Mapping LLM Personality Traits for Control and Safety

Large Language Models (LLMs) often exhibit consistent behavioral patterns, or "personas," which influence their generalization and safety. However, tools to effectively analyze and control these personas have been lacking. Researchers introduced "Persona Cartography," a novel approach that conceptualizes LLM personas as positions within a space of behavioral traits, utilizing the well-established OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) to describe these traits. The study involved training low-rank adapters to either amplify or suppress specific personality traits across six different LLMs. Evaluations, using LLM-judges, human-validated panels, and trait-specific benchmarks, confirmed that these adapters could largely monotonically adjust target traits, combine additively for mixed personas, and maintain core model performance at moderate scales. Crucially, the induced trait axes were shown to influence safety-relevant behaviors, such as frustration and sycophancy, providing a direct link between personality control and model safety. An unsupervised psychometric pipeline also recovered interpretable behavioral factors, bridging personality measurement, model editing, and safety considerations.