CALM-IT: Generating Realistic Long-Form Motivational Interviewing Dialogues with Dual-Actor Conversational Dynamics Tracking

TL;DR

CALM-IT tackles the instability of long-form mental health dialogues by explicitly modeling bidirectional conversational dynamics as a latent state-space for therapist-client interactions in Motivational Interviewing. The framework generates long-form, therapeutically coherent transcripts by continuously updating inferred alignment, mental states, and short-term goals for both agents, guided by MI expertise. Large-scale evaluations against strong baselines show CALM-IT achieves higher effectiveness and goal alignment, with markedly better stability across conversation lengths and more effective therapist redirections. The work demonstrates that explicit state modeling and state-informed control of interventions are key for high-quality, scalable synthetic therapeutic dialogues and offers a reproducible platform for stress-testing MI-supportive LLMs, with careful attention to ethics and evaluation rigor.

Abstract

Large Language Models (LLMs) are increasingly used in mental health-related settings, yet they struggle to sustain realistic, goal-directed dialogue over extended interactions. While LLMs generate fluent responses, they optimize locally for the next turn rather than maintaining a coherent model of therapeutic progress, leading to brittleness and long-horizon drift. We introduce CALM-IT, a framework for generating and evaluating long-form Motivational Interviewing (MI) dialogues that explicitly models dual-actor conversational dynamics. CALM-IT represents therapist-client interaction as a bidirectional state-space process, in which both agents continuously update inferred alignment, mental states, and short-term goals to guide strategy selection and utterance generation. Across large-scale evaluations, CALM-IT consistently outperforms strong baselines in Effectiveness and Goal Alignment and remains substantially more stable as conversation length increases. Although CALM-IT initiates fewer therapist redirections, it achieves the highest client acceptance rate (64.3%), indicating more precise and therapeutically aligned intervention timing. Overall, CALM-IT provides evidence for modeling evolving conversational state being essential for generating high-quality long-form synthetic conversations.

Figures (23)

• Figure 1: Architecture of the CALM-IT framework. Therapist and client agents interact through state-dependent turns, with latent conversational dynamics inferred and updated throughout the dialogue to guide generation and evaluation of long-form MI interactions.
• Figure 2: Percentage change in performance across metrics from short-form conversations (30 turns) to long-form conversations (100 turns), averaged across all pairs of short-long conversations. Positive or close-to-zero values indicate stable performance, whereas negative values indicate degradation in performance. (RQ1)
• Figure 3: Distribution of changes in client sustain talk at the greatest therapist-initiated redirection moment, for all four tested frameworks. Negative values indicate reductions in sustain talk following redirection, while positive values indicate increases. Solid vertical lines denote the mean change for each framework, and dashed lines represent no change (RQ2)
• Figure A1: Prompt for Patient System-1 Evaluation of Therapist turn.
• Figure A2: Prompt for Patient Rapport Delta Update.