Flattening Sociological Reasoning in LLMs in Life Outcome Inference

• In Event: Open Refereed Roundtables
  In Refereed Roundtable (90 minute): Table 2: Quantitative and Computational Methods

Abstract

Sociologists increasingly employ large language models (LLMs) to simulate human subjects and predict life outcomes, but it remains unclear whether LLM reasoning reflects sociological reasoning, which integrates multilevel factors and temporal dynamics. Using three reasoning LLMs (DeepSeek-R1, QwQ-32B, GPT-OSS-120B) on 1,949 cases from the Fragile Families and Child Wellbeing Study, we analyzed predictions of age-22 educational attainment and aspirations.
We uncover a homogenization bias: LLMs compress outcome variation toward modal categories, underestimating extremes and intergenerational mobility. Trace analysis identifies four mechanisms driving this compression: cognitive narrowing (reducing factor diversity), semantic proximity bias (favoring outcome-adjacent variables), temporal recency bias (overweighting recent data), and risk aversion (favoring moderate predictions). Collectively, these tendencies flatten multivariate and temporal complexity, shifting predictions away from heterogeneity. Non-reasoning models display similar patterns, suggesting this bias may be architecture-wide.
Our results indicate that LLMs follow specific types of inferential pathways that surprisingly simplify reasoning rather than accounting for the complexity of the life course, underscoring the need to audit inferential dynamics beyond accuracy when integrating LLMs into social science research.

Authors

• Ying Chu, National Taiwan University

• Chen-Shuo Hong, National Taiwan University