EURECOM

This work presents a first comprehensive analysis of the impact of vector coded caching (VCC) in multi-user multiple-input multiple-output (MU-MIMO) systems with multiple receive antennas and variable pathloss — two key factors that critically influence systems with inherent MU unicasting behavior. We investigate two widely adopted precoding strategies: (i) block-diagonalization (BD) at the transmitter combined with maximal ratio combining (MRC) at the receivers, and (ii) zero-forcing (ZF) precoding. Our analysis explicitly accounts for practical considerations such as channel fading, channel state information (CSI) acquisition overhead, and fairness-oriented power allocation. Our contributions span both analytical and simulation-based fronts. On the analytical side, we derive analytical expressions for the achievable throughput under BD-MRC and ZF, highlighting the performance benefits of equipping multi-antenna users with cache-aided interference management. Specifically, we develop a low-complexity BD-MRC optimization method that leverages matrix structure to significantly reduce the dimensionality involved in precoding computation, followed by solving the associated max-min fairness problem through an efficient one-dimensional search. In the massive MIMO regime, an asymptotic expression for the achievable throughput over Rayleigh fading channels is also derived. Simulations validate our theoretical results, confirming that VCC delivers substantial performance gains over optimized cacheless MU-MIMO systems. For example, with 32 transmit antennas and 2 receive antennas per user, VCC yields throughput improvements exceeding 300%. These gains are further amplified under imperfect CSI at the transmitter, where VCC’s ability to offload interference mitigation to the receivers ensures robust performance even in the face of degraded CSI quality and elevated acquisition costs.

International audience

The Cloud-Edge collaborative computing enables the deployment of latency-sensitive and data-intensive applications closer to end users. However, it introduces significant challenges for microservice placement, due to resource heterogeneity, limited edge capacity, and the need to satisfy storage requirements using aggregated resources across multiple nodes. To address these issues, we propose <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Diktopos</b>, a topology-aware, two-stage scheduling framework that jointly optimizes microservice placement and distributed storage volume allocation in cloud-edge networks. The joint optimization problem is decomposed into two subproblems: (i) microservice placement and (ii) distributed volume allocation, with the objective of minimizing computation, communication, energy, and storage costs. At its core, Diktopos employs a low-complexity, rank-based heuristic that ensures scalable and accurate placement across heterogeneous edge nodes. Simulation results show that our method achieves near-optimal placement decisions (within 1.67% of the optimal solution), and converges up to 5× faster than state-of-the-art approaches in large-scale deployments. Real-world experiments in Kubernetes environments demonstrate up to 53% latency reduction compared to the default scheduler, and up to 23% improvement over other baselines, confirming Diktopos' effectiveness in dynamic, resource-constrained edge scenarios.

We study fluid antenna (FA) port selection in underlay cognitive radio (CR) networks, where a single FA-enabled secondary transmitter serves a single-antenna secondary receiver under the underlay CR mechanism that caps the interference at the primary user via instantaneous power control. For each FA port, we formalize the resulting signal-to-noise ratio (SNR) and propose a CR-aware selection rule that prioritizes ports offering a favorable balance between the secondary-link gain and the power-control link to the primary user. This contrasts with the conventional main-channel maximization (MCM) rule, which ignores the power-control link imposed by underlay CR—a choice often adopted in non-CR scenarios. Through rigorous mathematical derivations, we establish the exact diversity orders of both FA policies and show that each attains a diversity order equal to the rank of the FA port correlation matrix. For comparison, we further consider a traditional multi-antenna transmitter employing maximum-ratio transmission under the same underlay constraint and prove that its diversity order equals the number of transmit antennas. Monte Carlo simulations over Rayleigh fading channels validate the analysis, demonstrating sizable outage gains of the proposed CR-aware selection over MCM, and performance comparable to that of traditional multi-antenna transmission.

• Introduce a new agentic paradigm- symbiotic agents -that pairs large language models with optimizers and examine it through the lens of trustworthy AI. • Implement two concrete agent designs for dynamic RAN control and multi-tenant SLA negotiations. • Leveraged open-source real-world platforms, including OpenAirInterface (OAI) for the 5G user equipments (UEs), RAN and Core Network, and FlexRIC for the Radio Intelligent Controller (RIC). • Utilized real-world datasets with RAN channel quality fluctuations from moving vehicles to emulate realistic mobility scenarios. • Proved that optimization algorithms maximize the decision accuracy of LLM agents with up to 5 times lower error. • Evaluated both larger and smaller models (SLMs) proving that SLMs retain sufficient near-RT performance (82 ms loop) with substantially smaller GPU overhead (99.9 % less). • Introduced and validated a novel LLM-driven architecture towards AGI networks leveraging the developed agents. Large Language Model (LLM)-based autonomous agents are expected to play a vital role in the evolution of 6G networks, by empowering real-time decision-making related to management and service provisioning to end-users. This shift facilitates the transition from a specialized intelligence approach, where artificial intelligence (AI) algorithms handle isolated tasks, to artificial general intelligence (AGI)-driven networks, where agents possess broader reasoning capabilities and can manage diverse network functions. In this paper, we introduce a novel agentic paradigm that combines LLMs with real-time optimization algorithms towards Trustworthy AI, defined as symbiotic agents . Optimizers at the LLM’s input-level provide bounded uncertainty steering for numerically precise tasks, whereas output-level optimizers supervised by the LLM enable adaptive real-time control. We design and implement two novel agent types including: (i) Radio Access Network (RAN) optimizers, and (ii) multi-agent negotiators for Service-Level Agreements (SLAs). We further propose an end-to-end architecture for AGI-driven networks and evaluate it on a 5G testbed capturing channel fluctuations from moving vehicles. Results show that symbiotic agents reduce decision errors fivefold compared to standalone LLM-based agents, while smaller language models (SLM) achieve similar accuracy with a 99.9 % reduction in Graphical Processing Unit (GPU) resource overhead and in near-real-time (near-RT) loops of 82 m s . A multi-agent demonstration for collaborative RAN on the real-world testbed highlights significant flexibility in service-level agreement and resource allocation, reducing RAN over-utilization by approximately 44 %. Drawing on our findings and open-source implementations, we introduce the symbiotic paradigm as the foundation for next-generation, AGI-driven networks-systems designed to remain adaptable, efficient, and trustworthy even as LLMs advance. A live demo is presented here https://www.youtube.com/watch?v=WQv61z1deXs&ab\_channel=BubbleRAN

In dense urban environments, traditional GNSS-based navigation for Unmanned Aerial Vehicles (UAVs) suffers from multipath interference and signal obstructions, compromising positioning accuracy and increasing risks of collisions and airspace violations. This paper proposes a novel machine learning-based system architecture for autonomous UAV parcel delivery, leveraging standardized 5G Network Exposure Function (NEF) and CAMARA Device Location API to achieve sub-meter location precision. Our approach integrates dynamic geofencing and predictive rerouting at the network edge, powered by a Random Forest-based collision prediction model that proactively adjusts UAV trajectories to avoid restricted zones in real time. Through simulations of six UAVs navigating dynamically updated no-fly zones, we demonstrate that our system significantly reduces time spent in restricted areas to near zero, compared to GNSS-only and rule-based methods, while limiting path-length inflation to approximately 30% for five of six flights. These results underscore the potential of combining 5G-enabled location services with edge intelligence to enhance safety and compliance in urban UAV operations.

5G New Radio (NR) is a key enabler of accurate positioning in smart cities and smart factories. This paper presents the experimental results from three 5G positioning testbeds running open-source OpenAirInterface (OAI) gNB and Core Network (CN), using Uplink Time Difference of Arrival (ULTDoA) with the newly integrated Location Management Function (LMF). The testbeds are deployed across both indoor and outdoor scenarios with O-RAN Radio Units (O-RUs), following a 3GPP-compliant system model. The experiments highlight the impact of synchronization impairments, multi-path propagation, and deployment geometry on positioning accuracy. To address these challenges, we propose tailored ToA and TDoA filtering as well as a novel position estimation method based on Particle Swarm Optimization (PSO) within the LMF pipeline. Moreover, we show a beyond-5G framework that leverages non-conventional measurements such as Channel Impulse Response (CIR) to train and test Artificial Intelligence and Machine Learning (AI/ML) models for data-driven positioning. The results demonstrate the feasibility of achieving 1–2 meter positioning accuracy in 90% of cases in different testbeds, offering practical insights for the design of robust 5G positioning systems. Moreover, we publicly release the datasets collected in this work to support the research within the 5G positioning community.

Current research in end-to-end speech deepfake detection predominantly centers around inputting “raw” waveforms to a deep architecture, such as RawNet2, and training the deep neural network to predict if the waveforms are fake. However, direct processing of waveforms could cause over-parameterization in the network, reducing its generalizability. To overcome this limitation, we propose a multi-level variational regularization framework integrating a modified Variational Autoencoder (VAE) with discriminative constraints. Specifically, we adopt an VAE with a deepfake discrimination constraint to regularize a RawNet2-based high-level feature map (HFM) extractor. Experimental results show that the proposed variational regularization leads to HFM features that improve the performance of AASIST, SE-Rawformer, and RawBMamba by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{3 6. 0 1 \%, } \text{1 0. 0 7 \%}$</tex>, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{6. 3 5 \%}$</tex>, respectively.

Next-generation wireless networks are poised to redefine our digital world, by integrating advanced artificial intelligence (AI) capabilities, enabling intelligent support for emerging applications such as Extended Reality (XR), Holographic Telepresence, Robotics, and Autonomous Mobility. In this environment, Consumer Electronics (CE) devices, such as smartphones, smartwatches, medical implants, and IoT devices, will play a crucial role in various real-world scenarios. However, due to their limited battery and processing capacities; training complex and large deep learning models on such devices is computationally intensive and can result in high energy and power consumption, and early termination of the training. Therefore, it is crucial to optimize the training process using lightweight models and offloading computation to more powerful servers while providing good learning performance. Distributed machine/deep learning algorithms offer viable alternatives in this context. They consist of deploying distributed agents that collaborate to train complex models without sharing local data. Federated Learning (FL) is a popular distributed learning paradigm in the literature. Nevertheless, as mentioned earlier, training the full FL model on computationally constrained nodes is not feasible.