The fifth generation (5G) and beyond wireless networks are envisioned to provide an integrated communication and computing platform that will enable multipurpose and intelligent networks driven by a growing demand for both traditional end users and industry verticals. This evolution will be realized by innovations in both core and access capabilities, mainly from virtualization technologies and ultra-dense networks, e.g., software-defined networking (SDN), network slicing, network function virtualization (NFV), multi-access edge computing (MEC), terahertz (THz) communications, etc. However, those technologies require increased complexity of resource management and large configurations of network slices. In this new milieu, with the help of artificial intelligence (AI), network operators will strive to enable AI-empowered network management by automating radio and computing resource management and orchestration processes in a data-driven manner. In this regard, most of the previous AI-empowered network management approaches adopt a traditional centralized training paradigm where diverse training data generated at network functions over distributed base stations associated with MEC servers are transferred to a central training server. On the other hand, to exploit distributed and parallel processing capabilities of distributed network entities in a fast and secure manner, federated learning (FL) has emerged as a distributed AI approach that can enable many AI-empowered network management approaches by allowing for AI training at distributed network entities without the need for data transmission to a centralized server. This article comprehensively surveys the field of FL-empowered mobile network management for 5G and beyond networks from access to the core. Specifically, we begin with an introduction to the state-of-the-art of FL by exploring and analyzing recent advances in FL in general. Then, we provide an extensive survey of AI-empowered network management, including background on 5G network functions, mobile traffic prediction, and core/access network management regarding standardization and research activities. We then present an extensive survey of FL-empowered network management by highlighting how FL is adopted in AI-empowered network management. Important lessons learned from this review of AI and FL-empowered network management are also provided. Finally, we complement this survey by discussing open issues and possible directions for future research in this important emerging area.

This paper designs a novel energy-efficient hybrid federated and centralized learning (HFCL) framework for training wireless traffic prediction models in aerial networks over distributed multi-access edge computing (MEC) servers where multiple network data analytics functions (NWDAFs) are embedded in the unmanned aerial vehicle (UAV)-aided MEC servers according to the framework standardized by the 3rd Generation Partnership Project (3GPP) specification. Here, UAVs can use federated learning (FL) with their part of local datasets, while the remaining local datasets can be offloaded to the centralized server for centralized learning (CL). To achieve such HFCL energy-efficiently for battery-constrained UAVs, we propose an energy-efficient computation offloading (ECO) scheme for the HFCL. We rigorously formulate analytical models for the overall energy consumption at MEC servers on UAVs and total latency for HFCL process based on the amount of offloaded datasets at each UAV. Here the energy consumption includes i) the transmission energy consumption for offloading the local dataset and ii) the energy consumption for FL processing, which includes local training and local model uploading. To balance between the overall energy consumption and total latency during HFCL process while preventing overload at UAVs, we propose a theoretical framework for the ECO problem that optimizes the amount of offloaded local datasets over multiple UAVs on aerial networks. Here, the ECO problem is formulated as the convex optimization problem under the continuous domain of the amount of offloaded datasets. Our numerical and simulation results show that our proposed framework can construct wireless traffic prediction models with an acceptable training accuracy in an energy-efficient manner over various benchmarks by striking a balance between energy consumption and overall training latency.

In this letter, to alleviate the training burden over Unmanned Aerial Vehicles (UAVs) for generating the mobile traffic prediction model collaboratively, we design a novel energy-efficient mobile traffic prediction framework empowered by Split Federated Learning (SFL) for UAV networks, termed E-SFL. Specifically, in the proposed E-SFL, to enhance latency and energy efficiency, the UAVs are organized into clusters based on their geographic locations and data patterns, and leaders are selected within each cluster considering their battery power levels and proximity to other clients. In each cluster, the proposed E-SFL conducts joint cut layer selection and computation management aiming for minimizing the maximum value of UAVs’ energy consumption during the SFL process. For this purpose, we rigorously formulated an analytical model of the overall energy consumption of UAVs, including both computing and networking energy consumption. The experimental results for two real-world mobile traffic datasets show that the proposed E-SFL surpasses previous state-of-the-art methods in terms of energy consumption with an acceptable accuracy loss.

Hybrid language models (HLMs) are inference-time architectures that combine the low-latency efficiency of small language models (SLMs) on clients (edge devices) with the high accuracy of large language models (LLMs) in centralized servers. Unlike traditional end-to-end LLM inference, HLMs aim to reduce latency and communication by selectively invoking LLMs only when the local SLM’s predictions are uncertain—that is, when the model exhibits low confidence or high entropy in its token-level probability distribution. However, when the SLM encounters ambiguous or low-confidence predictions during inference, it must offload token-level probability distributions to the LLM for refinement. This frequent offloading leads to substantial communication overhead, particularly in bandwidth-constrained environments. To address this challenge, we propose federated learning (FL)-enabled HLM (FedHLM), a communication-efficient HLM framework that integrates uncertainty-aware inference with FL. The key innovation lies in collaboratively learning token-level uncertainty thresholds that determine when SLM predictions require LLM assistance. Instead of relying on static or hand-tuned thresholds, FedHLM uses FL to enable distributed threshold optimization across clients while preserving data privacy. Additionally, embedding-based token representations are employed to facilitate semantic similarity comparisons during peer-to-peer (P2P) resolution, allowing clients to reuse tokens inferred by similar peers without efficiently involving the LLM. Moreover, we propose hierarchical model aggregation as a strategy to reduce redundant token transmissions. At the edge server level, client updates are aggregated to refine local routing policies, while global coordination across clusters further synchronizes decision boundaries. This layered approach ensures that repeated uncertainty patterns are captured and resolved locally, significantly reducing unnecessary LLM queries. Extensive simulations on large-scale news classification tasks demonstrate that FedHLM achieves over 95% reduction in LLM transmissions with negligible accuracy loss, highlighting its potential for scalable and efficient edge-artificial intelligence (AI) deployment.

To alleviate the substantial local training burden on clients in the federated learning (FL) process, this paper proposes a more efficient approach based on hybrid federated and centralized learning (HFCL), leveraging the Mobile Edge Computing (MEC) environment within wireless communication networks. Considering the existence of heterogeneous data types with different privacy levels -such as i) sensitive data, which can not be exposed, and ii) less-sensitive data, which can be exposed for centralized learning (CL)-we formulate an optimization problem aimed at achieving a balance between i) total latency, including computation and communication, and ii) the training burden on the MEC server. This balance is achieved by adjusting the set of participants in FL, taking into account client selection under different privacy levels. A multi-objective optimization problem is designed using mixed-integer nonlinear programming, which is generally recognized as NP-hard. We employ relaxation techniques in combination with the Mutas & Simulated Annealing Heuristic algorithm to develop a near-optimal yet practical algorithm. Our numerical and simulation results reveal that the proposed scheme effectively achieves a global model by striking a balance between the total time required for model convergence and the computational load on the MEC server. Furthermore, experimental results on three well-known real-world datasets demonstrate that the proposed scheme maintains an acceptable level of accuracy and loss.

Federated learning (FL) enables privacy-preserving adaptation of large language models (LLMs) across distributed clients. However, deploying FL in edge environments remains challenging because of the high communication overhead of full-model updates. Recent advances in parameter-efficient fine-tuning (PEFT), particularly low-rank adaptation (LoRA), have substantially reduced update sizes by injecting lightweight trainable matrices into pretrained transformers, thereby making FL with LLMs more feasible. In this paper, we propose LoRaC-GA, a communication-aware optimization framework that dynamically determines the optimal number of clients to participate in each round under a fixed bandwidth constraint. We formulated a max-min objective to jointly maximize the model accuracy and communication efficiency and solved the resulting non-convex problem using a genetic algorithm (GA). To further reduce the overhead, we integrated a structured peer-to-peer collaboration protocol with ${log}_{2}K$ complexity, enabling scalable communication without full connectivity. The simulation results demonstrate that LoRaC-GA adaptively selects the optimal client count, achieving competitive accuracy while significantly reducing the communication cost. The proposed framework is well-suited for bandwidth-constrained edge deployments involving large-scale LLMs.