This paper presents DINO-SLAM, a DINO-informed design strategy to enhance neural implicit (Neural Radiance Field – NeRF) and explicit representations (3D Gaussian Splatting – 3DGS) in SLAM systems through more comprehensive scene representations. Purposely, we rely on a Scene Structure Encoder (SSE) that enriches DINO features into Enhanced DINO ones (EDINO) to capture hierarchical scene elements and their structural relationships. Building upon it, we propose two foundational paradigms for NeRF and 3DGS SLAM systems integrating EDINO features. Our DINO-informed pipelines achieve superior performance on the Replica, ScanNet, and TUM compared to state-of-the-art methods.

Technology companies have built large-scale heterogeneous GPU clusters to support various workloads. However, their cluster machines are found underutilized with severe resource fragmentation. The main causes are myopic online scheduling of incoming tasks and complex placement constraints specified by users or systems. In this paper, we propose to use task migration as a measure to alleviate resource fragmentation. Our trace-driven analysis on a production cluster with 12k machines reveals that almost at any random snapshot, most of the concurrent tasks have long run-times and small migration times, thus justifying the feasibility of task migration. By making the complex constraints mathematically tractable, we formulate an integer linear programming problem. An efficient heuristic algorithm called Iterative Partitioned Defragmentation (IPD) is presented to perform task migration in multiple iterations of computation. We design and implement Defrag, an operational resource defragmentation system on Kubernetes, and deploy it on the production clusters. Trace-driven experiments show that Defrag can reduce up to 80% idle CPUs and 29% idle GPUs on average. Furthermore, Defrag can refine the performance of online scheduling strategies by reducing up to 57% idle CPUs and 35% idle GPUs at the time of execution. Our real-world experiment also demonstrates the effectiveness of Defrag.

Current approaches to few-shot 3D point cloud segmentation typically depend on heavily pre-trained and parameterized networks. These methods tend to introduce substantial computational costs and exhibit limited generalization ability when applied to new categories across different domains. To address these challenges, we propose a novel Non-Parametric network for few-shot 3D scene Segmentation, NP-3DS. Unlike existing methods that use fixed fusion schemes, our approach proposes a non-parametric intrinsic attention aggregation mechanism that adaptively integrates spatial and chromatic features, effectively preserving fine-grained structures in complex scenes. To mitigate prototype bias and improve cross-domain transferability, we propose an optimal transport prototype calibration strategy that dynamically aligns category representations between support and query sets. Furthermore, to restore structural fidelity during feature upsampling, we design a volume-inverse weighted trilinear interpolation method that captures local geometric correlations, enabling accurate recovery of sharp edges and fine details. Experiments on two benchmarks (S3DIS and ScanNet) demonstrate that NP-3DS significantly outperforms the state-of-the-art non-parametric method (gains +7.37% and +4.48% mIoU, respectively) and matches or surpasses parametric baselines. These results highlight the efficiency and generalizability of our method in a few-shot 3D scene segmentation.

Extended reality (XR) is becoming a crucial application in the shift from 5G to 6G networks, providing immersive experiences that combine real and virtual environments. However, the real-time transmission of XR video encounters challenges with high data rates and low latency, particularly due to its frame-by-frame delivery, which is highly sensitive to network fluctuations. To address this challenge, we propose a comprehensive framework for the collaborative transmission of real-time XR video and haptic feedback that emphasizes joint source-channel optimization (JSCO). Specifically, this framework involves two critical components: adaptive video quality selection at the source and efficient wireless resource allocation over the channel. By optimizing both the video content and the underlying communication resources, we aim to enhance the overall transmission efficiency and user satisfaction. We also propose a dual-layer deep reinforcement learning (DRL) algorithm for solving JSCO problem, termed as proximal policy optimization combined with multi-agent Transformer (PPO-MAT), which operates within a hierarchical framework to effectively tackle the issue of inconsistent time scales. Through extensive experiments, we demonstrate that our proposed method significantly outperforms existing baseline solutions, achieving at least a 14.39% enhancement in quality of experience (QoE) and a 67.71% increase in the number of satisfied users. These results underscore the effectiveness of our joint optimization approach, ultimately paving the way for more robust and responsive XR applications in dynamic network environments.

Beamforming design for extremely large-scale multiple-input multiple-output (XL-MIMO) systems is challenging due to prohibitive computational complexity and complex near-field propagation effects. To address this, this paper introduces a holographic beamforming paradigm that reformulates the design from optimizing variables at spatially discrete antenna locations to shaping a continuous electromagnetic wave function over the array aperture, effectively mitigating the growth of algorithmic complexity as the array scale increases. We apply this paradigm to the challenging dual near-field (DNF) scenario, where strong transceiver coupling severely degrades conventional iterative algorithms. In this case, we propose a novel Virtual Point Source (VPS) method, which approximates the ideal wave function with a single and analytically tractable spherical-wave. A rigorous geometric-optical analysis is provided to show that the optimal VPS location can be determined in a fully non-iterative manner, thus decoupling the coupled DNF problem. The proposed method is demonstrated in an intelligent reflecting surfaces (IRS)-assisted system, where simulation results show that our non-iterative approach achieves performance comparable to converged alternating-optimization (AO) algorithms, while incurring significantly lower complexity and avoiding convergence uncertainty. This work offers a new theoretical framework for holographic beamforming design in XL-MIMO systems.

Intelligent reflecting surface (IRS) is a promising technology, yet its physical design, particularly the element spacing, generally relies on the conventional half-wavelength ( lambda/2 ) rule from active arrays. This letter investigates the validity of this rule for IRS-aided systems with two-hop cascaded channels. In this letter, we first derive the spacing requirements from first principles by applying the Nyquist sampling theorem to both the reflection-angle and the cascaded-channel-angle domains. We then distinguish between Visible Grating Lobes (VGLs) and a more subtle phenomenon, termed Ghost Grating Lobes (GGLs). Our analysis confirms that while lambda/2 spacing is indeed the threshold to prevent VGLs, whereas a stricter quarter-wavelength ( lambda/4 ) spacing is the actual threshold required to avoid GGLs due to the doubled angular bandwidth of the cascaded channel. Simulation results are presented to validate our theoretical findings. It is shown that element spacing between lambda/4 and lambda/2 leads to significant channel correlation and system-level interference, even in the absence of VGLs. These results establish a refined two-threshold framework that provides a more complete guideline for IRS design, especially in interference-limited systems such as cell-free Multiple-Input Multiple-Output (MIMO) systems.

Credit card fraud brings serious losses to both cardholders and card issuers. To reduce losses caused by fraudulent behaviors, banking institutions establish credit card fraud detection (CCFD) models to identify potential fraudulent behaviors. To develop more effective fraud detection models, banking institutions need to collaborate on model training. Federated learning (FL) enables collaboration on fraud detection model training without exchanging data between banking institutions. Nevertheless, the distribution of transaction data in the real-world is heterogeneous among banking institutions, which may lead to convergence issues in global fraud detection models. Furthermore, the behavior and weights of the model may implicitly contain the cardholders' personal information, which makes existing federated models prone to transaction data leakage. In this article, we propose a heterogeneous credit card fraud detection model based on federated learning with membership privacy, called HFDFM. To protect the sensitive information of cardholders, we design a novel mechanism that ensures training-data confidentiality by minimizing the accuracy of the best black-box membership inference attack (MIA) against the model. Unlike previous FL frameworks that either optimize for client drift or for membership privacy, HFDFM simultaneously mitigates drift and provides certified membership privacy through a single min-max game. Additionally, we use the control variable to rectify the client drift in its local update in the scenario of heterogeneous data. Extensive experimental results on three mainstream real-world transaction datasets demonstrate that the proposed HFDFM has advantages in utility compared to 11 SOTA baselines, and the proposed HFDFM can mitigate the risks of MIAs (near random guess).

3D Gaussian Splatting (3DGS) has recently gained popularity in SLAM applications due to its fast rendering and high-fidelity representation. However, existing 3DGS-SLAM systems have predominantly focused on indoor environments and relied on active depth sensors, leaving a gap for large-scale outdoor applications. We present BGS-SLAM, the first binocular 3D Gaussian Splatting SLAM system designed for outdoor scenarios. Our approach uses only RGB stereo pairs without requiring LiDAR or active sensors. BGS-SLAM leverages depth estimates from pre-trained deep stereo networks to guide 3D Gaussian optimization with a multi-loss strategy enhancing both geometric consistency and visual quality. Experiments on multiple datasets demonstrate that BGS-SLAM achieves superior tracking accuracy and mapping performance compared to other 3DGS-based solutions in complex outdoor environments.

We consider structural equation models (SEMs), in which every variable is a function of a subset of the other variables and a stochastic error. Each such SEM is naturally associated with a directed graph describing the relationships between variables. When the errors are homoscedastic, recent work has proposed methods for inferring the graph from observational data under the assumption that the graph is acyclic (i.e., the SEM is recursive). In this work, we study the setting of homoscedastic errors but allow the graph to be cyclic (i.e., the SEM to be non-recursive). Using an algebraic approach that compares matroids derived from the parameterizations of the models, we derive sufficient conditions for when two simple directed graphs generate different distributions generically. Based on these conditions, we exhibit subclasses of graphs that allow for directed cycles, yet are generically identifiable. We also conjecture a strengthening of our graphical criterion which can be used to distinguish many more non-complete graphs.

The rapid growth of large language models (LLMs) and the continuous release of new GPU products have significantly increased the demand for distributed training across heterogeneous GPU environments. In this paper, we present a comprehensive analysis of the challenges involved in implementing 3D parallelism in such environments, addressing critical issues such as the need for symmetric tensor parallelism, efficient gradient synchronization in asymmetric pipeline parallelism, and the trade-offs between memory utilization and computational efficiency. Building upon these insights, we introduce AutoHet, a novel system that automatically identifies the optimal parallelism plan for distributed training on heterogeneous GPUs. AutoHet supports asymmetric 3D parallelism structures and facilitates fine-grained workload distribution. We propose a theoretical model that frames the device grouping and load balancing as an optimization problem to minimize per-iteration training time, thus effectively balancing computing power and memory usage across GPUs with diverse capabilities. To enable elastic training upon spot instance preemption, AutoHet presents an efficient recovery strategy that prioritizes to retrieve training states from local nodes, and only downloads the missing checkpoints from the cloud storage. Our extensive evaluation, conducted on three large-scale models and utilizing combinations of three different GPU types, demonstrates that AutoHet outperforms existing DNN training systems, achieving up to a 1.79× speedup in training throughput compared with Megatron-LM and Whale, and a 4.38× speedup of recovery speed compared to a spot instance baseline.

Extended reality (XR) technologies and applications have grown rapidly in recent years. In addition to providing immersive ultrahigh-definition (UHD) video, XR also allows for a haptic experience where devices can be remotely manipulated to accomplish tasks. However, varying numbers of XR users accessing the communication system can strain limited spectrum resources, posing challenges in resource allocation. Therefore, this article studies resource blocks (RBs) allocation problem in a downlink transmission scenario where real-time cloud XR video and haptic contents need to be transmitted simultaneously. We also consider the random variation in the number of XR users and propose an adaptive distributed multiagent deep reinforcement learning (DRL) combined with Transformer (ADMA-DcT) for dynamic RB allocation method. This method addresses the dynamic change in network input dimensions due to user number variability using a state division module and a self-attention mechanism in the encoder module. To our knowledge, this is the first work to study the RB allocation problem in Cloud XR transmission considering simultaneous transmitting of video and haptic services with a dynamically changing user base. Our extensive simulations show that the ADMA-DcT model, end-to-end trained, outperforms other benchmarks in successfully serving a larger number of XR users under varying user number conditions, demonstrating excellent adaptivity and robustness.

Traditional visual SLAM systems (dense and sparse) focus on building metric maps, but the internal representations are misaligned with human vision, making it insufficient for assisting robots in scene perception and interpretation. Conversely, aligning robot scene representation with human vision enables more intuitive human-to-robot commands and improves the generalization capability of deployed neural networks trained on natural images. Neural scene representation-based visual SLAM system, with its consistent and high-fidelity mapping, provides a novel way to assist robots in detailed scene depiction and comprehensive perception. However, end-to-end methods suffer from low accuracy in robot localization, which inevitably degrades mapping quality and limits their practical applications. In this paper, we propose a robust hybrid SLAM system, named OGS-SLAM, which integrates traditional visual SLAM with 3D Gaussian Splatting (3D GS) mapping. This system inherits the high localization accuracy of traditional SLAM while providing a scene model that aligns with human cognition, thereby offering a reliable foundation for downstream human-robot interaction tasks. Experiments demonstrate that our method outperforms state-of-the-art (SOTA) end-to-end SLAM systems in localization, mapping, and map semantic segmentation. Code will be available at: https://github.com/realXiaohan/OGS-SLAM.

As a key technology in intelligent driving, Cooperative Vehicle Infrastructure System is an advanced communication framework that enables cooperative and information exchange between vehicles and infrastructure. However, this system encounters challenges in meeting the low latency, ultra reliability, and high efficiency requirements of task execution of intelligent driving applications. Meanwhile, the deployment and combination of some key technologies supporting this system are not well-established. To address these issues, this article proposes a Digital Twin-based Cloud-native Vehicular Networks (DT-CVN) architecture to enhance the efficiency of virtual-reality integration in real-world vehicle traffic scenarios. In DT-CVN, the digital twins, which can bridge the physical space and cyberspace gaps in real-time, are implemented and deployed in a distributed manner by leveraging the distributed features of microservices based on cloud-native technology. DT-CVN employs the cybertwin as a smart communication agent in cloud-native vehicular networks, enabling efficient communication between cyberspace and physical spaces. Moreover, a case study is presented to demonstrate the effectiveness of DT-CVN. Simulation result shows the potential to address the challenges of integrating resources in vehicular networks with our proposed DT-CVN.

This paper explores the problem of instruction selection optimization in LLVM backend code for transforming platform-independent intermediate code into high-quality target platform instructions. The instruction selection process is divided into two problems: pattern matching and pattern selection. We propose a novel architecture-aware optimization strategy that leverages the features of digital signal processors (DSPs) to improve the efficiency and performance of instruction selection. Our approach involves analyzing the characteristics ofDSP architectures to guide the selection of optimal instructions. We evaluate our approach on a set of benchmarks and demonstrate significant improvements in both execution time and code size compared to existing LLVM optimization techniques. Our results show that architecture-aware optimization strategies can effectively enhance instruction selection in DSP compilers, leading to better performance and reduced energy consumption in real-world applications.

Intelligent driving has advanced significantly in recent decades, paving the path for the transportation of the future. Digital twin (DT) technology, which can bridge the physical and virtual space gaps in real time, plays an important role in the collaboration of vehicles and roads for intelligent driving. In this paper, we design the Digital Twin empowered Cloud-native Vehicular Networks (DT-CVN) architecture as well as the workflow for task execution aiming at executing intelligent driving tasks efficiently and reliably. In DT-CVN, we propose a design and implementation scheme for digital twins, in which different modules of the same digital twin entity can be deployed in different network locations in a distributed manner by taking advantage of the distributed features of microservices based on cloud-native technology. Furthermore, we design the modules reuse and requests aggregation mechanisms of digital twins invocation for task scheduling in DT-CVN, which can improve its efficiency even further. Then we model the task scheduling in DT-CVN into a combinational optimization problem and propose a deep reinforcement learning (DRL) based adaptive task scheduling algorithm. Simulation results show that the proposed scheme can improve the efficiency of task scheduling while reducing energy consumption.

With the introduction of sixth-generation (6G) communication, the attention to integrated sensing and communication (ISAC) is growing rapidly due to the diverse requirements for 6G communication services. However, the sensing performance in wireless scenarios is constrained by limited resources, such as the number of array antennas, spectrum, and power. Nevertheless, reconfigurable intelligent surface (RIS) is suitable for assisting ISAC because it can manipulate the incident signals with relatively lower power consumption. Therefore, this paper proposes a user localization and environment estimation problem with the assistance of orthogonal frequency-division multiplexing (OFDM) and RIS. Considering that carrier frequency offsets (CFO) will impact the location estimation in OFDM, we first investigate the joint estimation of time of arrivals (TOA) and 2-D direction of arrivals (DOA) without CFO estimation by taking advantage of matrix transformation. In order to increase the ability of environment estimation of wireless signals, we develop an efficient environment estimation method with the help of RIS to adjust the directions of the incident signals in different frequencies. Finally, by taking the space-time continuity of user movement and the geometry features of environment into account, we establish the probability transition model of the estimation process for user's location and environment by factor graph. The performance of our factor graph-based estimation algorithm is demonstrated by simulation results.

The rapid development of new communication applications such as virtual reality and video conferencing has brought new challenges to congestion control algorithms. In particular, these applications have specific requirements in terms of delay. Meeting specific delay requirements without high throughput loss is difficult, especially in dynamic networks. In addition, it is important that the proposed congestion control algorithms can be easily deployed. In this paper, we propose a congestion control algorithm, namely TD-BBR, to meet the delay requirements of different applications. TD-BBR is built on BBR and can adapt to various network environments without high throughput loss. we employ an online algorithm based on the recursive least squares method to predict future bandwidth. We design a simple and effective algorithm to adjust the congestion window (CWND) to meet the specific delay requirements according to the value of bandwidth prediction and the distance between the current delay and the target delay. We implement a real congestion control system through extended Berkeley Packet Filter (eBPF) technology and have deployed it in the Linux kernel without recompiling the kernel. Extensive experiments show that TD-BBR can effectively meet different delay requirements in most cases, decrease the 95th percentile delay, and avoid high throughput loss compared to other congestion control algorithms,

The future of networking must address the connectivity demands of billions of people and trillions of networked devices. The Intelligent Internet of Everything (IoE) is widely considered the next evolution of the Internet, yet it poses significant challenges to the current TCP/IP architecture. To overcome these challenges, the concept of the Cybertwin network has been proposed as a promising future Internet architecture. To facilitate the advancement of Cybertwin-related research, we have developed CTSim, a Cybertwin network simulator. Our demo demonstrates that the Cybertwin network could effectively enhances mobility, availability, and security in Internet of Things scenarios, making it a powerful tool for researchers in these fields.