New ontology reveals LLMs' psychotherapy approach and how to steer it

A new research paper introduces an ontology of ten therapeutic moves to analyze how large language models conduct psychotherapy sessions. The study found that models tend to over-use inquiry, neglect psychoeducation, and are heavily influenced by human-initiated strategies. However, exposing this ontology as a set of tools significantly reduced the deviation from human therapist behavior and improved turn-level alignment without requiring model fine-tuning. AI

IMPACT Provides a framework to analyze and potentially improve LLM interactions in sensitive domains like psychotherapy.

RANK_REASON Research paper published on arXiv detailing a new methodology for analyzing LLM behavior in a specific domain. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy
• MULTI-60

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →