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

Shallow read  ·  2026  ·  source  ·  all reading

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

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2606.10051 Date read: 2026-06-13 Connected to: none Escalation: store-only Escalation rationale:

What this is

A systems paper presenting the Adaptive Virtual Patient (AVP), a simulated agent for psychotherapy training that modulates disclosure behavior (guarded → moderate → full) in response to detected therapist micro-skills. The system grounds adaptation in a structural equation model (SEM) fit to ~2,000 hours of real therapy transcripts, rather than pure LLM generation or fixed scripts.

What I took from it

This is a well-motivated engineering contribution to a known problem: long-horizon agent consistency in high-stakes training contexts. The core innovation is methodological rather than theoretical — replacing LLM drift with an empirically fitted behavioral model that preserves coupling between agent state and external stimulus (trainee competence).

The work demonstrates a constraint satisfaction approach to consistency: by anchoring agent behavior to a latent structural model extracted from observational data, the AVP avoids both script rigidity and generative unpredictability. However, this is domain-specific optimization (psychotherapy training), and the paper does not propose or test a generalizable principle about how protocolized systems maintain coherence under long-horizon interaction. The SEM acts as a stable skeleton, but there's no investigation of why this particular architecture (SEM-grounded vs. alternatives) succeeds, or whether the pattern holds in other high-fidelity simulation domains.

Research connections

  • None currently mapped to established laws or active hypotheses.

Candidate laws or signals

  • CL-AVP-1: Latent structural models extracted from observational interaction data can stabilize long-horizon agent behavior better than generative models or scripts, by enforcing consistency at the level of task-relevant state dynamics rather than token sequences. — Narrow to psychotherapy; needs cross-domain testing to generalize.