LLM opinion diversity: more persona detail doesn't help, says study

A new paper by Qiyang Yao (arXiv:2607.20429) systematically investigates what actually drives opinion diversity in large language models. The study used a factorial experiment across 7 models and 100 real-user open-ended questions, separating two intervention dimensions: input conditioning (persona depth) and interaction architecture. The results challenge several common assumptions held by practitioners using LLMs for synthetic surveys and focus group modeling.

First, more persona detail does not monotonically increase diversity. The initial step of persona conditioning already captures the majority of the gain, and further elaboration with demographic details did not consistently improve diversity—on some models it even reduced it. Second, different interaction architectures (e.g., chain-of-thought, multi-agent debate) explore largely non-overlapping opinion regions. Combining multiple architectures yielded broader coverage than optimizing any single one. Third, low-cost alternatives like raising sampling temperature or adding diversity instructions had negligible effect compared to structured interventions. The work demonstrates that diversity is not a product of scaling any single dimension, but is highly sensitive to structural form and combination of interventions.