ChatAddiction: an LLM-Supported Chatbot for Substance Use Recovery

ACL ARR 2025 May Submission5557 Authors

Abstract: Substance use disorders (SUDs) affect over 36 million people worldwide, yet treatment adoption remains critically low due to stigma, resistance, and a lack of tailored support. While large language models (LLMs) show promise in mental health applications, existing systems lack integration of clinically grounded strategies such as cognitive behavioral therapy (CBT), limiting their impact in high-relapse domains like addiction recovery. We present ChatAddiction, the first multi-agent conversational framework designed to simulate and support long-term addiction treatment. ChatAddiction models therapeutic dialogues with dynamic patient profiles, real-world resistance factors, and a rich library of persuasive and behavioral interventions grounded in CBT and motivational interviewing (MI). To support training and evaluation, we construct a benchmark simulated recovery scenarios across Easy, Medium, and Hard levels, and adopt a two-stage training pipeline combining supervised fine-tuning (SFT) with direct preference optimization (DPO). Evaluation results show that ChatAddiction achieves a 41.5% gain in average motivation and \textbf{+0.49} in confidence over GPT-4o, while using 26% fewer turns to resolve hard cases. Additionally, ChatAddiction ranks higher in empathy, responsiveness, and behavioral realism based on automatically (GPT-4o judger) evaluation than GPT-4o and other LLMs. Our framework enables controlled, high-fidelity analysis of conversational strategies in addiction care and provides a scalable foundation for deploying emotionally intelligent AI therapists.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Chatbot, Addiction, Persuade

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 5557