LLM-Empowered Resource Allocation in Wireless Communications Systems

Authors:
Top Cited Papers
Open Access
Publisher website
Google Scholar
Abstract
The recent success of large language models (LLMs) has spurred their application in various fields. In particular, there have been growing efforts to integrate LLMs into wireless communications systems. Leveraging LLMs in this domain offers the potential to realize artificial general intelligence (AGI)-enabled wireless networks. In this paper, we investigate the feasibility and characteristics of an LLM-based resource allocation framework. To fundamentally assess the reasoning capability of LLMs, we formulate a representative resource allocation problem involving two transmit pairs, aiming to maximize either energy efficiency (EE) or spectral efficiency (SE). We propose a few-shot learning-based approach that prompts the LLM with channel conditions to infer resource allocation decisions, thereby eliminating the need for explicit model training. To address potential reliability concerns, we further introduce a hybrid strategy that combines the LLM’s output with a conventional low-complexity method, serving as a robust fallback. Through simulations, we demonstrate that the proposed framework achieves up to 96% of the optimal performance in EE maximization. Our performance evaluation also yields two key insights. First, specialized models such as CodeLLaMA demonstrate superior performance, likely due to the structural resemblance between programming tasks and resource allocation problems. Second, the proposed framework achieves more substantial gains in EE optimization, which can be viewed as a continuous approximation problem, compared to SE optimization, which is inherently more discrete and binary in nature. After confirming the feasibility of LLM-based resource allocation through extensive simulations, we address key technical challenges, such as latency, optimal architecture, explainability, and training methodology, that must be overcome for practical deployment, paving the way for the next generation of intelligent wireless systems. Furthermore, the proposed approach is adaptable to a wide range of optimization problems beyond wireless communications, making it applicable to various domains that require intelligent decision-making.

This publication has 41 references indexed in Scilit: