Rapid advances in large language models (LLMs) have enabled agentic artificial intelligence (AI) with strong reasoning and autonomous decision-making capabilities. Bringing these capabilities closer to users motivates Mobile Edge General Intelligence (MEGI), which can reduce cloud-offloading latency and keep data closer to its source. However, deploying LLM-based agentic AI reasoning in MEGI environments poses significant challenges due to the high computational demands of reasoning and the limited resources of edge devices. To address these challenges, we propose a joint optimization framework for efficient LLM reasoning deployment in MEGI. First, we systematically review enhancement methods to identify mechanisms suitable for edge adaptation. Subsequently, we present a distributed framework that synergizes reasoning enhancement via adaptive CoT prompting with scalable deployment through a distributed MoE architecture. The framework models reasoning depth as a dynamic network resource variable and optimizes it jointly with task-level expert activation and transmission power. This design adapts the active experts and reasoning complexity to task requirements and device capabilities. System-level evaluations show that dynamic CoT reduces energy consumption relative to the dense baseline while balancing accuracy and latency satisfaction.
In complex electronic warfare scenarios for future communication systems, radar jamming recognition faces severe challenges. Especially under low signal-to-noise ratio (SNR) conditions, the time frequency (TF) characteristics of jamming signals are easily submerged by noise. Traditional convolutional neural networks exhibit low recognition accuracy in such harsh environments, which limits their deployment in next-generation intelligent sensing systems. Our paper analyzes several typical jamming recognition scenarios and the corresponding recognition workflow. To effectively enhance feature robustness against noise, singular value decomposition (SVD) low rank approximation is employed to extract inherent signal features. On this basis, an end-to-end radar jamming recognition framework named ST-DCNet is proposed. Specifically, a dual-branch feature extraction structure embedded with the convolutional block attention module (CBAM) is constructed to strengthen the representation of critical jamming features. Furthermore, a dynamic weighted multi-loss function is designed by fuzing cross-entropy loss, center loss, and cosine embedding loss, which improves the intra-class compactness and inter-class separability of deep features. We argue that this method offers a novel perspective for radar jamming recognition in low-SNR scenarios.
The rapid scaling of massive multiple-input multiple-output (MIMO) systems toward future gigantic MIMO arrays imposes prohibitive power and hardware bottlenecks on conventional digital beamforming designs. To overcome these limitations, this paper introduces the microwave linear analog computer (MiLAC) into the front-end of massive MIMO systems, establishing a comprehensive cross-layer framework that unifies algorithmic design with physical hardware implementation. We first formulate the system model, detailing a systematic mapping from the target digital-domain combining matrix to the tunable electromagnetic admittance network, thereby realizing sophisticated beamforming directly in wave domains. Furthermore, to compensate for the inevitable non-idealities of practical analog components, we propose a novel hardware-in-the-loop adaptive calibration mechanism. Specifically, driven by gradient back-propagation, this closed-loop scheme iteratively mitigates admittance programming errors via pilot probing and error feedback. Case studies demonstrate that this strategy consistently minimizes the hardware-to-algorithm mismatch, translating enhanced matrix-level precision into substantial system-level sum-rate gains. Finally, we present several key challenges and research directions for future MiLAC-enabled wireless networks.
Collaborations among distributed on-device Large Language Models (LLMs) can overcome the limited capability of a single model, yet prevailing token-level exchanges suffer from additional pre-fill computation, and inherent information bottleneck. To address these challenges, this article advocates a paradigm shift from discrete tokens to KV caches, the rich continuous-space attention states that LLMs already maintain, as the native communication medium for cross-model collaboration. We systematically analyze why the token interface is insufficient on its own for efficient federated LLM inference and present FedCacheInf, a cache communication-enabled framework that encompasses cache projection for cross-model compatibility, flexible cache integration strategies, and compression techniques. Diverse numerical results demonstrate the superiority of cache communications over token-based approaches in accuracy and efficiency, while identifying the regimes in which token exchange remains preferable. We further outline open research directions to advance cache-centric LLM collaboration toward next-generation AI-native networks. By exploring the LLM-native communications, we wish to provide a new paradigm for this broad area.
Maritime Internet of Things (IoT) deployments increasingly rely on the space-air-ground-sea integrated network (SAGSIN) to connect underwater sensors with terrestrial and space backbones. However, the heterogeneous links along this path, ranging from bandwidth-limited and energy-hungry underwater acoustic channels to high-capacity optical links above the sea, make the transport of massive raw sensing data costly in both latency and energy, and difficult to sustain for unattended, battery-powered nodes. This article presents an edge-computing paradigm for SAGSIN built on two coupled ideas: a Multi-Layer Network Architecture (MLNA) that organizes the underwater, surface, aerial, and ground/space tiers, and Multi-Level Information Processing (MLIP) that progressively refines data from raw measurements toward compact, event-level representations as they ascend the network. We characterize the computation, transmission, and storage energy at each tier and show how distributing refinement across layers trades local processing cost against transmission and storage savings. We then discuss how MLNA-MLIP reduces latency, strengthens data privacy, improves service reliability, and manages energy to prolong network lifetime. A case study on offshore monitoring quantifies the resulting lifetime gains and identifies the optimal processing depth. Open challenges and future directions are outlined.
