MADP: Multi-Agent Deductive Planning for Enhanced Cognitive-Behavioral Mental Health Question Answer

TL;DR

The paper tackles Mental Health Question Answering (MHQA) by moving beyond single-agent CBT prompts to a Multi-Agent Deductive Planning (MADP) framework. MADP employs three specialized agents—Explorer $A_{EX}$, Empathizer $A_{EM}$, and Interpreter $A_{IN}$—to perform reverse ABC analysis (A → C → B → A) and to coordinate through three stages: multi-agent dialogue deduction, support planning, and response generation. It introduces the MADP dataset $D_{MADP}=\{ (p_i; k_i; v_i) \}$ and a MADP-LLM pipeline that uses instruction tuning and LoRA to fine-tune small open-source LLMs, guided by teacher models to produce planning-informed responses. Experiments across EMH and PsyQA datasets show consistent improvements in analytical depth, empathy, guidance, and overall quality, with strong cross-lingual performance and cost-efficient deployment. The work underscores the practical potential of structured, CBT-based multi-agent reasoning for personalized, responsible mental health support while highlighting ethical considerations and the need for human oversight in real-world applications.

Abstract

The Mental Health Question Answer (MHQA) task requires the seeker and supporter to complete the support process in one-turn dialogue. Given the richness of help-seeker posts, supporters must thoroughly understand the content and provide logical, comprehensive, and well-structured responses. Previous works in MHQA mostly focus on single-agent approaches based on the cognitive element of Cognitive Behavioral Therapy (CBT), but they overlook the interactions among various CBT elements, such as emotion and cognition. This limitation hinders the models' ability to thoroughly understand the distress of help-seekers. To address this, we propose a framework named Multi-Agent Deductive Planning (MADP), which is based on the interactions between the various psychological elements of CBT. This method guides Large Language Models (LLMs) to achieve a deeper understanding of the seeker's context and provide more personalized assistance based on individual circumstances. Furthermore, we construct a new dataset based on the MADP framework and use it to fine-tune LLMs, resulting in a specialized model named MADP-LLM. We conduct extensive experiments, including comparisons with multiple LLMs, human evaluations, and automatic evaluations, to validate the effectiveness of the MADP framework and MADP-LLM.

Figures (4)

• Figure 1: The diagram illustrates the limitations of single-agent methods in mental health support.
• Figure 2: (a) depicts the traditional ABC model in CBT, showing the sequential relationship between activating events A, beliefs/cognitions B, and consequences C. (b) presents the proposed multi-agent framework, illustrating the interactions among the three agents (arrows indicate the flow of information).
• Figure 3: An example of prompt for the MADP Framework.
• Figure 4: An overview of training MADP-LLM. It first utilizes help-seeking posts and MADP Prompt to generate MADP planning and answers, and then fine-tuning MADP-LLM.