Large Language Model-Empowered Resource Allocation in Intent-Driven Wireless Networks

Abstract

Although large language models (LLMs) have the potential to automate resource allocation in wireless networks, generating high-precision allocation schemes remains non-trivial for cloud-based LLMs due to their inability of accurate numerical processes. The heavy re-training overhead of LLMs for handling the dynamic features of wireless networks also poses a challenge. This paper presents a new LLM-empowered framework for intent-driven wireless resource allocation. The framework features a dual-expert LLM architecture, consisting of an instruction expert for generating resource allocation instructions and a policy expert for obtaining numerically-precise policies. Both experts encompass lightweight adapters and collaborate with the locally deployed LLM for resource allocation. Decoupling the generation of instruction and policy benefits efficient adaptation and generalization under diverse task requests and dynamic network states. We further propose a new two-stage fine-tuning protocol to enhance the adaptability of LLM-empowered resource allocation through distillation and reinforcement. Experiments on two major application cases demonstrate the superior adaptability and generalization abilities of the proposed LLM-empowered resource allocation framework with two-stage fine-tuning, compared to existing schemes.