Large AI Models for Wireless Physical Layer

TL;DR

Large AI models offer robust generalization and multimodal processing that can transform the wireless physical layer. The paper surveys two strategic families—pre-trained LAMs adapted from NLP/CV domains and wireless-native LAMs trained directly on wireless data—through a unified framework, representative use cases, and complexity analyses. It highlights practical considerations, including latency, hardware constraints, and the need for standardized datasets, while outlining future directions such as efficient architectures, interpretability, and large-small model collaboration. The findings suggest that LAMs can significantly enhance PHY tasks like channel prediction, CSI feedback, and beam prediction, enabling more adaptable and scalable next-generation wireless systems.

Abstract

Large artificial intelligence models (LAMs) are transforming wireless physical layer technologies through their robust generalization, multitask processing, and multimodal capabilities. This article reviews recent advancements in applying LAMs to physical layer communications, addressing obstacles of conventional AI-based approaches. LAM-based solutions are classified into two strategies: leveraging pre-trained LAMs and developing native LAMs designed specifically for physical layer tasks. The motivations and key frameworks of these approaches are comprehensively examined through multiple use cases. Both strategies significantly improve performance and adaptability across diverse wireless scenarios. Future research directions, including efficient architectures, interpretability, standardized datasets, and collaboration between large and small models, are proposed to advance LAM-based physical layer solutions for next-generation communication systems.

Figures (4)

• Figure 1: General framework of pre-trained LAMs for physical layer. The framework adapts LAMs using a preprocessing module to align wireless data with input space and output layers to match physical-layer tasks. Depending on the task, LAM parameters are selectively fine-tuned liu_llm4cp_2024 or frozen guo_lvm4csi_2025.
• Figure 2: Detailed architecture of the LLM4CP framework proposed in liu_llm4cp_2024.
• Figure 3: Main framework of the pre-trained LLMs for multimodal sensing-empowered beam prediction, i.e., M$^2$BeamLLM zheng_m2beamllm_2025, comprising three modules, akin to the general framework illustrated in Fig. \ref{['LLM4CP']}.
• Figure 4: A general framework of wireless physical layer-native LAMs. This framework trains an LAM from scratch for a specific guo_prompt2025catak_bert4mimo_2025 or multiple physical tasks alikhani_large_2025yang_wirelessgpt_2025.