Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Using Llama-3 to Refine Psychotherapy in Silico

Kristin Witte¹, Milena Rmus¹, Elif Akata¹, Eric Schulz²; ¹Helmholtz Zentrum München, ²Max Planck Institute for Biological Cybernetics

Presenter: Kristin Witte

Psychotherapy is deeply personal and often time-intensive. Although therapeutic interactions crucially impact treatment outcomes, strategies to improve them often rely on trial and error, often resulting in a long and costly process with minimal improvement for the client. This project explores the potential of Large Language Models (LLMs) as low-risk, cost-effective tools for enhancing therapy. Using Llama-3, we simulated therapist-client dialogues, supervised by an expert LLM. The Expert LLM iteratively refined the Therapist LLM's responses, which were subsequently rated by the Client LLM. To extend this framework to real-world data, we applied LLM therapy revision to real therapy transcripts. For each segment of a recorded session, we compared the Client LLMs rating of the real therapist's last response to ratings of an LLM generated response and an LLM-based revision of the actual therapist's response, assessing satisfaction based on prior conversation context. These comparisons revealed that LLM-generated responses were often rated more favorably than human responses, with responses revised based on LLM feedback rated most favorably. This suggests that LLMs could meaningfully support and enhance therapeutic interactions, and improve quality of treatment.

Topic Area: Methods & Computational Tools

Extended Abstract: Full Text PDF