Sign InOpen Brain
arXivPaperNeeds Review

Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy

A tool-exposed ontology steered models closer to human therapy patterns without fine-tuning, showing how explicit action vocabularies can improve agent behavior.

arXiv · Aug 21, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The researchers define **10 therapeutic moves**, validated with **5 licensed psychologists**. Frontier models used inquiry at up to **3× the human rate**, while exposing the moves as tools improved turn-level alignment by **7–9 percentage points**.

Practical Implication

For builders, this suggests representing desired behavior as explicit, callable actions rather than relying only on prose instructions. A compact action ontology can make agent behavior measurable and steerable without fine-tuning.

Agent-Ready Context
The researchers define **10 therapeutic moves**, validated with **5 licensed psychologists**. Frontier models used inquiry at up to **3× the human rate**, while exposing the moves as tools improved turn-level alignment by **7–9 percentage points**.

For builders, this suggests representing desired behavior as explicit, callable actions rather than relying only on prose instructions. A compact action ontology can make agent behavior measurable and steerable without fine-tuning.

The evidence concerns psychotherapy, where behavioral alignment is safety-sensitive and human distributions are not automatically ideal outcomes. The material does not establish whether the method transfers to coding agents.
Connected Context · Feed7 Judgment

This converts an expert-defined behavioral taxonomy into both an evaluation surface and a steering interface. It strengthens the case for encoding domain practice as observable actions rather than trusting prose instructions or model introspection, while narrowing the claim to turn-level therapeutic behavior: matching human move distributions does not itself establish clinical quality or transfer to other agents.

Trading Desks to Clinical Trials: Parallels in Applied Vertical AI — Ayush Bhardwaj, Allos AIThe expert-validated move ontology provides a concrete implementation of the vertical-AI requirement to encode expert workflow, while preserving that experts—not self-grading—must define useful behavior.Metacognition in LLMs: Foundations, Progress, and OpportunitiesExposing therapeutic moves as callable actions offers an externally measurable control mechanism where the metacognition survey warns that internal self-inspection may be unreliable.MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMsMedPRESS evaluates whether safe behavior persists across user pressure, complementing this paper’s move-level alignment with a multi-turn test of whether steering remains reliable under interaction.
Context Map
agentresearch#tool-use#agent-evals#agent-reliability
Uncertainty
The evidence concerns psychotherapy, where behavioral alignment is safety-sensitive and human distributions are not automatically ideal outcomes. The material does not establish whether the method transfers to coding agents.