Mental health issues not only affect individuals’ quality of life and social functioning, but also pose broader risks to social well-being. Cognitive Behavioral Therapy (CBT), a widely used approach in mental healthcare, focuses on cognitive restructuring—modifying emotional and behavioral responses by identifying, challenging, and transforming negative or distorted automatic thoughts. Although CBT has demonstrated substantial effectiveness in clinical settings, its widespread implementation remains constrained by resource scarcity, high costs, and limited therapist availability. With the rapid advancement of artificial intelligence (AI), AI-powered CBT-style support systems offer a promising wang to deliver scalable, low-cost psychological support.

This study proposes a novel CBT-CoT (Cognitive Behavioral Therapy with Chain-of-Thought reasoning) framework, which combines CBT’s structured psychological reasoning process with a multi-agent architecture driven by large language models (LLMs). The framework is organized into two key stages—cognitive assessment and cognitive restructuring—and implemented through four collaborative and functionally distinct specialized agents.

In the cognitive assessment stage, the Automatic Thought Extraction Agent structurally decomposes the user’s message into a standardized three-column record comprising the precipitating event, the associated automatic thoughts, and the resulting emotional and behavioral responses. The Automatic Thought Evaluation Agent then analyzes these automatic thoughts by examining whether they are facts or interpretations, their belief intensity, and the types of cognitive distortions, and determines whether the conversation should proceed to the next stage of structured reasoning.

If restructuring is required, the CoT Reasoning Agent matches appropriate cognitive restructuring techniques and generates alternative rational thoughts by employing reasoning pathways such as hypothesis verification and evidence analysis. This agent outputs a structured five-column record aligned with standard CBT practice. Finally, the Response Generation Agent integrates the five-column record and the selected restructuring strategy to produce a personalized, empathetic, and cognitively supportive response tailored to the user’s concern.

Building on the methodology of PsyQA, we constructed a new large-scale Chinese psychological question-answering dataset comprising over 20,000 question-answer pairs across nine key domains, including interpersonal relationships, emotional adjustment, academic stress, personal growth, and others. The CBT-CoT system adopts a modular architecture, in which the four agents cooperate to complete a full-cycle CBT process—from the extraction and evaluation of automatic thoughts to the generation of supportive language based on structured reasoning.

To identify the most suitable base model for CBT-CoT, we evaluated three state-of-the-art Chinese LLMs: DeepSeek-Chat, Qwen-Plus, and Doubao-1.5-pro. DeepSeek-Chat consistently outperformed the other models in both reasoning quality and response fluency, and was selected as the base model for all four agents in the CBT-CoT framework.

Evaluation was conducted using both automated and manual methods. An ensemble of three large language models served as automated evaluators, while trained annotators performed human evaluations on a stratified sample. The results showed that CBT-CoT significantly outperformed baseline methods across six dimensions of psychological support: empathy expression, cognitive restructuring effectiveness, depth of reasoning, appropriateness of strategy, logical coherence, and degree of personalization. The total scores of human evaluation and LLM-based automatic scoring across the six dimensions were 17.18 and 17.48, respectively, with an absolute difference of only 0.30 points. In comparison with human-written responses (total score: 7.97), the CBT-CoT system (total score: 17.48) consistently generated more structured, coherent, and therapeutically oriented replies.

In summary, the CBT-CoT framework operationalizes the core cognitive restructuring flow commonly emphasized in CBT practice through multi-agent collaboration. By dividing the process into cognitive assessment and restructuring stages, and by explicitly modeling the chain-of-thought reasoning path, the system provides structured, explainable, and supportive responses. The framework demonstrates robust generalizability across diverse psychological topics and user intents. Future research may further optimize the CBT-CoT framework to improve its accuracy and adaptability in addressing complex psychological issues, while ensuring the integration of ethical safeguards into real-world deployment scenarios.