Aligning large language models for cognitive behavioral therapy: a proof-of-concept study

Recent advancements in large language models (LLMs) have significantly impacted society, particularly with their ability to generate responses in natural language. However, their application to psychotherapy is limited due to the challenge of aligning LLM behavior with clinically appropriate responses. In this paper, we introduce LLM4CBT, designed to provide psychotherapy by adhering to professional therapeutic strategies, specifically within the framework of cognitive behavioral therapy (CBT). Our experimental results on real-world conversation data demonstrate that LLM4CBT aligns closely with the behavior of human expert therapists, exhibiting a higher frequency of desirable therapeutic behaviors compared to existing LLMs. Additionally, experiments on simulated conversation data show that LLM4CBT can effectively elicit automatic thoughts that patients unconsciously possess. Moreover, LLM4CBT is able to pause and wait until they are prepared to participate in the discussion for patients experiencing difficulty in engaging with the intervention, rather than continuously pressing with questions. The results demonstrate potential possibilities on designing LLM-based CBT therapists by aligning the model with appropriate instructions.