The Chameleon’s Limit: Investigating Persona Collapse and Homogenization in LLMs

Multi-agent LLM simulations require population diversity among agents. However, responses from LLMs with diverse human profiles cluster into classes, unlike human responses that uniformly fill the entire space. In this paper, we formalize this failure as Persona Collapse and quantify its population-level structure. We propose three diagnostic metrics: Coverage (how much of the persona space is occupied), Uniformity (how evenly agents spread across it), and Complexity (how rich the resulting behaviors are). Across ten LLMs on personality (BFI-44), moral reasoning, and self-introduction, collapse is both dimension- and domain-dependent: a model can look diverse on one axis yet be structurally degenerate on another, and can collapse the most in personality while being the most diverse in moral reasoning. Item-level diagnostics further show that variation tracks coarse demographic stereotypes rather than the fine-grained differences each persona specifies. Counter-intuitively, the models with the highest per-persona fidelity produce the most stereotyped populations. We will release our toolkit and data.