With the rapid evolution toward 6G dynamic edge networks featuring massive device connectivity and ultra-low latency requirements, the contradiction between diverse highdensity concurrent AI services and limited edge resources intensifies node failure risks. Existing fault-tolerant schemes struggle to adapt due to insufficient scalability and high operational costs. To address this, this paper proposes DRIFTER, a diffusionbased proactive migration framework for fault-tolerant AI service migration in edge networks. The framework employs a hybrid Graph Attention Network (GAT)-Transformer architecture for spatiotemporal fault diagnosis and integrates conditional diffusion models with Q-guided single-step inference to balance generation quality and efficiency. The online replay mechanism ensures robust adaptation to dynamic environments. Experiments demonstrate that DRIFTER achieves stable training over GAN-based methods and outperforms baselines with 14.8% latency reduction, 8.7% energy savings, 20.9% lower CPU variance and 17.8% fewer migration operations, while Q-guided inference delivers 42.2% faster scheduling than the ablated version, validating its effectiveness for intelligent fault tolerance in 6G dynamic edge networks.

Cross-domain computing power networks require joint optimization of request routing and computing resource allocation under dynamic conditions and administrative isolation. To this end, we formulate the problem as a partially observable Markov decision process and propose a privacy-preserving Transformer-MAPPO framework operating under the centralized-training-decentralized-execution (CTDE) paradigm. The proposed method jointly determines compute-domain selection, inter-domain skeleton routing, intra-domain refinement, and computing resource allocation, while guaranteeing runtime feasibility through candidate filtering and projection-based action correction. An SRv6-compatible realization is further developed to map the learned decisions into executable segment lists. Experimental results show that, compared with MAPPO, DDPG, and Greedy, Transformer-MAPPO achieves 4.16% reduction in average delay, 32.81% reduction in energy consumption, 13.38% reduction in maximum link utilization (MLU), and 44.89% reduction in SLA violation rate.

In edge computing power network environments, there is a growing demand to support compute-intensive Agentic AI Services, which are composed of interdependent functions represented as Directed Acyclic Graphs (DAGs). Nevertheless, the challenges posed by dynamic resource volatility and potential node failures significantly impact reliable task execution. Existing solutions (often reactive heuristics or GAN-based models) struggle to anticipate risks and overlook DAG dependencies. This paper introduces GADP, a Graph-Aware Diffusion Policy framework designed to facilitate proactive fault-tolerant DAG workload migration in large-scale edge computing systems. This paper presents GADP, a Graph-Aware Diffusion Policy framework for proactive, fault-tolerant DAG workload migration in large-scale edge systems. GADP integrates three key modules: a Transformer-GAT fault predictor for failure probability and type estimation; a DAG encoder that learns structure-preserving task embeddings via multi-round attention; and a diffusion policy generator that refines placement strategies through conditional denoising. Experiments on dynamic simulations with real workload traces show that GADP achieves 99.6% fault detection accuracy, 95.4% diagnosis F1, and over 60% fewer SLO violations, while consuming the least energy among baselines. These results demonstrate GADP's robustness and effectiveness in anticipatory migration under volatile edge conditions.

Driven by the low-altitude economy representing emerging productive forces, the Low-Altitude Internet of Intelligences (LAIN) emerges as an intelligent airspace architecture. Addressing poor adaptability and slow response in traditional strategies under dynamic distributions, this paper proposes a dynamic node deployment method based on the Actor– Critic deep reinforcement learning framework to maximize throughput. The established system model comprises ground stations, communication nodes, and users, adopting power-domain Non-Orthogonal Multiple Access (NOMA) to enhance spectrum efficiency and multi-user transmission. Formulating the problem as a Markov Decision Process (MDP), the proposed AC-based algorithm enables efficient decision-making in continuous action spaces, adaptively optimizing node positions and power. Simulations demonstrate throughput improvements of 38.18% and 50.38% over heuristic and DQN algorithms, respectively. Furthermore, the method exhibits superior stability under uniform and Gaussian distributions, maintaining performance advantages in high-density scenarios up to 40 users/km2.

As emerging services such as AI datacenter interconnection, immersive AR/VR, and autonomous driving demand greater transport flexibility and resource efficiency, Metro Transport Networks (MTNs) have become the critical infrastructure for next-generation metro backbones providing flexible and efficient resource orchestration. Nevertheless, the Routing and Slot Allocation (RSA) problem in MTNs remains challenging due to combinatorial complexity, high-dimensional coupling in the action space, and resource fragmentation. We propose EFP-RSA, a novel entropy-regularized reinforcement learning framework that addresses the RSA problem through factorized policy. It factorizes the joint optimization into coordinated path selection and slot assignment, alongside entropy regularization annealing to balance exploration with convergence. Compared with baselines, EFP-RSA reduces blocking probability by $\mathbf{3 5. 3 \% - 6 0. 7 \%}$ and increases slot utilization by $\mathbf{1 9. 4 \% - 5 8. 6 \%}$, while maintaining low load variance and high solution efficiency, demonstrating strong generalization and scalability in complex MTN scenarios.

Task scheduling for multimodal large language model (MLLM) applications in edge-cloud continuum faces critical challenges, as they simultaneously balance system delay, carbon emissions, and model accuracy requirements, while adapting to network conditions and varying energy availability. To address this, we formulate a joint optimization problem integrating MLLM task scheduling, resource allocation, and green energy utilization to minimize system delay and carbon emissions while meeting strict accuracy requirements. A Green-Aware MAPPO scheduling approach is proposed, which fuses graph attention networks (GAT) with multi-agent proxy policy optimization (MAPPO) to enable distributed decision-making in edge-cloud continuum. By modeling the problem as a partially observable Markov decision process (POMDP), the approach allows edge or cloud servers to capture complex resource dependencies through relation-specific attention mechanisms, even with limited local observations. Experiments across diverse network configurations show that compared with baseline algorithms such as MADDPG, MADQN, and DDPG, the proposed method reduces system delay by 28.3%-66.7%, energy consumption by 17.9%-47.4%, carbon emissions by 11.6%-63.5%, and overall SLA violation rate by 22.21%-73.7% and maintaining MLLM task accuracy above the required threshold. This work provides an effective solution for balancing performance and sustainability in MLLM-enabled edge-cloud continuum.

Recent progress in remote sensing (RS) technologies has led to an explosive growth of imagery data. Since the restricted capacity of satellite-ground links imposes severe constraints on real-time task delivery, onboard satellite compression technology is adopted. However, existing compression scheme fails to fully exploit the onboard computing resources of satellites, thereby inducing a misalignment between compression decisions and system resources. To address this issue, we propose a Hierarchical Resource-Aware Compression (HRAC) scheme and formulate an optimization problem to jointly optimize compression decisions and computing resource allocation for minimizing the transmission delay. The problem is solved within a reinforcement learning framework, and a Marginal Impact Prediction (MIP) network is introduced to overcome the challenge of delayed rewards. Simulation results demonstrate that our proposed algorithm achieves approximately 10%, 23% and 31% reduction in latency compared with the benchmark schemes.

Large language models (LLMs) have advanced the Text-to-SQL task significantly. However, existing methods still face challenges in handling natural language ambiguity, limited dataset content, and redundant database schema information. In this paper, we propose SRSPSQL, a dual-stage Text-to-SQL framework that integrates question semantic reconstruction and dynamic schema pruning to eliminate natural language ambiguity and minimize schema search space. Specifically, we design a novel semantic reconstruction mechanism leveraging LLMs’ summarization capabilities and concrete database content to eliminate ambiguities and align questions with database structures. Moreover, we develop a full-schema prompt strategy that adopts masked similarity matching for context-aware example selection and integrates database content information to minimize value errors. Furthermore, we propose a concise schema generation method that combines structured SQL information with dynamic schema pruning to focus on key schema and examples, thereby significantly improving the accuracy and efficiency of SQL generation. Experimental results demonstrate that SRSPSQL consistently outperforms baselines, achieving 86.4% and 62.13% execution accuracies (EX) on Spider and Bird, with robust gains on Spider-Syn and Spider-Realistic. Ablation studies further validate the individual and synergistic contributions of each component.

In the era of 5G and beyond, the hard-isolated channels enabled by time slot cross-connects in metro transport network (MTN) effectively meet the demands of emerging network services for low latency, low jitter, flexible bandwidth, and secure isolation. However, the dynamic arrival and departure of tenant virtual network request (VNRs) lead to resource fragmentation within the MTN transport network, resulting in inefficient resource utilization. To mitigate network resource fragmentation, we formulate the MTN dynamic channel orchestration problem and propose a fragmentation-aware MTN dynamic channel orchestration method. This method comprises two key components: a greedy graph-reconstruction-based channel mapping algorithm and a fragmentation-aware channel reconfiguration algorithm. The former optimizes MTN channel resource allocation to achieve static channel orchestration, while the latter, leveraging a simulated annealing-based channel reconfiguration strategy, dynamically adjusts channel allocations based on the static orchestration results, thereby reducing fragmentation levels. Compared to existing approaches, under varying network load conditions, the proposed channel mapping algorithm reduces the consumption of network time slot resources by 16.3% -20.6%, while the channel reconfiguration algorithm significantly lowers the levels of network fragmentation by 48.7% -79.5%, and reduces the running time by 91.3%–94.2% compared with the baseline.

Task allocation in computing power networks faces challenges such as complex task dependencies, multi-dimensional and heterogeneous resources with uneven distribution, and difficulty in global collaboration among large-scale nodes. Traditional single agent algorithms were difficult to cope with the dynamic resource collaboration issues in wide area environments, and often relied on global state information, while existing multi-agent reinforcement learning algorithms lacked targeted modeling of task dependencies and network topology, resulting in low global collaboration efficiency. To address these issues, a distributed collaborative allocation algorithms of dependency-constrained tasks in computing power networks was proposed. First, complex subtask dependencies were processed through layering, triangular dependency decoupling and multi-resource threshold merging. Then, relation-graph attention network (RGAT) was adopted to encode multi-resource relations and topology features, combined with multi-agent soft actor-critic (MASAC) for distributed optimization. Experimental results show that in multiple scenarios, the proposed algorithms significantly optimizes task completion time and energy efficiency compared to baseline algorithms.

The Digital Twin (DT) represents a rapidly advancing technological innovation within the Industrial Internet of Things (IIoT) domain. DT leverages the power of simulation, machine learning, and data mining to facilitate optimal decision-making for physical objects. However, the creation of a dynamic and living digital counterpart comes at a considerable cost. It requires continuous massive data updating and processing every time the physical object changes. As most data collected by IIoT devices are in their original form, such as images and videos, transmitting such data to remote cloud computing will result in large delays. Furthermore, data processing is often a computationally intensive operation, such as image recognition and video coding, making it impractical to perform processing tasks directly in IIoT devices. To overcome this problem, we introduced the Multi-access/mobile Edge Computing (MEC) architecture to enhance capabilities of DT-enabled IIoT devices. IIoT devices can leverage the extra computing resources in MEC to process raw data, transmitting only the calculation results to update the digital counterpart. To efficiently allocate resources between IIoT devices and MEC, we propose a double auction-based resource allocation scheme. The IIoT devices can purchase computing power from MEC, and an iterative double auction scheme is applied to achieve system efficiency within this market. Furthermore, we propose the Win or Learn Fast Algorithm Policy Hill Climbing (Wolf-PHC) algorithm, which enables agents to improve their strategies continuously through participation in auctions. Simulation results demonstrate that this algorithm accelerates the process of market equilibrium convergence.

The Mongolian Plateau grassland (MPG) is critical for ecological conservation and sustainability of regional pastoral economies. Aboveground net primary productivity (ANPP) is a key indicator of grassland health and function, which is highly sensitive to variabilities in large-scale atmospheric circulations, commonly referred to as teleconnections (TCs). In this study, we analyzed the spatial and temporal variations of ANPP and their response to local meteorological and large-scale climatic variabilities across the MPG from 1982 to 2015. Our analysis indicated the following: (1) Throughout the entire study period, ANPP displayed an overall upward trend across nine ecoregions. In the Sayan montane steppe and Sayan alpine meadow ecoregions, ANPP displayed a distinct inflection point in the mid-1990s. In the Ordos Plateau arid steppe ecoregion, ANPP continuously increased without any inflection points. In the six other ecoregions, trends in ANPP exhibited two inflection points, one in the mid-1990s and one in the late-2000s. (2) Precipitation was the principal determinant of ANPP across the entire MPG. Temperature was a secondary yet important factor influencing ANPP variations in the Ordos Plateau arid steppe. Cloud cover affected ANPP in Sukhbaatar and central Dornod, Mongolia. (3) The Atlantic Multidecadal Oscillation affected ANPP by regulating temperature in the Ordos Plateau arid steppe ecoregion, whereas precipitation occurred in the other ecoregions. The Pacific/North America, North Atlantic Oscillation, East Atlantic/Western Russia, and Pacific Decadal Oscillation predominantly affected precipitation patterns in various ecoregions, indicating regional heterogeneities of the effects of TCs on ANPP fluctuations. When considering seasonal variances, winter TCs dominated ANPP variations in the Selenge–Orkhon forest steppe, Daurian forest steppe, and Khangai Mountains alpine meadow ecoregions. Autumn TCs, particularly the Pacific/North America and North Atlantic Oscillation, had a greater impact in arid regions like the Gobi Desert steppe and the Great Lakes Basin desert steppe ecoregions. This study’s findings will enhance the theoretical framework for examining the effects of TCs on grassland ecosystems.

With the rapid development of the Internet of Things and machine learning, deep neural networks have been widely applied and developed in multiple fields. At the same time, thanks to the development of mobile edge computing, the deployment of DNN reasoning tasks sinks from the remote central cloud to the edge side. However, the failure probability of edge computing nodes is much higher than that of central cloud. Therefore, when an edge server fails, how to better reconstruct and resume the execution of DNN inference tasks is a key issue. To address the aforementioned issues, we propose an edge fault oriented DNN inference task reconstruction method. Firstly, we established a problem model for the redeployment of edge fault oriented DNN inference tasks based on cloud edge network environment. The problem model optimizes the reliability and delay weights of the reconstructed DNN inference tasks. Then we designed a meta heuristic method based on simulated annealing algorithm to find the optimal redeployment scheme for DNN inference tasks. Finally, we evaluated the performance of our proposed algorithm through simulation. The results showed that our algorithm outperformed Random, LB, and RMA algorithms by an average of 36.4

Large bearer networks are often managed by different operators, with each network forming an autonomous domain. These domains together constitute a multi-domain network. Metro transport network (MTN) is a new bearer network technology developed for 5G metropolitan bearer requirements. For the routing and time-slot allocation (RSA) problem in MTN, traditional algorithms are often based on heuristic rules or limited to single-domain environment, making them difficult to apply to more complex multi-domain scenarios. In view of the above issues, this paper proposes a hierarchical deep reinforcement learning (DRL) framework based on the asynchronous advantage actor critic (A3C) algorithm to solve the multi-domain MTN routing problem. And a time-slot allocation algorithm is proposed to optimize the time-slot resource utilization of the links. Compared to other classic heuristic algorithms, the proposed algorithm achieves an overall improvement of at least 42.1% in terms of the combined metrics of blocking rate and network load balancing, while the average utilization of link time-slot resources increases by 1.5%.

Personalized federated learning (PFL) enables collaborative model training across devices while adapting to heterogeneous data, but faces resource constraints on edge devices. Combining PFL with pruning techniques helps address these constraints. A challenge is that one-size-fits-all pruning strategies may ignore the varying importance of parameters for local data. To overcome this, we propose PLFL, a novel personalized lightweight federated learning framework. PLFL uses a hypernetwork at the server level to deliver personalized local models to clients and incorporates a federated pruning mechanism tailored to parameter importance, ensuring optimal performance and maintaining personalization. Experimental results show that PLFL achieves higher accuracy with lower computational costs and fewer parameters compared to state-of-the-art methods on heterogeneous datasets.

Time-Sensitive Networking (TSN) is a set of data link layer protocol specifications developed by the IEEE 802.1 Working Group to provide low latency and high reliability network communication for real-time applications. In order to facilitate research and testing validation of Time-Sensitive Networking, we have designed and developed a TSN simulation system based on NS3. This system simulates gate list configuration and traffic scheduling in time-sensitive network. In addition, the system supports dynamic changes to the topology and flow configuration while the network is running. It includes the following main functionalities: 1. Simplified gate list configuration. 2. Support for dynamic networks. 3. Support for embedding various traffic scheduling algorithms. After experimental verification, our system can support the verification and testing of TSN scheduling algorithm, and effectively improve the development efficiency of researchers.

Task scheduling decisions for multimodal large language model (MLLM) applications in multilayer computing power networks present a significant challenge, as they simultaneously balance system delay, carbon emissions, and model accuracy requirements while adapting to network conditions and varying energy availability. Thus, in this paper, we formulate the joint optimization problem of MLLM task scheduling, resource allocation, and green energy utilization to minimize system delay and carbon emissions while meeting accuracy requirements. We propose Green-Aware MAPPO, a novel approach that integrates graph attention networks (GAT) with multi-agent proximal policy optimization (MAPPO) for distributed decision-making in multilayer computing power networks. By modeling the problem as a partially observable Markov decision process (POMDP), our algorithm enables agents to capture complex resource dependencies through relation-specific attention mechanisms while maintaining high performance with limited local observations. Experiments in various network configurations demonstrate that Green-Aware MAPPO significantly outperforms baseline algorithms.