MAPR of GPT compared with optimal algorithm.

Large language models have revolutionized the field of natural language processing and are now becoming a one-stop solution to various tasks. In the field of Networking, LLMs can also play a major role when it comes to resource optimization and sharing. While Sumrate maximization has been a crucial factor for resource optimization in the networking domain, the optimal or sub-optimal algorithms it requires can be cumbersome to comprehend and implement. An effective solution is leveraging the generative power of LLMs for such tasks where there is no necessity for prior algorithmic and programming knowledge. A zero-shot analysis of these models is necessary to define the feasibility of using them in such tasks. Using different combinations of total cellular users and total D2D pairs, our empirical results suggest that the maximum average efficiency of these models for sumrate maximization in comparison to state-of-the-art approaches is around 58%, which is obtained using GPT. The experiment also concludes that some variants of the large language models currently in use are not suitable for numerical and structural data without fine-tuning their parameters.大语言模型（Large Language Model, LLM）已彻底革新自然语言处理领域，如今正成为各类任务的一站式解决方案。在网络领域中，大语言模型在资源优化与共享方面同样可发挥重要作用。尽管和速率最大化（Sumrate Maximization）一直是网络领域资源优化的核心考量因素，但其所需的最优或次优算法往往理解与实现难度较大。针对这类无需预先掌握算法与编程知识的任务，借助大语言模型的生成能力是一种高效解决方案。为明确此类任务中使用这些模型的可行性，有必要对其开展零样本（Zero-shot）分析。通过调整蜂窝用户总数与设备到设备（Device-to-Device, D2D）对总数的不同组合，我们的实证结果显示：相较于现有最优方法，这些大语言模型用于和速率最大化任务时的最高平均效率约为58%，该结果由GPT实现。本实验同时得出结论：当前使用的部分大语言模型变体，若不对其参数进行微调，便不适用于数值与结构化数据。