Agentic LLMs for Social Network Generation using Explainable Adversarial Re-prompting

High-quality, reliable social network data are essential for advancing research in social computing. However, the direct use of real-world social networks is often constrained by stringent privacy regulations and ethical concerns, rendering the development of Social Network Generation (SNG) techniques indispensable. While Large Language Models (LLMs) have driven advancements in social network simulation, a core challenge remains unsolved: the lack of comprehensive guidance for generating sufficiently realistic social networks. To address this, we introduce a paradigm shift by incorporating a discriminator based on social bot detection techniques for assessing network realism. Building upon this, we propose an innovative social network generation method that integrates agentic LLMs with eXplainable Adversarial Re-Prompting (X-ARP). The framework utilizes graph-prompted environment perception and decision-making, empowering agents to simulate autonomous behaviors, thereby constructing highly realistic social networks. Furthermore, X-ARP utilizes an explainable bot detector to evaluate network realism, incorporating a multi-level explainable model to extract multi-dimensional evidence from key nodes and features affecting realism. The extracted evidence is converted into explainable re-prompts and fed back into the generator, guiding the agent to refine behaviors deviating from real-world activities. In the ''Generate–Detect–Explain–Refine'' iterative adversarial loop, the generator and detector co-evolve, ensuring that the structural and semantic consistency of the generated network progressively aligns with that of real social networks. Extensive experiments demonstrate that X-ARP significantly outperforms existing methods, achieving only a 2.22% deviation in macro-structural metrics and improving semantic realism by 10.5%.