Empirical Modeling of Therapist-Client Dynamics in Psychotherapy Using LLM-Based Assessments

TL;DR

SEM showed that therapist empathy and exploration directly shaped client disclosure and emotional expression, whereas rapport may contribute to reductions in internal emotional distress rather than increased willingness to express it.

Abstract

Psychotherapy is a primary treatment for many mental health conditions, yet the interplay among therapist behaviors, client responses, and the therapeutic relationship remains difficult to untangle. This work advances a computational approach for modeling these moment-to-moment processes. We first developed automated methods using large language models (LLMs) to assess therapist behaviors (e.g., empathy, exploration), relational qualities (e.g., rapport), and client outcomes (e.g., disclosure, self-directed and outward-directed negative emotions). These measures showed strong alignment with human ratings (mean Pearson $r = .66$). We then analyzed nearly 2,000 hours of psychotherapy transcripts from the Alexander Street corpus using Structural Equation Modeling (SEM). SEM showed that therapist empathy and exploration directly shaped client disclosure and emotional expression, whereas rapport may contribute to reductions in internal emotional distress rather than increased willingness to express it. Together, these findings demonstrate how computational tools can capture core therapeutic processes at scale and offer new opportunities for understanding, modeling, and improving therapist training.

Figures (6)

• Figure 1: Hypothesized Relationships Hypothesized links among therapist behaviors, the therapist--client relationship, and client states. Nodes include therapist behaviors (empathy & exploration), the therapist--client relationship (rapport), and client behaviors (self-disclosure & negative emotion). This figure illustrates the following four hypotheses: H1: Client self-disclosure in a given utterance is predicted by the therapist’s immediately preceding empathy and exploration (encouragement of deeper discussion). H2: Client self-disclosure is shaped by the cumulative rapport built prior to the current session. H3: Higher therapist empathy and exploration are associated with greater expression of negative emotions. H4: Higher cumulative rapport between therapist and client is associated with greater expression of negative emotions.
• Figure 2: The automatic assessment process starts with designing LLM prompts to generate quantitative psychological assessments and validating the prompts with human annotators.
• Figure 3: The results show that rapport was associated with reduced negative emotions and a slight decrease in disclosure. Exploration increased disclosure but also elevated self-directed negative emotions. Empathy had the strongest effects, substantially enhancing disclosure while simultaneously increasing both forms of negative emotions.
• Figure 4: Principal Component Analysis (PCA) of eight utterance-level emotion variables (anger, contempt, disgust, enjoyment, fear, sadness, surprise, anxiety). The first three components explained most of the structured variance (PC1 = 37.98%, PC2 = 21.36%, PC3 = 13.55%; cumulatively 72%). A clear scree-plot elbow appeared after the third component, and the loading patterns indicated three interpretable dimensions. Accordingly, we retained three components.
• Figure 5: Correlation matrix of emotions