The performance of large language model (LLM)-based multi-agent systems (MAS) largely depends on effective communication topologies. Existing topology generation methods, however, typically learn communication topologies through black-box optimization driven solely by task-level rewards. While effective, such optimization provides little insight into why particular communication edges are selected, making it difficult to identify the critical communication subgraphs responsible for successful collaboration. To address this limitation, we propose E2-Explainer, a model-agnostic framework for providing interpretable explanations of communication topologies produced by arbitrary topology generators. Specifically, we formulate topology explanation as a causal attribution problem that identifies compact communication subgraphs supported by edge-level evidence of task preservation. We obtain this evidence with a Granger-style objective that measures how masking each communication channel changes the task outcome and the stability of the final response. The resulting budgeted subgraphs are then distilled into an amortized explainer, enabling efficient post-hoc explanation without repeated edge-level evaluations at deployment. Extensive experiments on multiple reasoning and coding benchmarks demonstrate that E2-Explainer identifies critical communication subgraphs that preserve successful collaboration. These subgraphs can also be executed directly to prune redundant communication edges, substantially reducing communication costs while maintaining competitive task performance.

Up to date, the application of digital twin (DT) in the Industrial Internet of Things (IIoT) has been continuously promoted and deepened and has become the focus of the industry. IIoT serves as the foundational infrastructure that enables pervasive connectivity, real-time data acquisition, and intelligent control within industrial environments. DTs provide enterprises with an empathetic, virtual environment that enables them to manage and operate their production facilities in a more efficient and intelligent manner. However, there is not a special summary and analysis of the combinability and combination mode of the two. Therefore, this article first sorted out the professional definitions, characteristics, and frameworks of IIoT and DT, and deeply analyzed the semantic context of data flow. Second, this article discusses the combinability and combination mode of IIoT and DT and summarizes the enabling technologies and tools at each layer. Finally, the applications status of DT empowered IIoT in different fields was summarized, and the challenges of the combined application of the two were analyzed.

This paper aims to improve energy efficiency (EE) of the integrated access and backhaul (IAB) aerial-terrestrial network, facilitating rapid and adjustable network infrastructure deployment. This is challenging, as interference generated by backhaul and access links degrades network throughput, and power imbalance between these links increases overall energy consumption. To this end, we jointly optimize aerial base station (ABS) deployment, user association, and downlink power allocation for both terrestrial base station and ABSs to maximize network EE. Specifically, using fractional programming, the EE maximization problem is transformed into a subtractive-form parametric problem, and then decomposed into ABS deployment and resource allocation subproblems. A hybrid algorithm combining particle swarm optimization and simulated annealing is proposed to solve the ABS deployment subproblem, determining ABS spatial configurations and updating power allocation given fixed user association. Meanwhile, a dynamic power allocation in response to network load is designed to solve the resource allocation subproblem. Furthermore, considering the quality of service requirements of ground users and the transmit power constraints of base stations, a joint EE optimization algorithm is proposed to enhance the network EE. Simulation results validate the effectiveness of the proposed methods in improving network EE, especially in scenarios involving more deployed ABSs.

The explosive growth of data generated by Internet of Things applications has driven large-scale distributed machine learning (DML) on IoT cloud and edge cloud platforms. DML workloads in modern AI clusters generate a large amount of bursty and iterative communication traffic, making communication a major bottleneck affecting DML job completion. Therefore, flow scheduling is a critical approach for optimizing DML. In practice, AI clusters need to handle jobs with different latency requirements simultaneously, including Service Level Objective (SLO) jobs with strict deadlines and best-effort jobs without deadlines. However, existing DML flow schedulers ignore deadline requirements, resulting in low deadline satisfaction rates. We observe that communication contention among different jobs is a key factor causing jobs to fail to meet deadlines. In this paper, we propose Allot, a deadline-aware flow scheduler that mitigates the impact of inter-job communication contention on the deadline satisfaction rate of jobs. Allot leverages the periodicity and predictability of DML workloads to infer the job completion time and the spatial distribution of traffic. Based on this information, Allot performs admission control and adopts the Contention-Penalty Earliest Deadline First (CP-EDF) strategy to prioritize flows. We further introduce an efficient priority compression algorithm to adapt to limited priority levels on practical NICs and switches. Simulation results show that Allot can improve the job deadline satisfaction rate by an average of 1.13× to 5.03× compared to existing solutions.

The Mixture of Experts (MoE) architecture is an important method for scaling Large Language Models (LLMs). It increases model capacity while keeping computation cost low. However, the ultra-large MoE models still have hundreds of billions of parameters, requiring massive memory/storage and leading to difficulties for deployment on resource-constrained edge platforms. Pruning or quantization alone can hardly address the issue, because of the super-aggressive compression ratio with significantly degraded accuracy and output quality. To facilitate the deployment of ultra-large MoEs on edge platforms, we propose a collaborative compression framework by combining expert pruning, mixed-precision quantization, and activation optimization. It can effectively reduce the storage footprint of the ultra-large MoE DeepSeek-V3 from 1.3TB to 103GB, while preserving high output quality with better accuracy than traditional uniform low-bit quantization methods. To the best of our knowledge, we are the first to deploy a compressed model from the ultra-large DeepSeek-V3 on the platform with a strict 128GB total memory limit. Our comprehensive experiments on multiple benchmarks under various memory constraints demonstrate the effectiveness of our method with smaller model sizes and higher accuracy than uniform low-bit quantization methods.

Small reservoirs are highly vulnerable to rapid cyanobacterial blooms, yet routine monitoring is constrained by the spatial–temporal trade-off of single-sensor satellite observations. To address this limitation, we fused Sentinel-2 MSI (10–20 m, ∼5 d) and Sentinel-3 OLCI (300 m, daily) data using the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) to generate a Sentinel-2-like dataset (10 m, daily) for Zhuwei Reservoir during the ice-free seasons (April–November) of 2016–2024. Validation against coincident Sentinel-2 scenes showed good agreement between the fused and reference reflectance fields (RMSE < 0.015 sr-1, PSNR > 37 dB, SSIM > 0.8). After excluding oligotrophic and severe-bloom samples that were not used for quantitative retrieval, the remaining 93 paired samples collected during ten field campaigns in 2023–2024 were used to compare several established Chlorophyll-a (Chl-a) retrieval algorithms. The Normalized Difference Chlorophyll Index (NDCI) showed the best performance and was therefore selected as the most suitable algorithm for moderate-bloom conditions (R2 = 0.72, RRMSE = 26.2%, and MAE = 11.89 μg/L). Application of the selected model to the fused record revealed a weak overall decline in reservoir-mean Chl-a (−0.313 μg/L/yr), pronounced summer maxima, and persistent hotspots in the northwestern, northern, and northeastern sectors. These results demonstrate the practical value of combining spatiotemporal fusion with locally evaluated Chl-a retrieval for gap-reduced monitoring of bloom dynamics in small reservoirs.

Information freshness plays a pivotal role in time sensitive applications within Mobile Crowd sensing (MCS) net works and is quantitatively characterized by the Age of Information (AoI). This paper considers an MCS network where end users submit requests to an edge server (ES) for status updates monitored by a group of energy-acquiring mobile devices operating with periodic sensing. In such a system, the ES must schedule transmissions among heterogeneous mobile devices while optimizing their sensing cycles to better align with user demands. A critical challenge arises from the mobile devices' limited battery capacity, which is not directly observable by the ES. The objective is to minimize the average weighted sum of required AoI and the energy consumption of the mobile devices. To address this issue, we model the problem as a Partially Observable Markov Decision Process (POMDP) and leverage reinforcement learning (RL) methods to find a solution. Key challenges include sparse rewards and a mismatch between user demands and mobile device sensing frequencies. To address these issues, we propose an innovative Hybrid Offline-Online Reinforcement Learning (HOORL) algorithm, which combines the strengths of offline and online RL techniques. Finally, simulation results show that the proposed HOORL algorithm achieves superior performance compared to relevant scheduling schemes, effectively reducing system costs while maintaining stable battery levels in mobile devices.

The sparsely-activated Mixture-of-Expert (MoE) techniques support scaling the parameter sizes of pre-trained models to the trillion-level without increasing the computational costs. However, in large-scale cloud computing environments, dynamic load imbalance caused by the random expert selection of samples draws a huge challenge for distributed training efficiency. To address these challenges, we propose a lighter-weight and lower communication overhead dynamic load balancing framework, called BalanceMoE, to accelerate MoE model training. BalanceMoE is based on two key novel ideas. Firstly, we model a worker-pair-based expert transfer mechanism that considers the tradeoff between the expert parameter communication and the time reduction obtained. We perform a theoretical analysis and design a highly lightweight algorithm to obtain a near-optimal load balancing solution for per-iteration time reduction. Then, we present our proposed scheme for parallelization of expert computing and transfer, which overlaps the parameter communication of transferred experts and the computing of non-transferred experts to reduce per-iteration training time. We implement BalanceMoE architecture on the PyTorch framework. Extensive experiments on two clusters demonstrate that at training speed, BalanceMoE can achieve up to 1.26x, 1.79x and 2.62x speedup compared to the state-of-the-art SmartMoE, FasterMoE and FastMoE, respectively. At memory usage, BalanceMoE saves up to 71% and 36% of memory compared to FasterMoE and FastMoE, respectively. At energy consumption, BalanceMoE saves up to 13% of the energy consumed within each training iteration compared to SmartMoE. BalanceMoE's code is available at https://github.com/ZJU-CNLAB/BalanceMoE.

With the rapid growth of Internet of Things (IoT) applications, Mobile Edge Computing (MEC) has become an important paradigm for providing low-latency and reliable computing services. Nevertheless, under dynamic and large-scale task arrivals, jointly improving users’ quality of experience (QoE) and maintaining long-term energy efficiency remains a key challenge. To address this challenge, this paper proposes a two-stage online task offloading and resource allocation mechanism, where the first stage optimizes task offloading decisions of mobile users with fixed resource allocation to establish an efficient baseline, and the second stage iteratively adjusts resource allocation according to these decisions to achieve joint optimization. By leveraging Lyapunov optimization, the proposed mechanism can effectively balance energy efficiency and users' quality of experience (QoE) under long-term energy constraint. Simulations conducted on real-world datasets further validate the effectiveness and practicality of the proposed mechanism in the dynamic MEC environment.​

Vertical Split Federated Learning (VSFL) allows participants to collaboratively train a better model with different features vertically partitioned in the same sample space, where the model is divided into bottom model and top model by the cut layer, trained by passive and active participants respectively. However, in the process, the labels owned by the active participant will still be inferred or stolen by curious or malicious passive participants. In this paper, we propose Casper, a causality-inspired defense mechanism with a confounder against label inference attacks in VSFL. Casper first analyzes the feasibility of optimizing the training process in VSFL at the intervention level from a causal perspective. It then introduces a confounder consisting of cut layer output reconstruction and label obfuscation to disrupt the direct causality between cut layer outputs and labels. Additionally, we integrate selective discrepancy training to further ensure model utility by strategically balancing training between active and passive participants. Extensive experiments conducted on four datasets across different tasks demonstrate that Casper effectively preserves label privacy while maintaining model performance, significantly outperforming current advanced defending methods in VSFL.

Driven by the ultra-high throughput requirements of 6G, wireless communications are migrating to centimeter wave (cmWave) bands to overcome the limitations of current spectral resources. Massive multiple-input multiple-output (MIMO) and orthogonal frequency division multiplexing (OFDM) systems aim to achieve high spectral efficiency in cmWave regimes but are often constrained by the heavy overhead of downlink channel state information (CSI) feedback. This paper proposes a deep learning scheme based on the multi-axis multi-layer perceptron for image processing (MAXIM) architecture for joint semantic CSI feedback and hybrid beamforming in multi-user cmWave MIMO-OFDM systems, which maximizes the downlink sum rate by end-to-end optimization. Specifically, distributed encoders at multiple user equipments (UEs) perform limited CSI feedback, while the decoder at the base station (BS) jointly designs the hybrid beamforming matrices without explicit CSI reconstruction. The uplink transmission is implemented via deep joint source-channel coding (DJSCC) to enhance CSI compression efficiency and noise robustness. Furthermore, considering the high correlation between vertical and horizontal polarization channels in dual-polarized massive MIMO systems, a cross-polarization interaction module is introduced at the UEs to exploit polarization correlations for joint CSI compression. Simulation results demonstrate that the proposed method improves the downlink sum rate under various signal-to-noise ratio (SNR) conditions with a limited number of feedback symbols, validating its robustness and superiority in multi-user dual-polarized cmWave MIMO-OFDM systems.

Current congestion control algorithms ignore the application-layer delay, where the untransmitted data waiting at source nodes degrades the delay performance of services. Moreover, differentiated priorities are necessary for the application services with various delay requirements, especially for mission-critical services. Different from the existing works only considering the network delay, in this paper, by adding the flow queueing delay at source nodes, we formulate the network utility maximization (NUM) problem with additional deterministic service delay constraints. We propose a general TCP-based two-timescale congestion window control (TCWC) framework with delay-aware priority to enhance traditional algorithms. Specifically, to handle the obstacle of new delay constraints, we transform them to the time-average stability of virtual queues. By solving the new NUM problem via Lyapunov optimization, we design a short-term congestion window adjustment strategy in each time slot. To further guarantee the service delay, we apply extreme value theory (EVT) to evaluate the priorities of different flows, and determine the long-term control of window update rates. We deploy the proposed framework in three classic algorithms including NewReno, Vegas and DCTCP. In addition, simulation results show that our TCWC framework can significantly reduce the average service delay and provide deterministic guarantees compared with time-aware TCP congestion control algorithms such as TIMELY and BBRv2.

The data-intensive nature of the Internet of Things (IoT) significantly challenges conventional communication systems. While over-the-air computation integrates communication and computation to reduce data aggregation burdens, it requires strict synchronization of transmitted signals, increasing system overhead. Here, we report an optical in-sensor wireless data aggregation paradigm that exploits photocarrier trapping in defect-dominated persistent photoconductance sensors to reduce dependence on synchronization constraints. Within a single communication window, the system reliably aggregates incoherent optical signals with relative timing offsets exceeding the signal duration by up to 120%, while achieving computation errors below 1.5% normalized mean squared error. This process achieves unbiased integration with pulsed low-frequency readout, indicating its potential for ultra-low power operation. A distributed temperature prediction model based on this paradigm shows faster convergence and competitive accuracy with a simplified architecture, demonstrating effectiveness in distributed edge sensing scenarios. This work provides a promising hardware implementation for low-power data aggregation in large-scale IoT edge networks.

The rapid proliferation of intelligent mobile devices has spurred the increasing interest in collaborative learning (CL). Existing CL methods rely on direct algebraic averaging of model parameters for knowledge transfer accross devices, which faces communication bottlenecks, struggles to scale into distributed settings, and has difficulty in handling model and data heterogeneity. To address these challenges, building upon the single-device deep mutual learning (DML), we propose a novel communication framework that interconnects multiple devices to form a natively distributed DML system. Unlike the classic CL, multi-device DML utilizes a distillation loss term to enable models to mutually and indirectly influence each other, thereby sharing knowledge by identifying common optimum point across devices. This mechanism enhances the distributed scalability, fully leverages on-device communication and computation resources, and effectively addresses model and data heterogeneity. We explore the integration of our proposed multi-device DML into a wireless system, termed wireless DML (WDML). Since knowledge sharing is hampered by communication bottlenecks, we analyze the corresponding challenges and opportunities for enhancing learning efficiency. Through a case study on a device-to-device based synchronous peer-to-peer system, we validate the advantages of WDML in energy efficiency and generalization. We conclude by discussing open issues that guide future research towards a more efficient, lower latency, more flexible WDML system.

Communication is essential in coordinating the behaviors of multiple agents. However, existing methods primarily emphasize content, timing, and partners for information sharing, often neglecting the critical aspect of integrating shared information. This gap can significantly impact agents' ability to understand and respond to complex, uncertain interactions, thus affecting overall communication efficiency. To address this issue, we introduce M2I2, a novel framework designed to enhance the agents' capabilities to assimilate and utilize received information effectively. M2I2 equips agents with advanced capabilities for masked state modeling and joint-action prediction, enriching their perception of environmental uncertainties and facilitating the anticipation of teammates' intentions. This approach ensures that agents are furnished with both comprehensive and relevant information, bolstering more informed and synergistic behaviors. Moreover, we propose a Dimensional Rational Network, innovatively trained via a meta-learning paradigm, to identify the importance of dimensional pieces of information, evaluating their contributions to decision-making and auxiliary tasks. Then, we implement an importance-based heuristic for selective information masking and sharing. This strategy optimizes the efficiency of masked state modeling and the rationale behind information sharing. We evaluate M2I2 across diverse multi-agent tasks, the results demonstrate its superior performance, efficiency, and generalization capabilities, over existing state-of-the-art methods in various complex scenarios.

This article proposes a new cache-assisted integrated sensing-communication-computation (ISCC) framework for efficient online federated edge learning (FEEL) in orthogonal frequency division multiple (OFDM) systems. The framework innovatively leverages both real-time sensory data and selectively cached historical data to enhance learning efficiency, particularly in scenarios with constrained availability of new data and limited capacity of device cache in broadband channels. However, this approach introduces critical challenges that the concept drift in historical data degrades model relevance, and the joint optimization of sensing, communication, and computation resources becomes more complex due to the interplay between new and old data utilization. To address these challenges, we first establish a comprehensive system architecture detailing online sensing, cached data utilization, local computation, and global model aggregation. Then, a rigorous theoretical convergence analysis using the first-order Taylor approximation is conducted. Via alternating optimization, a cache-assisted ISCC scheme is proposed to maximize per-round loss reduction under stringent latency, energy, and cache constraints. Extensive simulations on a human motion recognition task validate the theoretical convergence analysis and the superiority of the proposed loss-aware caching design over baselines.

Deterministic demand of mission-critical applications is essential in edge computing systems for realizing Industry 4.0. However, the conventional average-based network slicing schemes incur unexpected long-tail delay, resulting in the failure to meet strict deterministic delay guarantee. To resolve this issue, in this paper, we construct a two-scale network slicing framework for multi-access edge computing (MEC), aiming to meet deterministic delay requirements of multiple users with minimal resource usage. To derive the closed-form delay distribution, we decompose the offloading process and employ the weighted approximation to address the multi-UE coupling. Then we propose an iterative two-scale network slicing (TNS) algorithm to guarantee deterministic delay, and construct a TNS-based deep-unfolding neural network, called TNS-Net, to improve the solution in presence of inaccurate channel statistics. Moreover, for the training of TNS-Net with deterministic delay as the network input, we apply extreme value theory (EVT) to analyze the distribution characteristic of delay bound violation. Finally, simulation results demonstrate that our theoretical analysis provides a relatively accurate estimate and the proposed TNS-Net ensures better delay guarantee with lower resource consumption.

In personalized edge learning (PEL), data is essentially a form of labor. Nevertheless, existing resource-scarce edge devices tend to cache the data they sense in isolation. This creates data silos that significantly degrade the quality of edge learning models. To address this impasse, we propose a cache-aware collaborative federated system, where data sensing allocation is optimized based on both local learning requirements and scenario-specific sharing feedback. Specifically, we establish a unified system reward model that explicitly decomposes the overall utility into (i) accuracy gains from local PEL and (ii) data-sharing rewards under two coexisting data sharing paradigms. For the alliance paradigm, we derive an optimal allocation policy by analyzing the properties of sensing allocation rewards. For the reward-feedback paradigm, we further design an asymptotically optimal allocation strategy using the exact penalty method. Extensive simulations demonstrate that the proposed algorithms outperform state-of-the-art approaches.