Cau et al. trace social simulation evolution to LLM-powered digital twins

The paper, published as a chapter for the Encyclopedia of Social Network Analysis and Mining, systematically maps the progression of social simulation methodologies. It begins with traditional agent-based models (ABMs), where individual agents follow predefined behavioral rules to explore emergent social phenomena. The authors then transition to a new wave of simulations that integrate large language models (LLMs), enabling agents to exhibit more nuanced, context-aware behaviors without rigid programming. Finally, they introduce Social Digital Twins—high-fidelity, data-driven digital replicas of entire social systems that can be updated in real time and used for experimentation, prediction, and policy testing.

The work highlights a critical shift: while early ABMs focused on abstract, generalizable mechanisms (e.g., segregation, opinion dynamics), newer paradigms strive for realism by embedding real-world data and AI-driven agent cognition. The authors critically assess each approach—ABMs offer simplicity and interpretability, LLM-enhanced simulations bring behavioral richness but raise computational and ethical concerns, and Social Digital Twins promise unprecedented accuracy at the cost of complexity and data dependency. This chapter serves as a timely reference for researchers and practitioners aiming to understand the frontier of social simulation.