Survey-based Large Language Model agent: probing the gender attitude and change

• In Event: 40221 - Section on Mathematical Sociology Flash Talks
  In Refereed Roundtable (90 minute): Table 8: Survey-based Large Language Model agent: probing the gender attitude and change

Abstract

Large language models (LLMs) have demonstrated the ability to mimic human attitudes. However, prior methods primarily capture subgroup stereotypes, lacking the granularity needed to reflect individual-level characteristics. In this study, we integrate survey data with large language models to construct LLM agents, or "AI twins" that correspond to individual survey respondents. We evaluate the performance of these agents by assessing their ability to replicate human attitudes and predict attitude shifts in response to specific social changes. Using the reversal of the gender gap in education as a case study, our findings demonstrate that these survey-based LLM agents effectively predict human’s gender attitudes and capture shifts in attitudes induced by social transformation. This approach offers valuable insights into social attitude and lays the groundwork for LLM-based social simulations.

Author

• Mengzhen Jia, Sun Yat-Sen University