Large Language Models (LLMs) for Wireless Network Resource Management and Forecasting Applications

Abstract

Next-generation wireless networks require intelligent methods that can both optimize resource allocation and anticipate future network behaviour under rapidly changing conditions. Conventional optimization and deep learning approaches have demonstrated strong performance in specific wireless tasks, but they often rely on expensive retraining, large task-specific datasets, and fixed model assumptions that limit adaptability across new environments. This thesis investigates whether large language models (LLMs) can serve as training-free, prompt-engineered reasoning agents for wireless systems, with the goal of developing a unified methodology that adapts to diverse wireless problems without gradient-based retraining. The thesis first focuses on radio resource management (RRM), where constrained non-convex optimization problems arise in tasks such as beamforming, power control, and user association. To address these problems, this thesis proposes LLM4RRM (LLM for Radio Resource Management), a multi-agent LLM framework inspired by block coordinate descent, in which specialized solver agents optimize subsets of decision variables while an evaluator agent iteratively refines their outputs through structured feedback. An LLM-guided adaptive differentiable projection mechanism is introduced to improve feasibility under coupled constraints, together with a prompt-based uncertainty quantification framework for assessing solution reliability and consistency. The thesis then extends the same training-free agentic design philosophy to multivariate time-series forecasting through LLM4TSF (LLM for Time-Series Forecasting), where specialized agents model temporal components of correlated signals and an evaluator agent enforces statistical consistency across multiple series. Overall, the results demonstrate that prompt-engineered multi-agent LLM systems can provide adaptable, feasible, and scalable wireless intelligence without task-specific retraining.