From Rule-Based to LLM-Based Agents: A Calibrated Simulation Framework for Classroom Social Networks

Can large language model (LLM)–based agents reproduce the formation of real-world classroom social networks? While recent studies have explored LLMs as flexible decision-making agents, their ability to generate socially meaningful network structures remains unclear.In this study, we propose a controlled simulation framework that compares calibrated rule-based agents and LLM-based agents under identical contextual conditions. Using empirical data from high school classrooms, we first calibrate a rule-based interaction model to reproduce observed network properties, including density, modularity, community structure, reciprocity, and personality–degree correlations. We then replace the rule-based decision mechanism with an LLM-based agent without parameter tuning, enabling a direct comparison of emergent network structures.Results show that LLM-based agents can reproduce coarse-grained interaction patterns, such as edge density and overall connectivity, but fail to consistently recover higher-order structural properties, including modularity, community organization, and relational patterns. These findings suggest that while LLMs capture locally plausible interaction tendencies, they do not reliably encode the structural constraints governing real-world social networks.This work contributes a methodological framework for evaluating LLM-based agents in educational social systems and provides empirical evidence on their limitations in simulating socially grounded network formation processes.