New study reveals LLM personality is frame-dependent, not fixed

A new arXiv paper by Yuan Yuan (stat.ML, July 2026) investigates the dual nature of LLM personas, challenging the assumption that psychometric evaluations capture stable traits. The study uses GPT-4o simulating American and Chinese-American personas on the IPIP-50 questionnaire, manipulating question order to test whether within-instance correlation structure (geometric features on SPD manifolds) is intrinsic or frame-dependent. Key findings: aggregated Big Five scores show a 21% drop under randomization but remain relatively robust to frame changes, while geometric features collapse by 42% under frame misalignment yet recover substantially to 84% under shared frames—outperforming aggregated features (76% recovery). This collapse-recovery pattern suggests persona geometry is not a fixed property but a frame-dependent coordination pattern encoding information invisible to simple aggregation.

The paper establishes a dual-nature framework: frame-robust aggregated tendencies versus frame-dependent geometric structure. This necessitates frame-aware evaluation methods for LLM personas and challenges static trait conceptions in AI personality research. The findings have practical implications for developers using LLMs in role-specific applications (e.g., customer service, therapy bots), where consistent persona expression may require careful attention to prompt framing and question ordering to avoid unintended behavioral collapses. The study also opens questions about whether similar effects occur in other LLMs and whether frame-independent personas can be engineered.