Large Language Models for Resource Block Assignment with QoS Constraints in OFDMA Multi-Cell Open RAN

Abstract

Traditional approaches to resource block allocation in cellular networks rely on model-based optimization or deep reinforcement learning (DRL). Model-driven techniques lead to inefficient solutions in terms of sum-rate performance under strong interference and optimize slot-by-slot without considering historical user QoS insights. Meanwhile, existing DRL resource allocation approaches struggle to generalize across diverse network setups and often overlook user rate guarantees, diminishing their practical viability. As deep learning gains prominence in next-generation wireless networks under the artificial intelligence (AI)-native vision of Open RAN, it is essential to assess the practical feasibility of both deep learning and emerging large language models (LLMs) for radio resource allocation. In this context, we address the resource block allocation problem in a multi-cell orthogonal frequency-division multiplexing (OFDM) setting using an LLM-based approach. We propose a scalable multi-agent framework featuring independent resource optimization by each agent or xAPP. The framework integrates LLMs enhanced through parameter-efficient finetuning to perform QoS-aware resource block scheduling. Results show that the proposed method outperforms both traditional optimization and DRL baselines, reducing user rate requirement violations by up to 21 percentage points. It achieves superior sum rate performance and generalizes well across different network conditions, unlike DRL solutions.