AutoCBT: An Autonomous Multi-agent Framework for Cognitive Behavioral Therapy in Psychological Counseling

ACL ARR 2025 May Submission7008 Authors

Abstract:

Traditional face-to-face psychotherapy remains a niche practice, typically sought by individuals experiencing psychological distress. Online mental health consultation forums offer a viable alternative for those hesitant to seek help. In this context, large language models (LLMs) and cognitive behavioral therapy (CBT) jointly facilitate the development of automated online mental health consultation platforms. However, many automated mental health systems rely on rigid, rule-based agent workflows or single-prompt LLM responses, resulting in generic advice that lacks empathy and contextual awareness.
Inspired by the single-turn consultation style commonly found in online forums—where users, unlike in real-time multi-turn chat scenarios, are more willing to wait longer for thoughtful and in-depth replies—we developed AutoCBT, an autonomous multi-agent framework designed to improve the quality of automated mental health consultations. AutoCBT is built for single-turn consultation scenarios and introduces dynamic routing and supervisor mechanisms to generate high-quality responses.
Our research shows that AutoCBT consistently outperforms baseline models on key psychotherapy metrics, including empathy, cognitive distortion detection, and response relevance. Furthermore, we identify key challenges in implementing a multi-agent consultation framework, such as routing inconsistencies and LLM safety constraints. Our findings underscore the potential of AutoCBT as a scalable and effective AI-driven approach to mental health support.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: multi-agent, single-turn consultation, cognitive behavioral therapy, psychological counseling

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 7008