Representational Idiosyncrasy of Psychopathology Across Text Embedding Models: Implications and Practical Suggestions for Clinical Psychological Science

Le résumé fourni par la source

The integration of large language models (LLMs) into clinical psychological science has outpaced our understanding of how psychopathology is represented across the diverse model landscape. We analyzed how 21 text embedding models represented 405 items from the HiTOP-SR, a non-public psychopathology measure absent from training data. Exploratory factor analysis of item-level representations at global and domain-specific levels revealed variable structural consistency across models. Globally, structural congruence frequently fell below meaningful thresholds (CC < .85), although broad two-factor solutions showed strong cross-model convergence. Domain-specific congruence varied considerably, with some domains (e.g., Anankastia) showing strong convergence and others (e.g., Antagonism) showing weak convergence. These findings demonstrate that text embedding models may lack a shared representational framework for psychopathology, yielding structures that are idiosyncratic and sensitive to model-specific features and methodological choices. These inconsistencies underscore the need for careful model selection and rigorous evaluation of LLM-based tools.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.