Preference-Optimized LLM Counselors Trade Goal Persistence for Relational Attunement in Motivational Interviewing

A recent paper on arXiv (2607.28814) presents a dual-axis evaluation framework for assessing counselor responses in Motivational Interviewing (MI), utilizing the Motivational Interviewing Treatment Integrity (MITI) code. This framework features two axes: Goal Persistence (GP) and Relational Attunement (RA), creating a four-quadrant model where 'rolling with resistance' is characterized by high scores on both axes. The research examines whether preference optimization—penalizing specific failures like capitulation or confrontation—teaches LLM counselors to effectively roll with resistance or inadvertently leads to the opposite outcome. By employing the expert-annotated AnnoMI corpus, the authors generate topic-disjoint Direct Preference Optimization (DPO) data. An automatic evaluator, aligned with expert annotations, assesses the models. Results indicate that preference-optimized LLM counselors might sacrifice goal persistence for relational attunement, risking capitulation when trained to avoid confrontation, and vice versa. The study emphasizes the difficulty of maintaining balance between these dimensions in automated counseling and the necessity for sophisticated training data and evaluation techniques.