StoryMI: Steerable Multi-Agent Therapeutic Dialogue Generation

StoryMI introduces a multi-agent LLM framework that generates therapeutic dialogue grounded in patient narratives and dynamically controlled MI strategies. The system benchmarks six LLMs across 6,000 simulated dialogues and demonstrates that situational context and macro-level strategy control improve clinical adherence to motivational interviewing standards.

StoryMI addresses a critical gap in AI-assisted psychotherapy: the need for clinically grounded dialogue generation that adheres to established therapeutic methodologies. Traditional LLMs excel at fluent text generation but lack the domain-specific constraints required for motivational interviewing (MI), a evidence-based technique for behavioral change. This research bridges that divide by combining narrative grounding with multi-agent coordination, where separate agents handle therapist responses, client simulation, and strategic interaction management.

The framework's innovation lies in its two-level evaluation approach. Rather than relying solely on generic language metrics, the authors developed MI-specific assessments aligned with clinical standards, including human expert evaluation. This methodological rigor distinguishes the work from broader dialogue generation research and establishes a benchmark for therapeutic AI.

The implications extend beyond academic interest. Healthcare systems increasingly explore AI to address therapist shortages and expand access to mental health services. StoryMI's 1,000 questionnaire-story pairs and coverage of 12 MI codes and 13 symptom domains create a reusable foundation for clinical deployment. The open-sourcing of code and datasets accelerates adoption across research and commercial implementations.

Looking forward, the challenge shifts from proof-of-concept to real-world validation. Clinical trials comparing StoryMI outputs with human therapists remain essential before deployment in actual treatment settings. The framework's effectiveness on diverse populations and complex comorbidities also requires investigation. Success here positions AI-assisted psychotherapy as a mainstream tool, potentially reshaping mental health delivery infrastructure globally.

• →StoryMI uses narrative grounding and multi-agent coordination to generate motivational interviewing dialogue that adheres to clinical standards.
• →A two-level evaluation protocol combining lexical metrics with MI-specific clinical assessments validates therapeutic quality beyond standard NLP benchmarks.
• →The dataset of 6,000 dialogues across 12 MI codes and 13 symptom domains provides a reusable foundation for scaling therapeutic AI systems.
• →Benchmarking six LLMs reveals that situational context and macro-level strategy control meaningfully improve MI adherence and clinical plausibility.
• →Open-sourced code and data accelerate the path toward real-world deployment of AI-assisted psychotherapy in healthcare systems.