Language-Grounded Co-evolution of Opinions and Ties: Paired Rewiring with Large Language Model Agents

Abstract

Modeling how user opinions and social ties co-evolve is central to understanding polarization and community formation in online platforms, yet existing approaches often sacrifice either linguistic realism or network structure. We propose a simulation framework for studying the co-evolution of online networks and opinions, in which large language models (LLMs) handle opinion updating, content generation, and network rewiring. This design integrates linguistic signals into structural change under real-world constraints, enabling systematic comparison across LLMs and empirical validation on three real-world datasets. Results show that LLM-based simulations consistently outperform equation-based baselines across key network metrics and better reproduce observed stance shifts. Analysis of the reasoning behind model decisions reveals systematic differences in rewiring motives and highlights distinct tendencies across models, offering interpretive insight into the network structures they produce. In addition, varying rewiring probabilities and strategies demonstrate how platform-like controls shape polarization and cohesion. Together, these findings establish a language-grounded, empirically validated approach to modeling network–opinion co-evolution with greater accuracy and interpretability.