Recently, agentic AI systems based on large language models (LLMs) have emerged as a promising paradigm for autonomous decision-making across various domains. In wireless communications, agentic AI can serve as an intelligent network orchestrator, which dynamically manages communication resources and adapts to changing conditions. Effective operation of agentic AI requires seamless integration of multimodal sensing data from distributed sensors, which poses significant challenges in communication overhead. In this paper, we propose multimodal agentic wireless network (MAWN), a cooperative agentic AI framework that intelligently selects sensing modalities and autonomously plans computational workflows to address various wireless tasks (e.g., beam management). MAWN consists of two specialized LLM agents: the sensor manager agent (SMA) and the function manager agent (FMA). The SMA minimizes data transmission overhead by selecting only necessary sensors while considering data availability, while the FMA dynamically composes optimal tool execution sequences based on task requirements. By combining two LLM agents, MAWN adaptively gathers information and handles requests with reduced communication overhead and no manual configuration. Simulation results demonstrate that MAWN achieves substantial performance improvements across diverse communication tasks.

Beam management is an essential operation of next-generation (xG) wireless networks to compensate severe signal attenuation and ensure reliable communications over millimeter wave (mmWave) and terahertz (THz) bands. The primary objective of beam management is to determine beam directions that are properly aligned with the signal propagation paths. Conventional beam management techniques typically rely on geometric channel parameters such as angles, delays, and path gains. However, due to their lack of contextual awareness of the surrounding environment, these techniques fall short in handling the piecewise continuous changes in the geometric channel parameters caused by sudden obstructions or variations in scatterers. In this paper, we propose a novel beam management framework that utilizes environmental information extracted from sensor data (e.g., images) and pilot measurements to optimize beam directions. The main idea of the proposed scheme is to utilize visual channel parameters, including the positions of user equipment (UE), reflection points, and scatterers, which provide a direct visualization of the propagation environment. By tracking the visual channel parameters, dynamic changes in scatterers and propagation paths can be monitored over time. To analyze multimodal data and track these parameters, we exploit large multimodal model (LMM), a generative artificial intelligence (AI) model specialized in extracting correlated features across multimodal data and generating subsequent data. Simulation results show that the proposed scheme can accurately track the visual channel parameters and enhance data rate.

Recently, large language models (LLMs) and large multimodal models (LMMs) have been successfully adopted in various fields due to their outstanding adaptability. One application of LMM in wireless communication is mobility management in ultra-dense network (UDN) systems. By exploiting its powerful contextual understanding capability, LMM can obtain abundant environmental awareness and effectively manage user equipment (UE) mobility. In this paper, we propose an LMM-aided geometry-aware mobility management (LMM-GMM) that estimates the long-term future data rates of the UE and makes a handover decision to maximize the data rates. LMM-GMM utilizes a geographic information system (GIS) map image in LMM, thereby identifying the UE position and corresponding propagation paths. Using this information, LMM-GMM can accurately predict the future data rate in dynamic scenarios and enhance the handover performance. Simulation results demonstrate that the proposed LMM-GMM can effectively perform mobility management in the UDN environments.

Recently, large multimodal models (LMMs) have been successfully adopted in various fields due to their outstanding adaptability and reasoning abilities. Despite their potential to automate diverse tasks in communications systems, application to the physical layer remains underexplored. The primary reason is that the traditional physical layer relies on analytic channel measurements (e.g., pilot measurements), which capture only quantitative changes in transmitted signals and fail to characterize qualitative physical interactions (e.g., reflections and blockages) with the environment. In this paper, we propose an LMM-based environment-aware channel estimation framework that captures the contextual channel information by leveraging both perceptual sensor data and numerical pilot measurements. The main idea of the proposed scheme is to utilize visual channel parameters (VCPs), i.e., positions of user equipment (UE), reflection points, and scatterers. Since VCPs provide a direct visualization of the propagation environment, we can identify how signals propagate and physically interacts with the surrounding objects. To extract the VCPs and learn their probability distributions, we develop a reflection learning technique based on LMM. By incorporating these parameters, we establish a fundamental channel knowledge map (CKM) between the UE position and the channel. Simulation results demonstrate that the proposed scheme can effectively predict the channel throughout the wireless environments.

Millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems have been considered a promising solution to provide high-quality data services. To achieve full beamforming gain, the acquisition of accurate channel information is crucial for the success of massive MIMO systems. In recent years, numerous approaches have been suggested to acquire the downlink channel state information (CSI). However, a significant mismatch between the estimated channel and the actual channel for data transmission causes outdated CSI, leading to a severe degradation in spectral efficiency. In this paper, we propose a channel estimation technique for mmWave massive MIMO systems that obtains multipath components of the downlink channel from the previous channel sequence. To be specific, the proposed technique learns the spatio-temporal correlation be-tween multipath components by exploiting a Transformer-based framework. From the numerical results, we demonstrate that the proposed technique outperforms the conventional channel acquisition techniques in terms of normalized mean square error (NMSE).

In the 5G NR and 6G, vehicle-to-everything (V2X) communication has emerged as a crucial technology. In this paper, we present a novel task-oriented communications (ToC) approach that employs large multi-modal models (LMMs) for V2X tasks such as traffic management. The key idea of the proposed multi-modal task-oriented communications (MMToC) is to extract essential information relevant to vehicular tasks from the various sensing data, integrate the information, and infer the output of vehicular tasks using LMMs. Unlike traditional communication methods focusing on bit-level metrics such as bits per second and error rate, MMToC enhances task efficiency and performance by optimizing information transmission methods for specific vehicular tasks. MMToC exchanges only the essential feature vectors needed for vehicular tasks, thereby reducing communication overhead and improving task performance. Numerical results demonstrate that the proposed MMToC technique significantly increases average vehicle speed by more than 40% compared to conventional methods.

As a key technology to meet the ever-increasing data rate demand in beyond 5G and 6G communications, millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems have gained much attention recently. To make the most of mmWave massive MIMO systems, acquisition of accurate channel state information (CSI) at the base station (BS) is crucial. However, this task is by no means easy due to the CSI feedback overhead induced by the large number of antennas. In this paper, we propose a parametric CSI feedback technique for mmWave massive MIMO systems. Key idea of the proposed technique is to compress the mmWave MIMO channel matrix into a few geometric channel parameters (e.g., angles, delays, and path gains). Due to the limited scattering of mmWave signal, the number of channel parameters is much smaller than the number of antennas, thereby reducing the CSI feedback overhead significantly. Moreover, by exploiting the deep learning (DL) technique for the channel parameter extraction and the MIMO channel reconstruction, we can effectively suppress the channel quantization error. From the numerical results, we demonstrate that the proposed technique outperforms the conventional CSI feedback techniques in terms of normalized mean square error (NMSE) and bit error rate (BER).

Millimeter-wave (mmWave) massive multiple-input multiple-output (MIMO) systems have been considered as a promising solution to provide high-quality data services. To get the most of beamforming gain, an acquisition of accurate channel information is crucial for the success of massive MIMO systems. In recent years, numerous approaches to reduce the excessive feedback overhead due to the large number of antennas have been suggested. However, they do not consider the long-range dependency of the channel state information (CSI) resulting from the correlation between the large-scale antennas and subcarriers. In this paper, we propose a deep learning-based parametric CSI feedback technique for mmWave massive MIMO systems to improve the performance of CSI compression and reconstruction with low feedback overhead. To be specific, the proposed scheme learns the correlation between channel parameters by exploiting Transformer architecture. From the numerical results, we demonstrate that the proposed scheme outperforms the conventional channel feedback schemes in terms of the normalized mean square error (NMSE) and the feedback overhead reduction.