Intent-Driven LLM-Based Two-Level Control in AI-Native 6G Networks

We propose an intent-driven two-level communication framework for AI-native 6G networks. The upper level employs a Large Language Model (LLM) to interpret highlevel application intent and adapt control parameters based on aggregated system feedback. The lower level executes slotlevel resource allocation decisions using any mathematically grounded online controller. The two levels interact through a structured feedback loop that enables continuous alignment between application intent and network operation under dynamic conditions. As a representative case study, we adopt a Lyapunovbased Decision Support Framework (DSF) as the lower level controller for a smart city video upload scenario with dynamic leasing of Reconfigurable Intelligent Surface (RIS) and spectrum resources. The LLM updates the Lyapunov control parameters at each adaptation window through structured prompts that encode application intent and aggregated system feedback. Simulation results indicate that LLM-driven parameter refe4inement can improve the cost-progress tradeoff over rule-based and fixedparameter baselines under time-varying network conditions.