Xingqi Wu (University of Michigan-Dearborn), Junaid Farooq (University of Michigan-Dearborn), Yuhui Wang (University of Michigan-Dearborn), Juntao Chen (Fordham University)
The decentralized and modular architecture of open radio access networks (O-RAN) enhances flexibility and interoperability but introduces significant challenges in efficiently managing resource allocation. The disaggregation of network functions across distributed unit, centralized unit, and RAN intelligent controller (RIC) creates complexities in coordinating resources across multiple network slices, each with distinct and dynamic quality of service (QoS) requirements. Traditional machine learning (ML) approaches for resource management often rely on extensive offline training, which is impractical in the highly variable and real-time environments of O-RAN systems. This paper presents LLM-xApp, a novel large language model (LLM)-powered xApp framework for adaptive radio resource management in O-RAN systems. The proposed framework is based on intelligently prompting LLM agents to dynamically optimize resource allocation to different network slices. Experimental evaluations are conducted on the OpenAI Cellular (OAIC) platform showcasing significant improvements in average data rates as well as the reliability of the slices, demonstrating the potential of LLMs to enhance real-time decision-making in next-generation wireless networks.