An LLM-based Motivation-Aware Framework For AI Coaching For Behaviour Change

Paper Title

An LLM-Based Motivation-Aware Framework For AI Coaching For Behaviour Change

Publication Info

• Topic area: AI-driven digital health coaching for behavior change.
• Keywords: Motivational Interviewing, Large Language Models, behavior change, health coaching, physical activity, conversational agents, user motivation, therapeutic alliance, MI strategies, AI empathy.

Background and Problem

• Problem / challenge: Existing digital health coaching systems focus heavily on emotional support but neglect goal-setting behaviors, which are crucial for clients ready to take action. Many systems also fail to adapt strategies to users’ motivational states, limiting their effectiveness.
• Significance: Addressing this gap can improve the scalability and accessibility of behavior change interventions, especially for promoting physical activity and other health-related behaviors.
• Motivation and related work: Previous systems relied on rule-based or retrieval-based frameworks, limiting conversational flexibility. While LLMs enhance flexibility, they often lack controllability and explainability. Recent MI-based systems have shown promise but still struggle with balancing emotional support and actionable guidance, particularly for clients at different stages of readiness.

Solution

• Proposed approach: MI-aware coaching framework that dynamically adapts strategies based on users’ motivational states, leveraging MI principles and psychological theories like the Transtheoretical Model (TTM).
• Novelty:
  1. Adaptive coaching framework that personalizes strategies to users’ motivational states.
  2. Integration of MI principles with LLMs for enhanced conversational flexibility and adherence to therapeutic guidelines.
  3. Mixed-methods evaluation demonstrating significant increases in users’ readiness to change and insights into successful coaching behaviors.
• Procedure and key techniques:
  • Dialogue history is analyzed using three components: motivational language detection, phase management, and strategy selection.
  • Motivational states are classified into attitude (change/neutral/sustain) and strength (high/medium/low) using in-context learning (ICL).
  • Coaching sessions are divided into five phases (engaging, focusing, evoking, planning, concluding), with strategies tailored to each phase.
  • Response generation uses retrieval-augmented generation (RAG) to align with MI principles and user context.

Results

• Concrete findings:
  • MI-aware agent increased users’ readiness to change by 1.47 points on average, compared to 0.84 for the baseline.
  • MI-aware agent demonstrated higher adherence to MI technical skills (CEMI technical items: 3.59 vs. 3.46).
  • Participants who agreed on a change plan with the agent experienced the greatest motivational increases.
• Advantage over baselines:
  • 32% higher user engagement (236 vs. 179 words per chat).
  • Improved compliance with MI principles (72% vs. 50% adherence to collaborative strategies).
  • Lower occurrence of unsolicited advice (0.25% vs. 1%).
• Experiments / evaluation:
  • Mixed-methods user study with 140 participants recruited via Prolific.
  • Measures included readiness to change (Contemplation Ladder), perception of MI (CEMI), and therapeutic alliance (WAI).
  • Qualitative analysis identified key behaviors (plan agreement, cooperation) contributing to successful coaching sessions.
• Limitations and future work:
  • Single-session design limits insights into long-term behavior change.
  • Fixed phase durations may not suit all users; dynamic phase transitions could improve user experience.
  • Lack of multimodal inputs (e.g., non-verbal cues) restricts natural interaction.

Summary

This paper presents an MI-aware coaching framework that adapts strategies to users’ motivational states using LLMs and MI principles. The system demonstrated significant increases in users’ readiness to change and adherence to MI best practices, outperforming a baseline MI agent. Key findings highlight the importance of plan agreement and cooperation in successful coaching sessions. While the framework shows promise for scalable digital health interventions, future work should address long-term engagement, dynamic session management, and multimodal interaction capabilities.