From Representations to Behaviors: Exploring the Person-Situation-Behavior Triad in LLMs

This paper adapts Funder’s person-situation-behavior framework to analyze personality-like structure in large language models. It treats the person component as internal personality-related representations, situations as contexts that elicit trait-relevant responses, and behavior as response patterns in broader social tasks. Using contrastive behavior pairs and sparse autoencoder decomposition, the authors identify features associated with opposing trait poles. Feature-level interventions produce bidirectional shifts across diverse situations while preserving response validity. Social-intelligence evaluations show benefit-tradeoff patterns that the authors describe as consistent with human personality research. The abstract does not specify the evaluated models, datasets, traits, or effect sizes.