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

Area of Science:

• Artificial Intelligence in Mental Health
• Computational Psychiatry
• Natural Language Processing in Therapy

Background:

• Large language models (LLMs) show promise for natural language generation but face challenges in clinical psychotherapy alignment.
• Existing LLMs struggle to adhere to professional therapeutic strategies required for effective mental health interventions.

Purpose of the Study:

• To introduce LLM4CBT, a novel LLM specifically designed for cognitive behavioral therapy (CBT).
• To evaluate LLM4CBT's ability to align with expert therapist behavior and elicit therapeutic responses.
• To assess LLM4CBT's adaptability in patient interactions, particularly for those struggling with engagement.

Main Methods:

• Development of LLM4CBT, an LLM tailored for cognitive behavioral therapy (CBT) principles.
• Experimental evaluation using real-world conversation data to compare LLM4CBT with human expert therapists.
• Testing on simulated conversation data to assess the model's capability in eliciting automatic thoughts and adapting conversational pace.

Main Results:

• LLM4CBT demonstrated alignment with expert human therapists on real-world data, showing more desirable therapeutic behaviors than other LLMs.
• Experiments indicated LLM4CBT effectively elicits unconscious automatic thoughts in simulated patient conversations.
• The model showed an ability to pause and wait for patient engagement, improving interaction for difficult-to-engage patients.

Conclusions:

• LLM4CBT presents a viable approach for developing AI-driven CBT therapists.
• Aligning LLMs with specific therapeutic instructions is key to their effective application in mental healthcare.
• The findings suggest significant potential for LLM-based interventions in psychotherapy.