Large language models are increasingly used to stand in for humans in multi-agent simulations, such as generating survey respondents or modeling application users. But a new preprint, 'The Chameleon's Limit: Investigating Persona Collapse and Homogenization in Large Language Models,' argues that these models fall short in a fundamental way: they lack the population-level diversity that such simulations depend on.

The paper's title introduces the twin problems of 'persona collapse' and homogenization. The abstract's truncated opening hints that responses from LLMs — presumably under different assigned personas — fail to show the variation expected from a real population. The chameleon metaphor suggests that even when models try to adapt to a role, they end up blending into a homogeneous default.

While the full details of the study are not available in the abstract, the framing points to a practical limitation: simulations built on LLMs may systematically underrepresent the range of human perspectives. The paper's title suggests a hard limit to how well these models can mimic a diverse set of individuals, even when explicitly prompted to adopt different roles.