Cyber ranges require realistic background user activsity to support effective defensive training and experimentation. Traditional script-based user simulations are deterministic and easily detectable, while fully autonomous LLM-driven agents introduce significant computational overhead and scalability limitations.This paper proposes a hybrid user-simulation framework that combines probabilistic behavioral scheduling with event-driven LLM reasoning. User activity is modeled using Discrete-Time Markov Chains (DTMC), which coordinate high-level behavioral states while delegating contextual interaction generation to a local language model only when necessary.Analytical evaluation shows that the proposed architecture reduces LLM inference frequency from continuous per-timestep invocation to event-triggered execution only. Under a representative interaction probability of 5%, this corresponds to an approximate 95% reduction in LLM inference calls compared to fully LLM-driven agents, while preserving plausible behavioral variability. The decentralized design supports scalable deployment across large simulated endpoint populations with reduced computational and communication overhead.
- Címlap
- Publikációk
- A Hybrid Markov–LLM Architecture for Scalable User Simulation in Cyber Ranges