The proliferation of computation-and energyintensive AI/ML applications in 6G challenges resourceconstrained User Equipment (UE). Split AI, which partitions inference tasks between the UE and network, can reduce UE energy consumption. However, existing methods often neglect constraints on UE energy expenditure and lack mechanisms to adapt to heterogeneous, time-varying network conditions. This paper proposes EnSplit, a dynamic split inference framework that jointly optimizes AI/ML inference latency and UE energy consumption in 6G networks, by regulating the cost of UE energy expenditure at the mobile edge. In this work, AI/ML inference is modeled as a Directed Acyclic Graph (DAG) of neural network partitions, and the split problem is formulated as an optimization task that incorporates UE energy budgets. A Deep Reinforcement Learning (DRL) agent adaptively selects split points across the UE and network edge, accounting for both resource variability and UE energy constraints. Evaluation in a simulated environment with dynamic and heterogeneous computing and communication resources shows that EnSplit reduces inference latency by up to 47% and lowers UE energy consumption by up to 80% compared to baseline inference approaches, while respecting UE energy credit constraints.

This paper introduces Two-Way Ranging-based Cooperative Localization with an Open Ultra-Wideband Dataset from Bern-University (TWR-CLOUD-Bern), a multi-node Ultra-Wideband (UWB) Two-Way Ranging (TWR) dataset and, to our knowledge, the largest open dataset of its kind to date. Collected across 12 scans covering more than 500m2 with 40 Qorvo DWM3001C nodes, the dataset provides ground-truth coordinates, Time of Flight (ToF) measurements, and reception statistics. To support fair, transparent, and reproducible comparisons, we also release a companion evaluation framework. In this framework, 11 scans serve as training partitions for data-driven ranging-model development, while one independent scan with distinct node placements is reserved for testing. The framework fixes node selections and measurement configurations, while varying node counts and locations. It supports single-and multi-sample ranging and covers anchor-free cooperative, anchor-aided cooperative, and anchor-only non-cooperative localization under real-world node-to-node connectivity. The dataset, framework, and documentation are openly available through the Zenodo archival release, making TWR-CLOUD-Bern an immediately usable testbed for UWB localization research.

The growing complexity, scale, and heterogeneity of 6G wireless systems call for a shift toward AI-native architectures that are not only data-driven but also topology-aware, adaptive, and distributed. Graph Neural Networks (GNNs), with their native support for graph-structured data, are well-suited for modeling the irregular and dynamic relationships inherent in wireless communication systems. However, standalone GNNs may be insufficient to address key 6G challenges such as continual learning, data scarcity, and dynamic adaptation. This survey, therefore, explores the emerging synergy between GNNs and complementary AI paradigms, including deep reinforcement learning (DRL), federated learning (FL), meta-learning, generative models, mixture-of-experts (MoE), and world models, enabling hybrid AI-GNN architectures for intelligent control, predictive adaptation, and scalable optimization across the wireless stack. We systematically review how these GNN-centric AI models can support intelligent functionality in next-generation network architectures such as O-RAN, as well as core 6G domains, including edge computing for wireless systems, advanced MIMO, traffic prediction, and digital twins. The survey also highlights key challenges in hybrid AI-GNN adoption, particularly scalability, generalization across dynamic topologies, interpretability, and symbolic reasoning, and discusses emerging strategies such as graph causality learning, dynamic GNNs, and neurosymbolic integration to address them. By consolidating recent advances and outlining open research directions, this work positions hybrid AI-GNN architectures as a promising approach toward enabling intelligent, energy-efficient, and context-aware wireless systems for 6G and beyond.

We present ARTour, a latency-aware framework for real-time augmented reality (AR) tours that couples visual matching with prediction-driven prefetching on a glasses-companion platform. The system targets lightweight AR glasses, minimizing end-to-end (E2E) latency and on-device energy while sustaining reliable overlays in outdoor and indoor scenes. For outdoor landmarks, a centroid-based recognition module attends from a learned query over detector decoder embeddings to aggregate multiple sub-components into point-of-interest (POI)- level descriptors robust to partial views and geo-location noise. For indoor deployments, a fine-grained recognition module treats cross-attention weights over room-level candidates as task-specific similarity scores, enabling discrimination among visually similar artifacts in museum-like scenarios. A self-supervised next-POI prediction module learns user mobility patterns from trajectories and proactively preloads the corresponding metadata, turning most wide-area transfers into background traffic. Experiments on standard landmark, fine-grained retrieval, and next-POI benchmarks show that ARTour attains competitive or state-of-the-art accuracy, with mean Average Precision (mAP) scores of 91.5/94.3 on Oxford5K/Paris6K and leading performance on CUB-200- 2011, Stanford-Cars, and FGVC-Aircraft, while consistently improving Recall@k and Mean Reciprocal Rank (MRR) over recent next-POI baselines on the Foursquare NYC and Tokyo datasets. Using scripted replay and controlled interaction rather than human-participant QoE studies, system-level measurements on a glasses-companion prototype confirm median and 95thpercentile times-to-overlay (TTO) of 40 ms and 58 ms-p to 66% lower than a glasses-only baseline-and a reduction of 35% in glasses-side power, demonstrating that ARTour can deliver responsive and energy-efficient AR tours under realistic wireless and hardware constraints.

Mobile Mixed Reality Network (MMRN) is an emerging decentralized, infrastructure-free system that enables users to interact in a shared space blending real and virtual worlds via mobile Head-Mounted Devices (HMDs). These immersive experiences increasingly rely on Artificial Intelligence (AI) services such as gaze tracking to deliver responsive interactions. However, the limited computational and communication capabilities of HMDs make it challenging to support AI services locally, while the mobility and dynamic data distributions of HMDs hinder device-to-device collaboration. To address such issues, we propose FAST, a fully decentralized framework for scalable and resource-efficient AI service provisioning in MMRN. FAST performs online prediction with a time-varying loss function to enable model adaptation to evolving data streams, and exploits pairwise proximity constraints to identify shared interests among nearby HMDs, enabling selective device-to-device model sharing for improved local performance. We propose a decentralized online saddle-point algorithm that dynamically adjusts model complexity to balance accuracy and resource usage, with theoretical guarantees of sublinear regret and constraint violations under time-varying topologies. Simulations on real-world datasets show that FAST outperforms state-of-the-art baselines by up to 28% in accuracy and reduces computation and communication costs by 39% under user mobility and data heterogeneity.

Vertical Federated Learning (VFL) enables collaborative machine learning across institutions holding complementary features for shared entities while preserving data privacy. However, existing VFL systems face key limitations. Centralized architectures introduce single points of failure and communication bottlenecks, while decentralized approaches incur substantial communication overhead due to high-dimensional parameter exchange. Moreover, VFL performance often degrades as the number of participating clients increases, limiting scalability. This paper introduces VFL-GNN, a communication-efficient decentralized VFL framework based on Graph Neural Network (GNN) that reformulates vertical collaboration as a distributed graph learning problem. Instead of exchanging model parameters or embeddings, clients collaboratively learn cross-feature affinity structures by sharing lightweight adjacency matrices and coordinating through a ring-based consensus protocol. We prove exponential consensus convergence via spectral analysis and evaluate VFL-GNN on multiple benchmark datasets. Experimental results show that VFL-GNN reduces communication overhead while maintaining competitive performance even as the number of participating clients increases.

Transformer-based architectures have introduced end-to-end solutions for Multiple Object Tracking (MOT), seamlessly integrating object detection and association. However, their high computational demands—such as the need for feature map fusion across multiple frames—pose significant challenges to real-time deployment, limiting their practicality. In this paper, we present WT-MOT (Weight Table-based Multiple Object Tracking), a novel framework that addresses these limitations by leveraging the underutilized potential of attention weight tables for efficient object similarity evaluation. WT-MOT employs self- and cross-attention mechanisms to assess object similarity and directly assign identifications, integrating spatial, appearance, and temporal dimensions. By introducing the “frame embedding” concept, WT-MOT enhances the ability to distinguish objects across frames without relying on motion models or post-processing steps. Experimental results on the MOT17 and MOT20 benchmarks demonstrate the effectiveness of WT-MOT, achieving MOTA scores of 76.9% and 73.2%, respectively, setting new performance standards for Transformer-based MOT solutions. These findings highlight WT-MOT as a computationally efficient and robust tool for real-time MOT applications, paving the way for broader adoption of Transformer-based tracking methods in practical environments.

Edge computing enables distributed intelligence in resource-constrained IoT environments. However, traditional Federated Learning (FL) struggles with heterogeneous device capabilities, dynamic network conditions, and non-IID data distributions, resulting in straggler effects, slow convergence, and inefficient resource utilization. This paper proposes Resource-Aware Dynamic Split Federated Learning (RAD-SFL), a framework for efficient distributed training in heterogeneous edge environments. RAD-SFL introduces a dynamic model layer splitting mechanism that adaptively partitions model execution between client devices and edge servers based on real-time computation and communication conditions, and a group-and-reorder technique that organizes devices into balanced groups with similar data distributions to improve model convergence under non-IID settings. We validate RAD-SFL through experiments on widely adopted datasets using both a simulated environment and a real testbed with heterogeneous IoT devices. Results demonstrate that RAD-SFL reduces the training time by up to 66.4%, decreases device-side energy consumption by 52.5%, and improves global model accuracy by up to 30.5% compared to FL and SFL baselines.

Federated learning (FL) enables collaborative training across distributed edge devices while preserving data privacy, but faces critical challenges in heterogeneous environments: diverse client computational capabilities, limited energy budgets, constrained communication bandwidth, and non-uniform data distributions. Existing approaches treat client selection and model adaptation independently or rely on static configurations, limiting adaptability and leading to suboptimal trade-offs between model accuracy and resource efficiency. To address these limitations, we propose FedRAPS (Federated Learning with Resource-Aware Pruning and Client Selection), a deep reinforcement learning-based framework for federated orchestration that jointly optimizes client selection and adaptive model pruning. We formulate federated orchestration as a reinforcement learning problem to balance model accuracy, energy consumption, and communication latency. The agent leverages per-client resource profiles and global training metrics to jointly select participating clients and assign client-specific pruning rates. Experiments on CIFAR-10 and Fashion-MNIST demonstrate that FedRAPS achieves 17–21% energy savings and 8–21% training time reductions while maintaining similar accuracy compared to baseline FL methods.

Flexible Network Service composition is a fundamental enabler for the design of 6G networks, where network services are modeled as ordered Service Function Chains (SFCs) with heterogeneous Virtual Network Functions (VNFs). However, dynamic traffic generated by end users and dynamic network resource infrastructure utilization make online context-aware and resource-efficient SFC composition challenging. While Deep Reinforcement Learning (DRL) has been explored for this task, the multimodal nature of traffic and the variable-length inputs limit achievable performance. To address these challenges, we propose an attention-based actor–critic framework that integrates Transformer self-attention and encoding to capture variable-length SFC states and inter-VNF dependencies. The learned representations are then used by an actor–critic policy to sequentially select resource-aware composition actions, enabling adaptive and efficient service chain construction under dynamic network conditions. Extensive simulations show that our proposed transformer-augmented actor-critic DRL achieves faster policy convergence, lower bandwidth and computational resource consumption, and higher deadline satisfaction rates compared to state-of-the-art baselines.

Deploying federated learning (FL) across constrained residential networks is essential for decentralized edge applications. However, data heterogeneity, bandwidth limits, and evolving user behaviors challenge large-scale distributed learning. This paper introduces AdaX-Fed, an adaptive and explainable FL framework optimizing communication efficiency and predictive stability for residential network load forecasting. It leverages density-based clustering to isolate atypical profiles and form behaviorally cohesive groups, ensuring robust initialization for cluster-specific training. AdaX-Fed integrates Shapley Additive Explanations into an attention-based temporal backbone, providing dual-layer interpretability while minimizing explainability noise and edge latency. A stability controller monitors partition and explanation drift, triggering warm-started re-clustering when behavioral shifts are detected. Simultaneously, reliability-aware aggregation and adaptive quantization reduce communication overhead during updates. Evaluated on real-world datasets from Germany, Switzerland, and Mexico, AdaX-Fed reduces Mean Absolute Error by 36.5%, saves 480 MB of transmitted data, minimizes XAI latency to 59.96 ms, and ensures temporal reliability against state-of-the-art FL baselines.

Network slicing enables 5G and beyond networks to concurrently support diverse services, including emerging IoT applications, over a shared, softwarized infrastructure. Traditional resource management approaches are often reactive, resulting in suboptimal utilization and potential Service Level Agreement (SLA) violations under dynamic traffic conditions. We present Proactive Resource OPtimization for Heterogeneous nETwork Slicing (PROPHET), a framework for Proactive Resource Optimization in heterogeneous slice-enabled networks. PROPHET integrates an attention-based traffic forecasting model with a deep reinforcement learning (DRL) resource allocator based on Proximal Policy Optimization (PPO) to anticipate future slice demands and proactively adjust configurations. Evaluated using real LTE traffic traces representing enhanced Mobile Broadband (eMBB), ultra-Reliable and Low-Latency Communication (uRLLC)-like, and best-effort services, PROPHET improves SLA compliance and resource efficiency compared to reactive baselines. The proposed framework is equally applicable to IoTdriven slices, supporting massive connectivity and ultra-reliable low-latency requirements envisioned for next-generation smart environments.

Deep reinforcement learning (DRL) has been used very successfully for a variety of applications in the field of vehicular networking. We investigate the generalizability of a DRL-based scheduling approach for task offloading in vehicular edge computing (VEC). We propose a hybrid scheduling policy that combines heuristic task prioritization with a learned DQN-based resource selection. Two deep Q-learning (DQN) architectures are compared: a conventional multilayer perceptron (MLP) with fixed-size inputs; and a permutation-invariant architecture that uses a lightweight transformer encoder. Trained exclusively on a simple straight-road scenario, the permutation-invariant DQN achieves perfect selection of the fastest processing vehicle when tested in an unseen realistic scenario. In contrast, the MLP-based DQN suffers severe degradation due to positional bias and padding artifacts. Results demonstrate that set-based, permutation-invariant representations are essential for sample efficiency and generalization in VEC environments with high variability.

In emergency rescue scenarios, such as fires where smoke and dust reduce visibility, millimeter-wave (mmWave) radar offers advantages over vision-based localization due to its penetration capability and adaptability to harsh environments. Nevertheless, signal sparsity, multipath interference, and the high cost of data annotation limit its effectiveness in loop closure detection, while conventional trajectory optimization algorithms often lack adaptability, reducing localization accuracy. To address these challenges, this work presents mmDif-DRL, a cooperative and autonomous localization system based on mmWave radar. First, this work proposes a multimodal loop closure detection model by integrating mmWave point clouds and geomagnetic signals. In this model, a Transformer encoder is employed to capture temporal dependencies, and a bidirectional cross-attention mechanism enables effective multimodal fusion. Second, this work proposes a diffusion-based data generation strategy to provide sufficient training samples for the loop closure detection model, while also reducing the need for real data collection and manual annotation. Third, this work proposes a deep reinforcement learning (DRL)-based trajectory optimization strategy, integrated with a Graph-based Simultaneous Localization and Mapping (GraphSLAM) algorithm to achieve robust global optimization. This approach balances adaptability in complex environments with accurate localization. Experiments show that mmDif-DRL achieves an average loop closure detection F1-score of 94.44% and reduces the average localization error to 0.21 m across three indoor scenarios, demonstrating its robustness under challenging conditions.

Mobile Internet of Things (IoT) applications increasingly demand low-latency service delivery and efficient content dissemination; yet, conventional cloud-centric solutions often suffer from excessive backhaul delay. We present an energy-aware floating architecture that enables peer-assisted edge computing and cooperative content caching across mobile IoT devices. The framework comprises three key components: 1) floating service, where selected user nodes (UNs) temporarily act as local servers (LSs) and adapt in real time to node mobility to sustain seamless service provisioning; 2) cooperative caching, which supports dynamic, proximity-based content exchange among neighboring devices; and 3) a mobility-aware service-migration predictor that forecasts user movement to guide timely service handover and cache preloading. A radio duty-cycle energy model records transmit (TX), receive (RX), and idle (IDLE) activity to estimate device-side power expenditure. Extensive simulations show that the proposed system reduces end-to-end latency by up to 87.5% for more than 90% of service requests, while the cooperative caching component lowers retrieval delay by up to 58.1% for 28% of communications. Crucially, the floating mechanism increases the average device radio duty cycle by only 1-3 percentage points, corresponding to an incremental power cost of just 3.6-24.6 mW per node. These results underscore the practicality of energy-aware, peer-assisted edge collaboration for next-generation mobile IoT deployments.

G mobile networks are increasingly using Artificial Intelligence to manage highly dynamic environments characterized by time-varying traffic demands, user mobility, and heterogeneous resources. The dynamic behavior of User Equipment makes timely and accurate control decisions challenging, while distributed data exchange introduces communication overhead and privacy concerns. These challenges call for scalable and communication-efficient learning mechanisms for Radio Access Network (RAN) orchestration. In this paper, we propose DERRIC-FRL, a decentralized Federated Reinforcement Learning framework to orchestrate RAN intelligent controllers. DERRIC-FRL jointly optimizes controller placement and user power allocation through a selective two-level aggregation mechanism, reducing data exchange to only selected orchestrators and controllers while preserving user privacy and improving overall network performance. Specifically, our method significantly reduces total training communication costs by 34% across inter-domain connections, and up to 77% across intra-domain connections, compared to the FedAvg approach. Furthermore, DERRIC-FRL improves user throughput by up to 53% and 61% compared to the DERRIC and FedAvg baselines across a broad range of simulated scenarios.

Open-Radio Access Network (O-RAN) facilitates the scalability of cellular networks by introducing a RAN Intelligent Controller (RIC) component whose functions can be flexibly distributed over large-scale 6G networks. Artificial Intelligence (AI) is effective in optimizing RIC placement in 6G O-RAN, mitigating the limited adaptability of non-data-driven methods in complex time-varying network conditions. However, the centralized orchestration of current approaches for RIC placement hinders scalability. This work introduces a data-driven DEcentralized Reinforced RAN Intelligent Controller orchestration (DERRIC) method for 6G networks, leveraging the online learning capabilities of decentralized multi-agent Reinforcement Learning (RL) orchestration to solve the RAN Intelligent Controller Placement Problem (CPP). DERRIC is a two-layer network management scheme with decentralized orchestrators that adapt to network conditions, deploy controllers, and allocate resources. These orchestrators manage distributed controllers to optimize RAN parameters, such as user transmission power. DERRIC's main goal is to increase the system's overall user Packet Delivery Ratio (PDR) by optimal controller deployment and operation. Optimal controller deployment reduces controller-user latency and accelerates user-transmission-power control decisions, leading to further enhancement to user PDR. We show that DERRIC reduces the controller-user latency and power consumption by up to 66% and 29% and increases user PDR by up to 14% compared to state-of-the-art baselines in a broad range of simulated scenarios.

The rapidly evolving realm of Extended Reality (XR) demands high bandwidth and low-latency communication to support immersive experiences such as high-resolution 360-degree videos and real-time interactions in virtual reality gaming. In this study, “resources” are defined as digital assets essential for XR applications, divided into “static resources” (immutable media files such as textures and video segments) and “dynamic resources” (real-time user data and interactive elements crucial for user interactions). A primary challenge in XR environments is optimizing the delivery and caching of these resources within existing network infrastructures to enhance the Quality of Experience (QoE) for users. We introduce a novel hybrid network architecture that integrates resource caching, user-to-user communication, and central server oversight. This architecture not only ensures reliable delivery but also significantly reduces communication latency. Preliminary experiments, conducted under conditions where each node in the network has a 10% chance of failing at any given time, demonstrate that our approach enhances the delivery efficiency of static resources by 68%, affecting 38% of communications, with an increase in latency observed in 4% of cases by 22%. For dynamic resources, it reduces latency in 89% of the cases by an average of 30%, though 8% of cases experienced a 36% increase in latency. These results affirm the effectiveness of our architecture in enhancing user experience in XR environments under challenging network conditions.

The Cloud Continuum Framework (CCF) logically integrates distributed extreme edge, far edge, near edge, and cloud data centers in 6G networks. Deploying VNFs over the CCF can enhance network performance and Quality of Service (QoS) for modern delay-sensitive applications and use cases in 6G networks. Deep Reinforcement Learning (DRL) has shown potential to automate Virtual Network Function (VNF) migrations by learning optimal policies through continuous monitoring of the network environment. In this work, we leverage Deep Reinforcement Learning to optimize network control policies that continuously update VNF placement for optimal Service Function Chain (SFC) deployment in time-varying user traffic scenarios. By leveraging dynamic VNF relocation, this approach seeks to improve network performance in terms of latency, operational costs, scalability, and flexibility. This study addresses the gap in existing solutions by jointly considering network performance requirements and migration costs, providing a more comprehensive strategy for efficient VNF deployment and management. We show that our proposed DRL-based VNF deployment method achieves a 28.8% lower delay and a 34% lower migration overhead compared to state-of-the-art baselines in a broad range of large-scale simulated scenarios, showing the proposed method's scalability features.

Networks face growing challenges such as Distributed Denial-of-Service attacks, malware propagation, unauthorized access and irregular traffic patterns. Traditional centralized or static anomaly detection methods often struggle with scalability, latency and privacy concerns, making them less effective in dynamic mobile environments. This research introduces a decentralized federated learning algorithm that integrates Multi-Agent Reinforcement Learning (MARL) with tree-based classification models, using XGBoost (Extreme Gradient Boosting) in the experimental setup to enhance anomaly detection while preserving privacy. MARL coordinates agent behavior to adapt to dynamic network conditions, promoting decentralized collaboration and resilience without raw data sharing. Each agent trains a lightweight tree-based model locally for efficient and scalable detection. Feature selection is performed collaboratively using reinforcement signals, while trust-weighted aggregation enables decentralized global learning without a central server. The system is designed to address key constraints such as intermittent communication, which is modeled through agent dropout in the experiments, as well as computational overhead and scalability in edge network environments. Experimental results using real-world intrusion detection datasets show that the proposed method achieves up to 93.5% accuracy, 86.1% precision and 85.7% F1 score, outperforming baseline methods by as much as 23.3% in F1 score under non-IID scenarios while maintaining low bandwidth and energy usage across dynamic agent participation and varying attacker models.