ArchiveSince 2012, the growth rate of intelligent computing has outpaced Moore’s Law, achieving an order-of-magnitude increase per year, which poses unprecedented challenges to the architecture of computer. Ultra-large-scale intelligent computing clusters have been proposed as national new infrastructure in “Report on the Work of the State Council” in 2026. Considering the characteristics of human intelligence, i.e “small computing power, large network”, the enabling technologies of optical transmission and switching with lower power consumption, and new architectures and mechanisms driven by technological innovation, we present ten preliminary judgments on overseas intelligent computing clusters development trends from four aspects: cluster architecture, computing nodes, network technology, and technical ecosystem. This paper aims to offer insights and contribute to the national implementation of ultra-large-scale intelligent computing clusters as new infrastructure.
Computility federations coordinate heterogeneous resources across independent providers while preserving administrative autonomy. This autonomy limits federation-level visibility, forcing gateways to route from coarse-grained and delayed performance summaries rather than real-time internal states. Under such weak and lagged observability, deadline-miss risk for latency-critical tasks can become structurally concentrated in a small number of domains and time windows, even when federation-wide averages remain moderate. To address this problem, we propose per-domain SLO feedback (PDSF), a low-dimensional closed-loop mechanism that updates bounded routing tolls from window-level summaries and biases snapshot-driven self-routing without modifying domain-local schedulers. Theoretical analysis shows that, under a weak monotonicity assumption on routing response, PDSF exhibits boundedness, forgetting, and local stability under delayed feedback. Trace-driven experiments on a multi-domain emulator using Google cluster workloads show that, compared with deployable baselines, PDSF significantly alleviates upper-tail overload severity, reduces worst-domain risk exposure, and protects latency-critical traffic without noticeable degradation to best-effort traffic. Sensitivity analysis further shows that snapshot staleness and routing concentration jointly determine the mechanism’s operating boundary and clarifies how stale measurements and aggressive routing response amplify risk while reducing the effectiveness of low-dimensional feedback.
AI computility integrates geographically distributed intelligent computing resources via high-speed networks, providing efficient and elastic infrastructure for large-scale intelligent applications. Multi-agent debate (MAD), as a distributed multi-expert collaborative paradigm, connects multiple large language model (LLM) agents through computility networks to engage in iterative discussions from distinct roles, effectively enhancing collaborative reasoning quality. However, when facing jailbreak attacks—a typical class of attacks targeting LLMs—the security of the collaborative reasoning paradigm in MAD remains unexplored. To address this gap, a jailbreak attack method targeting MAD is investigated under a semi-black-box threat model. A structured prompt-rewriting template is designed that systematically integrates four strategies—narrative encapsulation, role-driven escalation, iterative refinement, and rhetorical obfuscation—to exploit the interactive mechanisms of multi-agent debate for inducing harmful outputs. Experiments on four mainstream MAD frameworks with GPT-4o, GPT-4, GPT-3.5-turbo, and DeepSeek demonstrate that the proposed method increases the average harmfulness of generated content by 28.14% to 80.34%, with attack success rates reaching 80% in certain scenarios. Building on these findings, preliminary defense experiments are further conducted across the input, generation, and propagation stages, providing initial evidence for the feasibility of lightweight multi-stage protection in MAD systems. These findings expose potential security vulnerabilities within multi-agent debate, validating the necessity of developing robust and targeted defense strategies before the widespread deployment of multi-agent collaborative systems under AI computility.
The increasing parameter scale and structural complexity of deep neural networks (DNNs) pose significant challenges for their efficient deployment in cloud-edge-end collaborative computing architectures, where edge and terminal nodes must support real-time inference under stringent constraints on computation, storage, inference latency, and communication overhead. Although model quantization effectively reduces resource consumption, traditional approaches based on fixed rules or empirical heuristics exhibit notable limitations in accuracy preservation, training stability, and adaptability across diverse architectures and application scenarios, particularly under very low bit-width conditions (≤4 bit). To mitigate these issues, learning-based quantization incorporates learnable components and diverse supervisory signals, substantially improving accuracy preservation, optimization robustness, and hardware adaptability under low-bit conditions. In this paper, we systematically review representative studies on learning-based quantization, centered on two primary paradigms: post-training quantization (PTQ) and quantization-aware training (QAT). We summarize performance enhancement strategies including learnable parameter modeling, reconstruction and approximation optimization, auxiliary supervision, and learning-driven methods under data and hardware constraints. We further organize existing technical approaches and their interrelationships from the perspectives of quantization process stages and learning signal mechanisms, and analyze the applicability of different strategies under varying data conditions, model scales, and application constraints. Finally, we discuss future research directions in learning-based quantization from three aspects: error-analysis-based optimization of quantization parameters and learning signals, the trade-off between accuracy and resources in learning-based quantization, and a unified quantization framework across model architectures.
MoE (mixture of expert) architectures have become a widely recognized paradigm for scaling large language models (LLMs), and have been adopted by many frontier models deployed at the network edge. However, deploying MoE-based LLMs on resource-constrained edge devices remains challenging due to their massive parameter sizes, which often exceed the memory capacity of individual platforms. Existing in-situ MoE serving systems adopt expert swapping strategies that offload inactive experts to secondary storage and load them on demand. As LLMs support increasingly longer context windows with correspondingly longer user inputs, the prefilling phase faces a critical bottleneck: long input sequences activate nearly all experts, and the resulting full-scale expert swapping generates substantial I/O traffic, leading to prolonged TTFT (time to first token) latency. Typical edge environments, such as smart homes and smart factories, comprise multiple trusted interconnected devices, including tablets, laptops, and smart home appliances connected via local area networks. These devices collectively form a pervasive edge computility network (ECN), where each node carries underutilize computing power, memory capacity, and I/O bandwidth. This motivates us to harness the aggregated resources of the ECN through coordinated cross-node scheduling. This paper presents Sirius, a collaborative MoE inference system for edge computility networks that pools the collective resources across ECN nodes to accelerate the prefilling phase and reduce TTFT. Specifically, Sirius proposes three key techniques: 1) an interleaved sequence partitioning strategy that resolves workload imbalance in causal attention computation and maximizes communication-computation overlap; 2) an adaptive token scheduling mechanism that jointly considers computation overhead, swap latency, and expert buffer cache states to fully utilize collective I/O bandwidth while achieving load balance; 3) a speculative expert prefetching scheme that exploits the high similarity between adjacent-layer hidden states to overlap expert loading with computation, effectively hiding I/O latency. Experiments on real-world ECN platforming with Mixtral-8x7B and Qwen3-30B-A3B demonstrate that Sirius achieves 1.2–4.0 times inference throughput improvement over state-of-the-art baselines.
With the increasing popularity of deep neural network (DNN) applications, Chiplet-based accelerators have become critical infrastructure for supporting large-scale DNN inference services on server platforms. However, the real-time and dynamic characteristics of parallel multi-DNN workloads pose severe challenges to scheduling algorithms, while existing algorithms suffer from limitations in decision overhead, service level agreement (SLA) assurance, and dynamic adaptability. To address these issues, we propose MAOSA, a multi-agent reinforcement learning based online scheduling algorithm for multi-DNN workloads on multi-Chiplet accelerators, aiming to maximize the overall SLA satisfaction rate. The proposed algorithm partitions DNN tasks into DNN blocks and models them as agents. It utilizes the attention mechanism to generate scheduling strategies in parallel, which include execution order and PE-level resource allocation, thereby significantly reducing decision overhead. Furthermore, a cooperative reward function considering time slack is designed to optimize SLA, and a periodic unified preemption mechanism is introduced to adapt to dynamic load changes. Experimental results show that, compared with state-of-the-art algorithms, the proposed algorithm improves the SLA satisfaction rate by up to 18% and system throughput (STP) by 31.9%, reduces tail latency by 28.7%, and decreases scheduling execution time by a factor of approximately 11, effectively enhancing system performance and service quality.
Deploying large language models (LLMs) in mobile and edge computing environments faces multiple challenges, including limited on-device resources, scarce wireless bandwidth, and the continual evolution of cloud-side models. Although speculative decoding based edge-cloud collaborative inference can reduce end-to-end generation latency by using a lightweight draft model at the edge and a target model in the cloud for parallel verification, existing methods usually rely on a tightly coupled relationship between the two models. In practical systems, frequent updates of the cloud model often require repeated synchronization of the edge-side draft model, leading to substantial communication overhead, higher latency, and limited scalability. To address this issue, ECSpec is proposed as a communication-efficient edge-cloud collaborative inference framework for model-evolving scenarios. Its key idea is a shared backbone architecture that enables a single static edge-side draft model to remain compatible with a family of continuously evolving cloud-side target models, thereby avoiding repeated retraining or weight downloading at the edge and significantly reducing communication and maintenance costs. In addition, a channel-aware adaptive speculation mechanism is introduced to dynamically adjust the draft length according to real-time channel conditions and energy budgets, achieving a better tradeoff between inference efficiency and energy consumption. Experimental results demonstrate that ECSpec delivers more stable and efficient collaborative LLM inference under communication-constrained edge environments.
Video anomaly detection is of great practical value in the fields of intelligent video surveillance, public security, etc. With the advancement of deep learning, conventional pixel-based video anomaly detection methods gradually expose their vulnerability to noises, high computation cost, while skeleton-based approaches have emerged as a research hotspot due to their robustness and efficiency. This study provides a detailed explanation of the fundamental concepts, learning paradigms, and methodological workflows of human skeleton-based video anomaly detection methods, and systematically reviews and classifies the representative skeleton-based studies published in the recent five years into different categories, including prediction-based method, reconstruction-based method, hybrid reconstruction and prediction-based method, hybrid reconstruction and clustering-based method, as well as unclassified method. An in depth analysis is given to explore the underlying principles and innovations of each type of method. Moreover, this study summarizes existing benchmark datasets and evaluation metrics for skeleton-based video anomaly detection, as well as gives a comparative performance analysis of the mainstream methods across the benchmarks. Currently, skeleton-based anomaly detection methods face significant challenges in skeleton feature extraction, modeling and understanding of multi-person interaction, etc. Therefore, this study further proposes scientifically grounded future directions with regard to robust skeleton feature extraction algorithm, multimodal information integration and heterogeneous data fusion, etc., aiming at exploring the positive factors for constructing robust video anomaly detection framework, as well as analyzing the existing problems and applicable research directions of these series of video anomaly detection methods in different aspects, including adaptation ability, generalization ability, running efficiency, etc.
Multi-modal 3D object detection is a pivotal task in autonomous driving and environment perception. It achieves accurate localization and recognition of objects in 3D space by integrating data from multiple sensors, such as LiDAR (light detection and ranging) point clouds and camera images. To address the common challenges of large scale variations and insufficient multi-scale feature extraction in existing methods, we propose a multi-modal 3D object detection method based on dynamic scaling convolution. For this method, we design a dual-branch multi-modal fusion architecture, which consists of a VoxelNet-based 3D backbone and a novel dynamic scaling network. The two branches collaboratively process LiDAR point clouds and image-derived virtual point clouds to achieve cross-modal feature alignment and adaptive fusion. The core of the dynamic scaling network is a dynamic scaling module, which integrates 3D and 2D dynamic scaling convolution operators specifically tailored for multi-modal feature processing. This module adaptively adjusts the sampling locations and the number of valid sampling points within the receptive fields of 3D and 2D branches, thus effectively capturing structural and semantic variations of multi-scale objects. Extensive experiments on two benchmark datasets demonstrate the proposed method outperforms more than ten prevailing methods in terms of AP (average precision). Specifically, it achieves mAP3D (3D mean average precision) scores of 86.87%, 46.68%, and 68.39% for the car, pedestrian, and cyclist categories on KITTI test set, and achieves a mAP (mean average precision) of 71.8% on nuScenes test set. These results demonstrate the superior detection performance and strong generalization ability of the proposed method for multi-modal 3D object detection tasks.
With the development of artificial intelligence technology, vision-based object recognition, as an important component of the Internet of things perception layer, can be applied in various cargo sorting scenarios when combined with sorting robots. However, in complex environments where objects are disorganized, relying solely on object recognition algorithms may not be sufficient to distinguish target objects from interfering ones, as the difference between them may be the spatial distance. This can negatively impact the robot’s normal grasping actions. To address this issue, we propose a lightweight object recognition method for complex environments—3Deus. This method combines deep learning with 3D point cloud data processing. It uses a depth camera to synchronously capture both RGB images and 3D point cloud data in the recognition scene. The method eliminates the plane where interfering objects are located using the 3D point cloud, and distinguishes the plane containing the target object from the stacked interfering objects with an improved RANSAC algorithm. It also employs a plane expansion mechanism to ensure the integrity of the plane point cloud where the target object is located. Finally, the object recognition algorithm is used to obtain the recognition bounding box of the target object and construct its mapping relationship with the specific point cloud cluster. We deploy 3Deus on edge computing devices and validate its effectiveness and robustness by constructing different experimental scenarios. Experimental results show that compared with the traditional random sampling consensus algorithm, 3Deus can effectively distinguish the target objects from interference objects, achieving an accuracy of over 95%. Moreover, by utilizing the plane expansion mechanism, the recognition accuracy of the target object is improved by 30.48%.
With the rapid development of deep learning-based voice conversion (VC) technology, attackers can compromise the security of automatic speaker verification (ASV) systems by tampering with the source speaker’s speech signal to mimic the target speaker’s acoustic characteristics. The existing anti-spoofing countermeasures primarily focus on distinguishing between genuine and spoofed speech, yet underutilize the implicit source speaker identity information concealed within the spoofed speech. This limitation hinders their ability to meet practical demands in scenarios such as forensic investigation and fraud traceability. Currently, research on source speaker identity verification and discrimination remains insufficient, and existing methods suffer from limited generalisation capabilities. To address this, we construct a new voice conversion spoofing dataset specifically designed for source speaker verification, named VCSD-SSV (voice conversion spoofing dataset for source speaker verification). This dataset encompasses six VC algorithms (i.e., Seed-VC, FACodec, DDDM-VC, Diff-HierVC, TriAAN-VC, and Free-VC) and includes 400 speakers, totaling 327 677 speech samples (46 811 genuine samples and 280 866 converted samples). Building upon this, we propose a feature generalisation enhancement-based source speaker verification model, FGE-SSV (source speaker verification based on feature generalised enhancement). This model extracts multi-level source speaker timbre features through a feature enhancement module and cascaded instance-batch normalization squeeze-and-excitation blocks. Experimental results demonstrate that FGE-SSV maintains consistently low equal error rate (EER) (0.2%~1.0%) in within-domain results across all six VC methods in VCSD-SSV dataset and also achieves relatively low EER across the six VC methods in SSTC2024 dataset. Furthermore, FGE-SSV demonstrates strong generalization performance on unseen voice conversion methods, with its EER values outperforming those of comparative methods.
A novel unsupervised multi-class anomaly detection model, termed DiffAD, is proposed to tackle the critical challenges of annotation scarcity and insufficient detection accuracy plaguing visual anomaly detection tasks in complex industrial scenarios. The model employs a progressive feature reconstruction strategy, with its core innovation centered on a dedicated feature reconstruction component named DADE. Breaking new ground, DADE decomposes the feature reconstruction process into three sequential phases, namely chaos, pre-refinement and strong refinement, which synergistically enhance the quality of reconstructed features and ensure the stability of the entire reconstruction workflow. A key standout of DADE lies in its seamless integration of difference convolution attention and a detail enhancement mechanism. It combines difference convolution with multi-head self-attention and incorporates residual dense connections, thereby enabling the component to markedly strengthen the model’s ability to capture subtle image variations and high-frequency details that are critical for anomaly identification and substantially elevate the precision of pixel-level anomaly localization. Extensive empirical evaluations are conducted on four representative benchmark datasets, including MVTec-AD, VisA, MVTec-3D and Uni-Medical. The experimental results consistently demonstrate that DiffAD achieves significant performance superiority over state-of-the-art mainstream models across both image-level and pixel-level anomaly detection metrics, fully underscoring its remarkable practical application value and great potential for widespread deployment in the domain of unsupervised visual anomaly detection.
To tackle the dual challenges of spatiotemporal feature redundancy and limited generalization in violent action recognition, we present RDT-MoE, a lightweight yet effective framework that couples redundancy-diminishing token (RDT) with a mixture-of-experts (MoE) architecture. The RDT module employs spatially and temporally decoupled attention to disentangle motion from appearance, and uses a dynamic Top-k token selection strategy to suppress nearly 50% of background redundancy, keeping only the most salient motion cues for subsequent encoding. Building on these compact tokens, we develop a LoRA-guided expert system and train it with a two-stage strategy, where we first jointly optimize the encoder and LoRA adapters to learn strong shared representations and stable adaptation signals, and then freeze the backbone to focus on refining the gating function and expert composition, which further improves robustness, convergence stability, and generalization across diverse scenes. To address limitations of existing benchmarks, we also introduce diverse violence action dataset (DVAD), a higher-resolution, multi-view violence dataset that is 50% larger than RWF-2000 and better reflects real-world variability. Extensive evaluations on RWF-2000 and DVAD, together with component-wise ablations, show that RDT-MoE outperforms VideoMAE by 2.00% and 2.34% in accuracy, respectively, while reducing parameters by about 50%, demonstrating strong efficiency, generalization, and practical applicability. This design supports deployment on resource-limited surveillance devices and maintains stable performance across diverse scenes consistently.
Brain-inspired spiking neural networks (SNNs) have shown considerable promise due to their strong spatiotemporal encoding capabilities and event-driven computation. Current mainstream SNN approaches can be divided into two categories: converting pre-trained artificial neural networks (ANNs) into SNNs (ANN conversion methods) and directly training SNNs. However, ANN conversion methods require large time steps to reduce conversion errors, while directly trained SNNs suffer from poor representational capacity. To address these challenges, we propose a high-performance spiking neural network based on hybrid model conversion. The proposed method decouples model training and inference. During the training phase, a portion of the spiking neurons in the SNNs are strategically replaced with ReLU activation functions to enhance the feature learning ability, thereby constructing a high-performance hybrid model that integrates both ANN and SNN components. Since the ANN part of the hybrid model cannot perform event-driven computation, we further introduce a reparameterization technique to convert the entire inference process of the hybrid model into event-driven computation without performance loss. Therefore, the proposed method combines the advantages of high performance and event-driven computation. Experimental results show that the proposed method achieves classification accuracies of 97.31% and 83.34% on CIFAR-10 and CIFAR-100, respectively, 70.89% on ImageNet, and 82.71% on the neuromorphic dataset CIFAR10-DVS, outperforming state-of-the-art methods.
Deep learning methods have achieved remarkable progress in point cloud completion tasks due to their fast inference and good generalization capabilities. However, existing methods pay less attention to the local geometric information of point clouds, such as normal vectors, vector angles, and higher-order relationships. To address this issue, we propose a geometric point cloud completion network based on hypergraph convolution to generate detailed and complete point clouds. The framework mainly consists of two core components: a geometric feature extractor and a coarse generation module based on hypergraph convolution. Specifically, the former fully extracts the normal vectors, the difference vectors between the center points and the sampled points, and their angles, and integrates them as position encodings into the Transformer structure, enabling the encoder to more fully learn the geometric characteristics and rotation-invariant features of the point cloud. Additionally, in the coarse generation module based on hypergraph convolution, by introducing hypergraph modeling to capture the higher-order correlations between points and using the cross-attention mechanism to explore the interaction between per-point features and higher-order neighborhood features, it guides the generation of point clouds with more geometric structural rationality. Extensive experimental results on multiple datasets verify that the proposed method has superior performance.
Multimodal recommendation methods have introduced features such as text and images to enrich item representations, significantly enhancing recommender system performance. In recent years, large language models (LLMs), with their extensive world knowledge and powerful reasoning capabilities, have shown great potential in the field of recommendation. However, effectively integrating LLMs with multimodal information to construct LLM-based multimodal recommender systems still faces several challenges. Firstly, there is a significant semantic gap between the textual input format of LLMs and the multimodal information, which limits the model’s ability to fully understand multimodal features. Secondly, there exists an information imbalance between different modalities, where naive feature fusion can lead to performance degradation. To address these challenges, we propose an LLM-based multimodal recommendation algorithm with adaptive modality fusion (AMFRec). The framework introduces multimodal hybrid prompt templates and projection modules, aligning non-textual features with the LLM token space and recommendation objectives through instruction tuning. To mitigate modality imbalance, a multimodal weight dynamic allocation strategy based on gating mechanism is designed, enabling adaptive adjustment of modality weights by combining mixture-of-experts projection and gating network. In addition, the algorithm requires training only the projection module and gating network, without updating the LLM itself, thereby significantly reducing training costs. Extensive experiments on multiple datasets demonstrate the effectiveness of the proposed algorithm.