Study reveals distinct behavioral fingerprints in large language models using psycholinguistic analysis

A new research paper published on arXiv (arXiv:2604.16755v2) demonstrates that large language models exhibit measurable individuality in their responses. Using crossed random-effects models—a statistical method common in psychometrics—the study analyzed 74.9 million ratings from 10 open-weight LLMs across 14 psycholinguistic norms. These models evaluated over 100,000 words to distinguish between genuine behavioral dispositions and random noise. On average, 16.9% of response variance was attributed to stimulus-specific individuality, significantly exceeding statistical null expectations. Cross-norm prediction analyses confirmed that each model possesses a unique, coherent fingerprint. This research addresses a critical gap in understanding LLM behavior as these systems become integrated into daily life for roles ranging from decision support to companionship. The findings suggest that behavioral differences among models reflect stable, identifiable characteristics rather than mere response biases or stochastic variation. The study’s methodology applies psychometric inventories and cognitive paradigms previously used to profile LLM dispositions, but with enhanced precision for separating systematic effects. This work contributes to the growing literature on AI behavioral analysis, providing a framework for assessing model individuality in practical applications.