Abstract: With the rapid advancement of large language models and generative agent technologies, social simulation is shifting from traditional agent-based modeling driven by hand-crafted rules to a new paradigm characterized by semantic generation, contextual interaction, and experiment orchestration. Compared with classical approaches, LLMs-based agents exhibit clear advantages in language expression, contextual understanding, role-playing, and complex behavior generation. These capabilities enable large-scale social simulators to achieve higher behavioral realism, richer interaction mechanisms, and broader application potential. The significance of large-scale social simulators is therefore no longer limited to “constructing a runnable virtual society”, but lies in establishing a new type of experimental platform and software system for complex social systems that is developable, verifiable, and deployable. This article analyzes this transformation from five perspectives: technological evolution, system development, trustworthy validation, application and governance, and future directions. Using AgentSociety as a representative case, this article further discusses its potential value in policy pre-evaluation, opinion dynamics forecasting, risk governance, urban governance, and the study of complex social systems. This article argues that LLMs-driven large-scale social simulators are transforming social simulation from a research tool into application-level infrastructure, and are poised to become a key technological foundation bridging artificial intelligence, social science, and major societal needs.