Semantic communications (SemCom) provide efficient transmission for mobile edge computing (MEC) services by extracting critical semantics from raw information. Although widely adopted in various scenarios, existing single-modal SemCom systems struggle to efficiently support edge multimodal data transmission. Additionally, mobile end users have varying communication requirements across different modalities. However, existing work lacks the ability to generate personalized communication policies tailored to diverse intents (Typically, communication policies include bandwidth allocation and modulation and coding schemes, etc.). In this paper, we propose an edge Cognitive SemCom Agent (CSCA) to facilitate edge multimodal SemCom. Specifically, CSCA leverages an edge Large AI Model (LAM) to realize modality alignment and natural language intent understanding. Moreover, we develop a communication planning module to realize the planning capability, which generates personalized wireless communication policies based on LAM’s environment and intent cognition. Particularly, to assess the efficiency of communication policies in multimodal SemCom and capture intent competition, we present a novel indicator named cognitive SemCom quality indicator (CSCQI). Then, we use the denoising diffusion probabilistic model to optimize the generation policy. Extensive experimental results demonstrate that CSCA achieves an average improvement in intent satisfaction rate and semantic accuracy by 42.19% and 29.75% respectively, while reducing communication delay by 33.40% .

In cell-free multiple-input multiple-output (MIMO) systems, beam alignment is critical to achieving high spectral efficiency and reliable communication. However, traditional optimization-based methods often suffer from high computational complexity and sensitivity to dynamic channel conditions, especially in decentralized architectures with distributed access points (APs) and mobile users. To address these challenges, a novel scheme for beam alignment optimization that leverages generative diffusion models (GDMs) is proposed in this paper. By learning the underlying distribution of optimal beamforming configurations from historical channel state information (CSI) and user positioning data, the proposed approach generates high-quality precoding and combining matrices that minimize beam alignment errors while maximizing signal-to-noise ratio (SNR). The system model integrates a conditional diffusion process, where CSI and user locations serve as input to guide the generation of beamforming solutions. The framework operates in two phases: an offline training stage that learns the latent distribution of optimal beam alignment, and an online inference stage that rapidly adapts to real-time channel variations. To ensure practicality, the design incorporates power constraints and feedback mechanisms to dynamically refine beam configurations. Simulations demonstrate significant improvements in beam alignment accuracy and communication performance, particularly in high-mobility scenarios.

Artificial Intelligence Generated Content (AIGC) technology, with its highly efficient automation and intelligent algorithms, is transforming the way information is produced. However, the resource-intensive nature of large generative AI models, combined with traditional cloud-based inference approaches, leads to significant bandwidth consumption and unpredictable communication latency. On the other hand, directly deploying AIGC models on edge nodes faces challenges such as limited computational resources and storage space. To address these issues, we proposes an AIGC inference framework based on edge node collaboration, and develops an optimization model using model splitting methods. Building on this, we also designed a deep reinforcement learning algorithm incorporating Graph Attention Networks (GAT), which adaptively adjusts model partitioning points and resource scheduling strategies to effectively deploy AIGC models. Finally, simulation results show that the proposed algorithm improves task success rates by an average of 20% compared to Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO) algorithms, achieving a balance between computation and communication latency in edge environments.

With the rise of AI-Generated Content (AIGC) services in wireless networks, efficient and high-quality distributed training of Large Language Models (LLMs) has become essential for enabling the large-scale application of next generation AI technologies. However, the extensive parameters of LLMs impose significant demands on memory, computing power and communication resources in heterogeneous networks. To efficiently utilize the dispersed network resources, this paper presents a First-Pipeline- Then-Federated Learning (FPTFL) approach with a hybrid parallel scheduling strategy to facilitate the training of Transformer-based LLMs. We propose a block-wise splitting mechanism to partition the Transformer's encoder into distinct segments, which are deployed cross individual devices. The encoder parameters and intermediate smashed data are uploaded to the edge server, where the whole model is updated through federated aggregation. Particularly, we develop a fine-grained computation-efficient method based on pipeline parallelism, enabling the segments to cooperatively train the entire encoder. An optimization problem is formulated to determine the LLM segments and the number of micro-batches under network resource constraints, with the goal of minimizing the total latency of LLM training services. Simulation results demonstrate that our approach enables Transformer-based model training on resource-constrained devices, preserves model performance, and reduces waiting time.

The distribution of computing resources in the Computing Power Network (CPN) is uneven, leading to an imbalance in resource supply and demand within domains, necessitating cross-domain resource scheduling. To address the cross-domain resource scheduling challenge in CPN, this paper presents an Improved Multi-Agent Actor-Critic (IMAAC) resource scheduling approach leveraging Data Processing Unit (DPU) offloading. Initially, we introduce a cross-domain resource scheduling architecture tailored for CPN by leveraging DPU offloading. Specifically, we delegate certain functionalities of the Multi-Agent Deep Reinforcement Learning (MADRL) Agent to DPUs, aiming to mitigate communication costs incurred during the generation of cross-domain scheduling decisions. Second, we introduce the parallel experience ensemble and multi-head attention mechanism in the Multi-Agent Actor-Critic (MAAC) framework to compress the state-space dimensionality of agent association across domains. Finally, we introduce the parallelized dual-policy network structure to mitigate training instability and convergence challenges within the actor and critic networks. Experimental results showcase that IMAAC achieves noteworthy reductions of 5.98%~13.56%, 23.54%~33.55%, and 41.17%~58.88% in total system delay, energy consumption, and the number of discarded tasks, respectively, compared to benchmark experiments.

Foundation models (FMs), known as pre-trained models, have garnered significant interest in Industrial Internet due to their remarkable performance and robust generalization capabilities in downstream tasks. However, with the increasing requirements of computing infrastructure and data privacy protection for large foundation models, existing learning frameworks face challenges such as data privacy leakage, poor scalability, and deployment difficulties. To address these issues, this paper proposes a novel collaborative Transformer Block (TB)-wise training framework based on Federated Learning (FL), which consists of three stages: pre-training, graph regularization, and personalized training. To tackle the challenge of statistical heterogeneity in distributed data, we design a Graph Convolutional Network (GCN)-based update operator that captures local training representations. Besides, we conduct an analysis based on feature similarity to enhance the interpretability of our algorithm. We choose popular vision Transformer models for the experiments, extensive results demonstrate that our framework can jointly train multiple clients to build a foundation model while improving the single client's personalized performance. The proposed method outperforms state-of-the-art frameworks under various data distributions and system heterogeneity settings, highlighting its robust performance.

Aiming at the ultra-low latency service demand of power emergency Internet of Things (PEIoT), a multi-slice network architecture for ultra-low delay transmission of emergency Internet of Things was designed, and a PEIoT slice resource reservation and multi-heterogeneous slice resource sharing framework was proposed. The proposed framework adopts the deep reinforcement learning method to realize the automatic prediction and allocation of real-time resource requirements among heterogeneous slices. Simulation results show that the method based on resource reservation enables PEIoT slice to explicitly retain resources and provides a better level of security isolation. Deep reinforcement learning can ensure the accurate and real-time update of resource reservation and effectively consider the resource utilization rate and the differentiated service quality requirements of slices. The comparison with two existing algorithms shows that Dueling DQN has better performance advantages.