Intent-based networking (IBN) automates network configuration by translating high-level user intents into executable policies. In this work, we focus on intent-driven network topology deployment, a critical scenario in IBN. While large language models (LLMs) show significant promise in enhancing intent interpretation and policy translation, current approaches grapple with abstract intent understanding and the prohibitive costs associated with large-scale proprietary models. To address this limitation, we propose a lightweight IBN agent powered by a novel fine-tuning strategy for open-source medium-scale LLMs, integrating curriculum learning (CL) with low-rank adaptation (LoRA). To address the scarcity of training data for topology deployment, we construct an expert-verified dataset tailored to intent-driven scenarios for robust model fine-tuning. Our proposed fine-tuning algorithm dynamically schedules training samples based on intent ambiguity, thereby enhancing model stability and generalization, particularly under limited data conditions. Experimental results demonstrate that the proposed method achieves a superior task success rate, outperforming GPT-4 (with prompts) by 2.4 times, while maintaining superior explainability and function-call similarity. Crucially, it accomplishes this with significantly lower economic inference costs compared to GPT-4’s API usage. This research provides a practical and efficient solution for intent-driven topology deployment in IBN systems, effectively balancing performance, cost, and scalability.
With the rapid advancement of artificial intelligence (AI) technologies, the rendering paradigm has shifted from traditional mesh-based local illumination techniques to neural representation methods that support global illumination. However, conventional mesh-based rendering pipelines are ill-suited to the requirements of this emerging paradigm and often fail to provide the adaptability and expressiveness demanded by modern rendering approaches. Consequently, a coordinated optimization of software and hardware components is essential to accommodate these evolving needs. To address these challenges, we focus on a fusion architecture that integrates 3D Gaussian with mesh representations and propose the F2IRE (fusion flexible rasterization engine), a generalized rasterization engine developed through a software-hardware codesign approach to considerably improve global rendering performance. F2IRE serves as the core of a novel global illumination neural rendering pipeline that effectively fuses traditional mesh representations with 3D Gaussian splatting (3D-GS). Experimental results demonstrate that, for mesh-based rendering scenarios, the F2IRE effectively balances the computational workload between the rasterization engine and neural network components, enabling high pixel interpolation rates. In 3D-GS tasks, the F2IRE-based pipeline achieves a 21
Software technology is undergoing a paradigm shift driven by two converging trends. First, the scope of software responsibility has expanded significantly: as “software-defined everything” becomes a reality, software has evolved into the integration core of sociocyber-physical systems (SCPSs). Second, the capabilities and development methods of software are being greatly enhanced by recent breakthroughs in artificial intelligence (AI). This article presents perspectives and observations on software engineering in this era of rapid progress. We aim to outline a set of foundational challenges in engineering SCPSs and highlight the need for innovative software solutions that extend beyond AI technologies alone. Specifically, we examine the need for a new paradigm that can address the complexities introduced by SCPSs, which challenge conventional paradigms through the blurring of system boundaries, continuous lifecycle evolution, and the embracing of inherent uncertainty. We highlight new engineering principles of socio-technical co-design, cyber-physical integration, and development-operation convergence, and a knowledge- and data-driven approach to taming uncertainty. Emerging proposals, including digital humanism, agentic SCPS, ubiquitous operating system, and continuous quality assurance, are discussed alongside possible extensions to existing technologies. We then outline key research directions for both runtime support and quality assurance. On the runtime side, we argue for a new generation of software infrastructure for SCPSs, including unified hardware abstractions, scalable and resilient runtime systems, AI-enabled system management, and human-centric operating system primitives. On the assurance side, we highlight the need for new approaches combining unified socio-cyber-physical modeling, specification of both technical and non-technical properties, data-driven simulation and testing, continuous verification, and runtime monitoring under uncertainty. Furthermore, we present specific challenges within key application domains, including intelligent vehicles, smart manufacturing, and smart cities. In doing so, we aim to stimulate discussion within the software engineering community and encourage support from industry and government to address the critical engineering and governance issues inherent to this new generation of systems.
Low Earth orbit (LEO) satellite networks have gained significant attention due to their flexibility in deployment, global coverage, and high data rates, making them a critical component of future space-air-ground integrated networks. However, their broad coverage introduces wireless security vulnerabilities, while stringent onboard power budgets and the energy supply demands of satellite terminals challenge their energy efficiency and sustainability. To address these cross-domain issues, this paper first derives a secure energy efficiency entropy formula from an information-theoretic perspective, which focuses on security, power utilization, and energy harvesting. Guided by this formula, we integrate movable antennas (MAs), hybrid precoding, and simultaneous wireless information and power transfer (SWIPT) to dynamically shape beams, control transmit power and wireless power transfer to terminals. To mitigate empirical model mismatches to practical scenarios (channel characteristics, user attributes, and hardware constraints), we propose a generative AI agent framework that automatically constructs system configurations based on natural language descriptions. The agent employs a semantic router and retrieval-augmented generation (RAG) to recommend modeling parameters for optimization objectives, channel characteristics, user attributes, and radio-frequency chain topologies. Based on the customized model, we formulate a joint optimization problem for hybrid beamforming and MA positioning to maximize energy efficiency under secrecy rate and energy harvesting constraints. As the problem is NP-hard, a multi-agent deep reinforcement learning (DRL) approach is developed to optimize the transmit power, analog/digital precoders and MA positions. Simulations show that the proposed multi-agent AI framework can translate requirements into system configurations, effectively avoiding configuration pitfalls, and the DRL-based solution significantly outperforms conventional fixed-antenna and random-positioning baselines in energy efficiency. This work paves a new path from entropy-based theoretical modeling to AI-based optimization in building efficient, secure, and energy-self-sufficient LEO satellite IoT infrastructures.
Information diffusion prediction comprises macroscopic and microscopic prediction tasks. The former aims to estimate the overall impact of the information propagation process, while the latter focuses on predicting the next participating user, both of which are crucial to understanding how information spreads in a network. Hence, as a recent trend, it is important to investigate multi-scale prediction models. Nevertheless, conventional methods primarily focus on extracting shared features but fail to consider the complex interactions between users and cascades, as well as the feature similarity of the same user across tasks. In this connection, we propose a multi-dimensional feature interaction (MDFI) prediction model with two novel designs. Specifically, to model the latent hierarchical interactions between users as well as users and cascades, we construct a heterogeneous weighted user-cascade graph at the macro-level. At the micro-level, we account for the social homophily of users in the social network and develop the cascade neighbor contrastive learning as a constraint under the multi-task learning framework. Experimental results on four publicly available datasets demonstrate that MDFI exhibits superior performance to all competing models.
Spiking neural networks (SNNs) are gaining attention for energy-efficient, event-driven computing due to their ability to encode information through discrete spikes over multiple time steps. Their use of a biologically plausible neuron model makes them well-suited for asynchronous and sparse data processing in edge workloads. However, this temporal and stateful behavior introduces burst-driven execution patterns and inter-step dependencies, posing significant challenges for GPU-based deployment. Existing GPU schedulers, optimized for dense and feedforward DNNs, fail to fully exploit the irregularity and sparsity of SNNs, leading to poor resource utilization and unpredictable latency under dynamic workloads. To address these challenges, we propose MPL-schedule, a multi-preemptive, priority-aware scheduling scheme that dynamically coordinates heterogeneous SNN tasks on GPUs. Our approach formulates the scheduling process as a reward maximization problem under resource constraints, integrating adaptive priority computation, overhead-aware time-slicing, and state-preserving preemption. We evaluate MPL-schedule across diverse scenarios, including task complexities, spike arrival patterns, and time-slice configurations. Compared to state-of-the-art baselines, our method improves throughput by up to 15.0
Wireless indoor localization and tracking systems often implicitly assume that all anchors are continuously available. However, this assumption is often impractical under bandwidth and energy constraints. With limited wireless resources, tracking systems must balance localization accuracy with sensing and communication overhead. Therefore, anchor scheduling is crucial as it enables the system to activate only the most informative anchors for a better accuracy-resource tradeoff. This paper proposes a self-supervised framework for joint tracking and anchor scheduling by first utilizing a physics-guided deep state space model to infer target positions and uncertainty without ground truth trajectories. Building on the learned latent representation, we further design an actor-critic-based reinforcement learning scheduler that proactively selects a resource-budgeted set of informative anchors for future ranging. Experiments on real-world ultra-wideband datasets demonstrate that the proposed method achieves a mean absolute localization error of 0.458 m, outperforming classical filters and learning-based baselines. By activating only 3 of the 8 anchors, the proposed method reduces the number of ranging requests by 62.5
The deployment of large-scale satellite networks demands high-capacity and stable inter-satellite communication links, thereby driving the adoption of coherent optical satellite communication (COSC). However, relative satellite motion introduces Doppler shifts, which severely degrade link performance. To address this challenge, we discover the periodic correlation between Doppler shifts and the Gardner timing error detector (TED). Based on this correlation, we propose a novel frequency-offset estimation (FOE) algorithm that estimates frequency offset by computing the Gardner TED gain. The proposed algorithm surpasses the conventional FOE’s half-baud-rate limitation and maintains high estimation accuracy under strong noise conditions. We conduct 25-Gbaud DP-QPSK transmission experiments to evaluate the performance of the proposed FOE and to further examine its role in Doppler-shift compensation. The experimental results demonstrate that the proposed algorithm achieves a Doppler-shift estimation range 1.9 times that of conventional FOE algorithms. In addition, providing noise-robust Doppler-shift estimation improves receiver sensitivity by 0.6 dB and ensures successful signal demodulation even at an OSNR of 10 dB, which is where conventional FOE algorithms fail.
This paper proposes a hierarchical attitude planning method to address the complex attitude planning problem for multi-rigid-body spacecraft equipped with multiple laser communication terminals under constraints. The proposed method comprises three key steps: attitude decomposition, constrained path search, and trajectory generation. First, attitude decomposition determines the desired spacecraft platform attitude and feasible terminal joint angles (satisfying range-of-motion limits) corresponding to the desired terminal pointing directions. Next, the multi-heuristic A* algorithm is employed to perform a search for a feasible attitude and joint angle sequence within the high-dimensional configuration space satisfying all pointing constraints. Finally, a novel Lie group SO(3)-based trajectory planning method is applied to generate smooth and dynamically feasible attitude trajectories connecting the discrete sequence, by solving a sequence of quadratic programs. Numerical simulations under different operational constraints demonstrate the effectiveness of the proposed method in generating feasible and smooth trajectories.
Group communication among first responders is critical to the success of public safety missions. The operational environment of such missions introduces unique requirements, including highly dynamic group membership, tolerance to communication disruptions, the need to maintain confidentiality, and enable auditability. However, commercial group communication applications are typically designed for networks that provide timely and reliable delivery, which do not necessarily satisfy these requirements. Thus, a group communication solution specifically designed for public safety missions—one that accounts for all of these requirements—is needed. We present NameShield, a secure and efficient framework for dynamic group communication in public safety and emergency response scenarios. NameShield adopts a holistic, layered architecture that integrates security, information dissemination, and networking to support reliable and confidential communication among highly dynamic groups of first responders. At the information layer, NameShield builds on a role-based graph namespace to realize an efficient publish/subscribe communication model that naturally captures organizational structure and group dynamics. The network layer of NameShield is intentionally decoupled from specific communication technologies and can operate over any substrate that provides multicast, making it suitable for network conditions during public safety missions. To maintain message confidentiality, NameShield incorporates a message-oriented key-policy attribute-based encryption (KP-ABE) mechanism that enables fine-grained, per-message access control and supports frequent membership changes without requiring expensive group rekeying. Together, these design choices allow NameShield to provide strong security guarantees while remaining flexible, scalable, and robust under the operational constraints of emergency response environments.
While deep neural networks (DNNs) are increasingly deployed on graphics processing units (GPUs) at the network edge, the performance interference caused by co-locating and executing multiple models on a single GPU is often overlooked, leading to high inference latency and excessive energy consumption. Pursuing the best performance faces several challenges, including characterizing the non-linear interference-incurred latency, dispatching the unpredictable inference workloads, and balancing the different trade-offs related to performance and accuracy. In this study, we first construct an interference-incurred latency model via real-world profiling and measurements. With insights from this model, we formulate a time-varying integer program to minimize the long-term total cost of the edge AI inference system, including the inference latency, the inference error rate, the query-dispatching communication cost, and the energy consumption, subject to resource and workload constraints. We then propose a set of polynomial-time online algorithms that continuously make fractional control decisions for each time slot based on the feedback from the previously applied decisions and strategically round these decisions into integers. We conduct a formal theoretical analysis and prove that both the time-averaged gap between the total cost incurred by our online decisions and the total cost of the series of one-shot offline optimums and the time-averaged constraint violation gradually diminish over time. Extensive experiments on real-world testbeds and datasets demonstrate that the proposed algorithms can reduce the total cost by 40
