Search papers, labs, and topics across Lattice.
This survey paper comprehensively reviews the application of Large Language Models (LLMs) to decision-making in wireless communication systems, addressing the challenges of accuracy, generalization, and collaboration in future 6G networks. It summarizes open-source communication datasets and LLM-based methods for data generation to tackle the lack of supervised annotations, and analyzes LLM architectures optimized for communication tasks. The paper structures the decision-making process into five key dimensions: prompt learning, chain of thought reasoning, inference mechanisms, decision frameworks, and multi-agent coordination, providing a task-adaptive pipeline.
LLMs can revolutionize 6G wireless communication decision-making, but only if we address the data scarcity problem and optimize reasoning pipelines for dynamic channel modeling and edge deployment.
Future 6G wireless systems will feature massive connectivity, complex tasks, limited resources, and heterogeneous architectures, posing significant challenges to decision-making regarding accuracy, generalization, and collaboration. Although large language models (LLMs) have demonstrated impressive capabilities in natural language processing and reasoning tasks, their application in wireless communication remains at a fragmented and exploratory stage. Given the potential of LLMs, this work provides a comprehensive overview of LLM-enabled decision-making. Firstly, in response to the lack of supervised annotations in communication scenarios, we summarize currently available open-source communication datasets and further explore LLM-based methods for data generation and augmentation, addressing a key gap in prior reviews on training data. We then analyze the evolution of LLM architectures for communication tasks, focusing on optimizations within the transformer architecture and emerging alternatives such as state space models and neuro-symbolic hybrid architectures. The discussion emphasizes their adaptability to challenges like dynamic channel modeling and edge deployment, providing a structural-level reference for selecting models in practical applications. Moreover, the paper provides a comprehensive analysis of the core techniques for decision-making. Unlike prior works that organize discussions around specific communication task types, we structure the reasoning process into five key dimensions: prompt learning, chain of thought reasoning, inference mechanisms, decision frameworks, and multi-agent coordination. This forms a task-adaptive decision-making pipeline aligned with the unique requirements of communication systems. Finally, we discuss deployment challenges and future directions, offering theoretical insights and practical guidance for the integration of LLMs into next-generation communication decision systems.
