Fine-Tuning Large Language Models for Optimal Resource Management in D2D Wireless Networks
Tayyib Ul Hassan, Aamina Binte Khurram, Attiya Waqar, Arsalan Ahmad, Syed Ali Hassan, Haejoon Jung · 2025
With the advent of sixth-generation (6 G) networks, efficient management of bandwidth and transmit power has become an increasingly important concern. With increasing network complexity and uncertainty, traditional optimization methods face multiple challenges. Large language models (LLMs) offer a flexible alternative to these methods due to their adaptability across varied conditions. In this study, we formulate a resource allocation problem involving multiple device-to-device (D2D) pairs and develop an LLM-based approach to maximize either energy efficiency (EE) or spectral efficiency (SE). We evaluate three LLM adaptation methods, namely fine-tuning, retrieval-augmented generation (RAG), and few-shot prompting, and identify fine-tuning as the most effective for the study under consideration. Finetuning achieves 94.24 % of optimal SE and 86.51 % of optimal EE, outperforming RAG and few-shot prompting in terms of performance. To further test its robustness, we examine finetuning under varied path loss distributions and increased system complexity, and observe how the LLM's predictions respond to these conditions. Additionally, fine-tuned models like Phi-3 Mini achieve inference times of less than one second (0.66s) and have a considerable time complexity advantage over exhaustive search, RAG, and few-shot prompting.