AI4C8 – AI/ML Solutions for Communications
Friday, 5 June 2026, 9:00-10:30, room Sala 4 (1st floor)
Session Chair: Francisco Luna (Univ. Málaga, ES)
Deep Learning Based Precoding Without Explicit CSI for Wideband THz UM-MIMO Systems
Akram Najjar, Mohammed El-Absi and Thomas Kaiser (University of Duisburg-Essen, Germany)
This paper proposes an unsupervised deep learning (DL)-based hybrid analog precoding framework incorporating true time delay (TTD) and phase shifter (PS) elements to mitigate the severe beam squint effect in wideband terahertz (THz) ultra-massive multiple-input multiple-output (UM-MIMO) systems. The proposed framework operates directly in the measurement domain, bypassing explicit channel state information (CSI) reconstruction at inference stage while jointly optimizing TTD and PS coefficients to maximize average array gain across wideband frequencies for multiple users. Specifically, we design a model-based loss function that guides the training process to maximize wideband average array gain without requiring labeled precoding coefficients. Through this design, and unlike conventional approaches that rely on accurate CSI, the proposed framework learns to map pilot-induced measurements to hybrid precoding coefficients in a single step. Simulation results validate that the proposed DL-based precoding framework effectively mitigates the beam squint effect, achieving near-optimal performance with moderate computational complexity.
A Feasibility-Shielded Agentic AI Framework for 6G Self-Healing Core Networks
Yen-Ting Chen, Tuck-Wai Choong, Yu-Chiao Jhuang and Jenq-Shiou Leu (National Taiwan University of Science and Technology, Taiwan)
The transition to 6G networks requires a shift from rigid automation to intent-driven cognitive autonomy. Although large language models (LLMs) offer promising reasoning capabilities for network orchestration, their stochastic nature introduces operational risks such as malformed commands and parameter errors, which compromise the reliability of autonomous management systems. This paper proposes a cognitive digital twin (CDT) framework that integrates agentic AI with a deterministic feasibility shield for reliable autonomous fault remediation. Operating on a modified OODA (Observe-Orient-Decide-Act) loop, our agent leverages a structured domain knowledge base to diagnose faults and validates candidate actions via short-horizon lookahead simulation before execution. Experimental results in a 3GPP-compliant 6G Core simulator demonstrate that the proposed framework improves the packet delivery ratio (PDR) by 38.83 percentage points during critical link failures compared to baseline systems, achieving a deterministic mean time to repair (MTTR) of 50 ms in simulation time. The multi-layer validation pipeline comprising schema enforcement and simulation-based verification achieved a 100% action success rate in the physical environment. This paper demonstrates that locally deployable open-source models (e.g., qwen3-30b, glm-4.7-flash) achieve diagnostic performance comparable to or exceeding proprietary commercial models when constrained by our framework, offering a viable path toward sovereign, privacy-preserving network autonomy.
LLM-Guided Reinforcement Learning for Adaptive Inter-Slice Resource Prioritization in 6G O-RAN
Martino Chiarani (UPC Universitat Politècnica de Catalunya & Iquadrat Informatica S.L, Spain); Swastika Roy (Universitat Politècnica de Catalunya & Iquadrat Informatica SL, Spain); Kostas Ramantas (Iquadrat Informatica, Greece); Christos Verikoukis (University of Patras, Greece)
The growing complexity of next-generation networks makes intelligent and adaptive management essential for O-RAN-based network slicing. Traditional Reinforcement Learning (RL) solutions in xApps automate resource allocation but struggle to incorporate high-level operator intents into their optimization process. To address this limitation, we propose an LLM-guided RL framework that bridges the gap between human intent and autonomous network control. In this approach, the LLM-based agent in the rApp interprets operator prompts and contextual network information to dynamically adjust the reward function of the RL agent operating in the xApp, ensuring that resource allocation decisions across slices reflect strategic objectives. The simulation results show that the proposed LLM-guided RL approach achieves higher cumulative reward while maintaining closer alignment with the operator objectives, creating the basis for more intelligent and dynamic O-RAN management.
Decentralized Spatial Reuse Optimization in Wi-Fi: An Internal Regret Minimization Approach
Francesc Wilhelmi, Boris Bellalta, Miguel Casasnovas, Aleksandra Kijanka and Miguel Calvo-Fullana (Universitat Pompeu Fabra, Spain)
Spatial Reuse (SR) is a cost-effective technique for improving spectral efficiency in dense IEEE 802.11 deployments by enabling simultaneous transmissions. However, the decentralized optimization of SR parameters—transmission power and Carrier Sensing Threshold (CST)—across different Basic Service Sets (BSSs) is challenging due to the lack of global state information. In addition, the concurrent operation of multiple agents creates a highly non-stationary environment, often resulting in suboptimal global configurations (e.g., using the maximum possible transmission power by default). To overcome these limitations, this paper introduces a decentralized learning algorithm based on regret-matching, grounded in internal regret minimization. Unlike standard decentralized “selfish” approaches that often converge to inefficient Nash Equilibria (NE), internal regret minimization guides competing agents toward Correlated Equilibria (CE), effectively mimicking coordination without explicit communication. Through simulation results, we showcase the superiority of our proposed approach and its ability to reach near-optimal global performance. These results confirm the not-yet-unleashed potential of scalable decentralized solutions and question the need for the heavy signaling overheads and architectural complexity associated with emerging centralized solutions like Multi-Access Point Coordination (MAPC).
Empowering Next-Generation AI Through Cognitive Cloud-Edge-IoT Continuum: Architecture, Management, and Challenges
Engin Zeydan (CTTC, Spain); Yekta Turk (Aselsan, Turkey); Kapal Dev (Munster Technological University, Ireland)
The increasing complexity and diversity of Artificial Intelligence (AI) workloads, from generative models to real-time decision-making systems, require a transition from traditional centralized architectures to a seamless, cognitive cloud-edge-IoT continuum. This paper presents a comprehensive and unified reference architecture for AI-driven orchestration and infrastructure management across heterogeneous computing environments. We introduce novel mechanisms to optimize AI training and inference workflows across distributed edge, cloud, and high-performance computing (HPC) infrastructures, while ensuring trust, privacy, and energy efficiency. Our approach integrates virtualization, federated learning, data compression, and conditional computing techniques to support scalable, secure, and context-aware AI applications. Finally, we analyze and classify the main architectural and operational challenges involved in enabling AI-native processing across a heterogeneous computing continuum.
