Back to feed
arXiv cs.CL
arXiv cs.CL
7/24/2026
More Is Not More: What Matters for Diversity in LLM Opinions?

More Is Not More: What Matters for Diversity in LLM Opinions?

Short summary

This study uses a factorial experiment to disentangle what drives diversity in LLM opinions across 100 real-user questions and 7 models. Key findings: initial persona conditioning captures most diversity gains while further demographic detail can reduce diversity; different interaction architectures explore non-overlapping opinion regions so combining them yields broader coverage; and low-cost tricks like raising temperature or adding diversity instructions produce negligible effects compared to structured interventions.

  • •Initial persona conditioning captures most diversity gains; more detail doesn't help and can hurt
  • •Combining multiple interaction architectures covers more opinion space than optimizing one
  • •Raising sampling temperature and adding diversity instructions have negligible effect

Generated with AI, which can make mistakes.

#research-breakthrough#ai-tools
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Stochastic Sampling is Epistemically Shallow: The Dimensionality Gap Between Temperature Variation and Model Diversity in LLMs
arXiv CS.AI
Prompt Robustness Is Task-Dependent: Comparing Objective and Belief-Style Questions in LLM Evaluation
arXiv cs.CL
When Reasoning Narrows the Move: Diversity Collapse in LLM Game Play
arXiv cs.CL
What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
arXiv cs.CL