Can LLM Agents Assist Dynamic Network Simulation? A Case Study on Email Networks and Phishing Synthesis

Siqi Miao, Yuhong Luo, Ziyang Chen, Hans Hao-Hsun Hsu, Mufei Li, Kaiqing Zhang, Pan Li

Keywords: Large Language Models (LLMs), Multi-Agent Systems, Dynamic Network Simulation

Abstract:

Simulating dynamic networks is crucial for understanding complex systems and for applications ranging from policy evaluation to data synthesis. However, traditional rule-based models often fail to capture micro-level patterns, while deep learning approaches are typically designed for single-step prediction rather than open-loop long-horizon generation. Although recent large-language-model (LLM)-based agent systems excel at simulating plausible micro-level behaviors in domains like task-solving and gaming, their capacity to generate dynamic networks faithful to real-world data remains underexplored. This work investigates the potential of LLM agents as high-fidelity, data-driven dynamic network simulators. As a proof of concept, we conduct our study on two public email networks (Enron and IETF). We propose an evaluation framework that assesses simulation fidelity across micro-, meso-, and macro-level structural and temporal dynamics. A comprehensive benchmark comparing LLM agents against classical point-process models and dynamic Graph Neural Networks reveals that LLM agents excel at generating plausible local interactions but struggle with preserving global structure, a limitation that we show can be mitigated by using Hawkes processes for guidance. We have studied long-horizon generation robustness and
demonstrated our framework's utility in a case study synthesizing realistic phishing attacks. Our results highlight a path toward high-fidelity dynamic network simulation with LLM agents for critical downstream applications. Our code is available at \url{https://anonymous.4open.science/r/DNSL-DE37}.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 23356