Influence in Motion: Tracing Persuasive Dynamics via Multi‐Agent Networks

A bottom–up, reproducible multi-agent simulation framework is presented for investigating opinion dynamics via LLM-based agents embedded in an endogenously evolving small-world network. The study simulates cognitive–affective agents deliberating the real-world controversy over rounds, generating natural-language exchanges, instantaneous appraisals, and systemic reflections. Network topology co-evolves through agent-driven edge rewiring. Community detection and intra- community stance-variance analyses reveal three temporal phases—turbulence, coalescence, consolidation—with opinion variance declining. Correlations between closeness centrality and Influence Score uncover heterogeneous influence patterns, exemplified by positively, negatively, and negligibly correlated agent archetypes. These results demonstrate that LLM-based generative agents can (1) reproduce key opinion and network structural dynamics (H1), (2) self-organize into stable communities with constructive opinion aggregation via endogenous rewiring (H2), and (3) convert network structure into persuasive influence through adaptive discourse strategies (H3). This framework bridges micro-level cognitive–affective processes and macro-level network phenomena, offering a versatile platform for computational communication research.