LLMs Struggle to Capture Human Personality Diversity, Study Finds

New research indicates a significant limitation in how large language models (LLMs) simulate human personalities, a phenomenon termed "persona manifold collapse." This occurs when attempts to provide more detailed or expressive persona specifications to an LLM lead to a reduction in the model's ability to represent diverse behaviors and latent personality traits. Instead of improving fidelity, richer descriptions can cause the model's internal representations to converge, diminishing the distinctiveness between different simulated personas. The study systematically evaluated this effect across various LLM architectures and scales, finding that increased persona complexity consistently reduced the separation between personas in the model's latent space. This also translated to weaker differentiation in downstream simulation tasks, meaning the models struggled to accurately reproduce human subgroup disagreements. Surprisingly, simple personas defined by age and gender often yielded more accurate predictions in simulation tasks compared to highly detailed "Ideal Customer Profiles." The findings suggest that simply adding more descriptive detail to personas does not guarantee improved simulation fidelity and can even degrade performance. The research emphasizes the need for a more nuanced, representation-aware approach to persona construction, rather than just increasing expressivity, to overcome these inherent limitations in LLM-based human simulation.