportunities are scarce and where existing role-play systems lack grounded, adaptive behavior.

What this means for practice

• Learners. Sustain empathic and exploratory responding rather than landing one good opening question: across 1,033 turns from 20 clinicians and trainees the adaptive virtual patient's disclosure climbed steadily in response to therapist empathy and exploration, while a prompt-only baseline using the same LLM and persona stayed flat.
• Learners. Treat exploratory questioning as the primary lever. In the underlying psychotherapy corpus, exploration did roughly 78 times the work of empathy in driving disclosure change, even though the system's deployed weighting treated it as only 3 times as important.
• Learners. Do not judge a guarded patient's responsiveness from a short exchange. With the guarded Sam persona, perceived adaptivity reversed even though trajectory coherence held, because calibrated slow disclosure is psychologically correct but perceptually subtle; the authors call for longer sessions before such judgments are made.
• Learners. Pair simulated practice with human supervision and varied case formats, since the system's authors warn that a virtual patient rewarding empathy and exploration implicitly teaches trainees that those are the skills that matter, and that trainees may optimize for what the system rewards rather than for broader clinical judgment.

Limitations

• The evaluation rests on N = 20 clinicians and trainees across 80 sessions (1,033 turns); each participant committed roughly 90 minutes to four sessions of about 10 turns each, and the authors state this sample leaves the skill-sensitivity interaction test underpowered.
• Only two personas were tested, and the guarded persona produced a reversal in perceived adaptivity; the short evaluation window may be too short to surface longer-horizon effects of reduced defensiveness.
• Both the adaptive and the static systems used the same underlying large language model, so the authors frame the contribution as the architectural separation of behavioral state from language generation — the state-output alignment figures specifically may shift with other model families.
• The dynamics module is parameterized from a single psychotherapy corpus and carries that corpus's distribution of therapist-client behavior forward, so the authors warn that deployment in populations under-represented in the corpus may produce dynamics that do not match real interaction patterns; the framework is also unvalidated in any second domain, and the deployed 3:1 exploration weight is an integer approximation of the SEM-implied ratio rather than an optimized parameter (the ablation indicates the optimum exceeds 3:1).

Citation

Chen, A., Jin, S., Bao, C., Wang, C., Kraut, R. E., Wu, T., & Zhu, H. (2026). The Empirically Grounded Adaptive Virtual Patient for Psychotherapy Training: Disclosure That Responds to Therapist Micro-Skills.