LLM personas can mimic human survey responses for stable attributes

A new arXiv paper explores the effectiveness of using large-language models (LLMs) to create digital personas that can substitute for human survey respondents. Researchers found that while these personas can improve alignment with human response distributions, particularly for questions related to stable attributes and values, they struggle with individual predictions and capturing complex respondent structures. The study suggests that retrieval-augmented architectures show the most promise, but performance is highly dependent on the inherent variability of the survey questions and the underlying structure of human responses, rather than solely on the LLM used. AI

IMPACT Digital personas may offer a cost-effective alternative for certain types of survey research, but human validation remains crucial for subjective or complex data.

RANK_REASON Academic paper on LLM capabilities [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Hugging Face
• large-language models
• LISS panel
• Mumin Jia

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →