LLM Agents Fail to Simulate Realistic Network Dynamics
A new study from arXiv (2605.12507) evaluates whether Large Language Model multi-agent systems can simulate realistic dynamic networks, using email networks with phishing synthesis as a case study. The authors find that existing frameworks generate plausible micro-level interactions but fail to capture emergent macroscopic topologies essential for modeling information propagation and cybersecurity threats. To address this, they propose two extensions: data-driven event triggers for sustained long-horizon interactions and Hawkes processes for temporal activation dynamics. The approach preserves macroscopic network fidelity without compromising micro-level realism.
Key facts
- arXiv paper 2605.12507 evaluates LLM multi-agent systems for dynamic network simulation
- Existing frameworks fail to capture emergent macroscopic topologies
- Two extensions proposed: data-driven event triggers and Hawkes processes
- Focus on email networks and phishing synthesis as case study
- Goal is to improve simulation of information propagation and cybersecurity threats
- Approach preserves macroscopic network fidelity
- Published on arXiv with cross announcement type
- Study explores whether LLM agents can replicate realistic structural and temporal dynamics
Entities
Institutions
- arXiv