Multiagent Large Language Model Framework for Psychotherapy Fidelity Assessment in Motivational Interviewing and...

Psychotherapy training is difficult to scale because manual rating of motivational interviewing (MI) and cognitive behavioral therapy (CBT) sessions is time-intensive, requires trained raters, and is subject to rater variability. Large language models (LLMs) may support simulation-based training and rubric-guided scoring, but early-stage evidence is needed before such systems can be applied to real learners.

This study aimed to conduct a simulation-based evaluation of a multiagent LLM framework for generating and scoring stylized MI and CBT training encounters.

We conducted a cross-sectional evaluation of a multiagent framework comprising Student, Patient, Evaluator, and Feedback agents. The Student agent conducted synthetic MI or CBT encounters with Patient agents derived from structured profiles. The Evaluator agent scored transcripts using study-specific MI and CBT scoring forms. Internal discrimination was tested across prompt-engineered novice, intermediate, and expert Student agent profiles using 133 MI and 102 CBT Patient profiles. Preliminary external grounding was assessed using 133 annotated motivational interviewing (AnnoMI) transcripts with coarse, metadata-derived, high/low session-level quality labels. Agreement with a pragmatic human-rater benchmark was evaluated using 16 independent human raters per modality. Criterion-level comparisons used paired Wilcoxon signed-rank tests with Benjamini-Hochberg false discovery rate correction at q<0.05. Agreement analyses used intraclass correlation coefficients (ICCs) with bootstrap 95% CIs. A secondary prompt-augmentation sensitivity analysis tested whether appending criterion-referenced feedback text to the novice Student-agent prompt shifted subsequent Evaluator-assigned scores.

Evaluator scores increased across prompt-defined Student-agent competence levels. For MI, overall mean scores increased from 1.18 (95% CI 1.16-1.20) for novice profiles to 1.75 (95% CI 1.69-1.81) for intermediate profiles and to 3.57 (95% CI 3.48-3.66) for expert profiles. For CBT, overall mean scores increased from 0.83 (95% CI 0.78-0.88) to 2.18 (95% CI 2.10-2.26) and 4.38 (95% CI 4.28-4.48), respectively. Criterion-level planned contrasts were significant after false discovery rate correction. On AnnoMI transcripts, Evaluator scores aligned with metadata-derived high/low session labels, with 91.7% classification accuracy at the prespecified threshold. Human interrater reliability was an ICC(2,1) of 0.866 for MI and 0.769 for CBT. Evaluator-vs-human-consensus agreement was an ICC(2,1) of 0.959 for MI and 0.934 for CBT. In the secondary prompt-augmentation analysis, overall MI scores shifted from 1.18 to 1.44, and CBT scores shifted from 0.83 to 1.01.

This proof-of-concept study suggests that a rubric-guided multiagent LLM framework can score stylized synthetic MI and CBT transcripts along expected prompt-defined competence gradients and align with preliminary external and human-rater benchmarks. The study is innovative in separating synthetic learner, patient, scoring, and feedback-generation roles within a single simulation workflow, extending prior LLM work from plausible dialogue generation toward rubric-linked scoring. The findings support further development of scalable simulation tools for psychotherapy training research. Prospective studies with human trainees, standardized or real patients, and formal psychometric testing are required.