LLM persona expression comprises frame-robust aggregates and frame-dependent geometry

A study analyzing GPT-4o simulating American and Chinese-American personas reveals that LLM persona expression consists of two dissociable components: aggregated features and geometric features. Researchers constructed within-instance correlation matrices from IPIP-50 responses to analyze geometry on SPD manifolds under manipulated question orderings.

• Aggregated features, represented by Big Five scores, degrade by 21% under randomization but remain frame-robust.
• Geometric features collapse by 42% under frame misalignment but recover substantially to 84% under shared frames, surpassing aggregated features which reach only 76%.

The findings establish a dual-nature framework for LLM personas, challenging static trait conceptions and necessitating frame-aware evaluation methods.