Adapting Large Language Models for Cognitive Spectrum Allocation in Coexisting GEO-LEO Satellites

Cognitive Satellite Radio (CSR), the integration of Cognitive Radio (CR) into satellite networks addressing the growing demand for efficient spectrum use, necessitates intelligent spectrum allocation. However, the dynamic nature of satellite environments poses significant challenges for real-time spectrum management. This work investigates the integration of Large Language Models (LLMs) to CSR, leveraging their reasoning and generalisation capabilities. We propose a compact LLM-based framework for Concurrent Spectral Access (CSA), comprising a state encoder, a Low-Rank Adaptation (LoRA) reinforcement learning pipeline, and a task-specific language model head. Using simulated CSR scenarios, we evaluate the proposed framework in five LLMs and benchmark against a Deep Q-Network (DQN)-based Dynamic Spectrum Allocation (DSA). The results indicate that LLMs offer competitive performance and scalability, suggesting their promise in enabling adaptive, efficient CSR networks.