Large Language Models (LLMs) have achieved remarkable performance across various artificial intelligence applications. However, current LLMs cannot be deployed directly on edge nodes due to their large number of parameters. Fortunately, model compression technology has been proposed to reduce the computational workload and memory usage of LLMs, enabling further edge-based LLM services. However, existing research typically concentrates on isolated compression algorithms and lacks a comprehensive perspective on how to leverage these techniques for practical, end-to-end LLM deployment in edge environments. In this survey, we review edge-oriented LLM compression techniques and software–hardware co-design strategies to enable efficient LLM deployment on resource-constrained edge systems and guide future research in this area. First, we analyze techniques for LLM compression from the perspective of cloud–edge collaborative intelligence, including model quantization, parameter pruning, and knowledge distillation. Second, we present several hybrid model frameworks tailored to dynamic, heterogeneous edge environments, based on model architecture, application scenarios, and combination selection. Third, we further refine a four-layer software–hardware codesign and an overhead-aware LLM deployment optimization. Finally, we discuss the challenges of current model compression approaches and offer insights into future research directions, with a focus on edge-based LLM services.

Federated Learning (FL) faces critical challenges in heterogeneous and resource-constrained environments, including device diversity, high communication overhead, and training delays. Therefore, we propose FedKDC, a federated learning framework that integrates knowledge distillation with data compression to jointly optimize server bandwidth, client computation resources, and compression ratios, thereby minimizing training latency. In particular, FedKDC employs a Generative Adversarial Network (GAN)-based generator to produce synthetic data for knowledge transfer across heterogeneous models without sharing raw data, mitigating privacy risks. Then, FedKDC uses a loss-driven adaptive compression mechanism to adjust the minimum compression threshold based on training stability, reducing communication volume while maintaining accuracy. In addition, we further discuss the problem of resource allocation under system constraints, and uses Particle Swarm Optimization (PSO) algorithm to solve it. Based on the three real world datasets (i.e., Fashion-MNIST, CIFAR-10, and CIFAR-100), the experimental results demonstrate that FedKDC reduces communication cost by up to 17% and training time by 8%. This shows that FedKDC is effective for large-scale heterogeneous FL deployment while maintaining the accuracy of the model.

3D object detection plays a critical role in autonomous driving perception systems. While existing multimodal approaches typically employ independent feature processing streams followed by direct Bird’s Eye View projection for modality fusion, they encounter three critical limitations: insufficient cross-modal complementarity, feature misalignment across modalities, and inefficient computational workflows. To address these challenges, this paper proposes DCI-PRNet, a dual cross-modal interaction and reasoning framework that establishes deep synergistic relationships between 3D LiDAR point clouds and 2D multi-view images. The core innovations of DCI-PRNet lie in its dual cross-modal interaction module and multi-level progressive reasoning module. The dual cross-modal interaction module enables iterative feature refinement through alternating attention mechanisms and residual feature updating, effectively aligning spatial-semantic representations between point clouds and images. The multi-level progressive reasoning module implements detection refinement through cascaded decoder layers, where each stage progressively enhances detection confidence and localization precision via cross-modal aggregation. Experiments on the nuScenes dataset demonstrate significant performance improvements over conventional methods, achieving 71.7% mAP and 74.2% NDS.

The advent of 6G and the Metaverse has created a need for efficient real-time data processing and low-overhead communication. To address this challenge, we propose SemCom-MN, a semantic communication-enhanced Metaverse framework integrating an Edge Service Provider (ESP), Edge Sensing Units (ESUs), and Virtual Service Providers (VSPs). ESUs capture physical-world data, ESP manages semantic information, and VSPs create immersive virtual environments. To improve utility under heterogeneous information and computational requirements, we model semantic information trading as a three-stage Stackelberg game and prove the existence of a Nash equilibrium. Furthermore, to overcome high-dimensional dynamics and slow convergence in semantic trading, we develop a Diffusion Game Algorithm (DGA) combining strategic exploration with a game-theoretic denoising mechanism, achieving robust convergence. Simulation results show DGA increases system utility by 8.49%–33.94%.

Vehicle edge computing can effectively ensure the quality of experience for user vehicles (UVs), but road side units (RSUs) with limited resources may not be able to handle intensive tasks under high traffic conditions. In this case, worker vehicles (WVs) with idle resources can share resources to alleviate the pressure on RSUs. However, selfish WVs may be reluctant to share idle computation resources without any rewards. In addition, the optimization problems in previous research are relatively simple and cannot be applied to complex scenarios. To address the above challenges, we propose an incentive-driven partial offloading framework aiming to maximize social welfare. In particular, the computing service provider (CSP) managing RSUs first determines resource prices and offloading rates with UVs, while also determining contract terms with WVs. Then, it generates the optimal task scheduling strategy and notifies the UVs to offload tasks to the corresponding WVs. Considering that maximizing social welfare is a mixed-integer nonlinear programming (MINLP) problem, we design the hybrid proximal policy optimization (HPPO)-based task offloading and resource allocation algorithm (HORA) with a hybrid action space to directly solve the original problem. Finally, extensive simulation results show that HORA outperforms other baseline methods across various scenarios, and the contract terms meet the constraints of individual rationality (IR) and incentive compatibility (IC).

Barrier coverage problem is one of hot research topics in directional sensor networks (DSNs). Directional sensors could zoom their sensing ranges, within which the event detection probability decreases with the increase of the distance between the location of event and the sensor. Although the probability that an event is detected by a single sensor may be below the required criteria, the detection probability achieved jointly by two sensors can be above the required criteria. In this work, we study the barrier coverage problem of DSNs, taking into account the directional nodes' ability of adjusting their working directions and sensing ranges. We mainly propose a barrier construction scheme, which schedules the nodes to form multiple barriers by jointly determining which nodes jointly forming a barrier, and their respective working directions and sensing ranges. Comparing to existing works, the proposed scheme is able to form more barriers, leading to the increase of service lifetime.

We present the design and implementation of a new jumping robot for mobile sensor network. Unlike other jumping robots, the robot is based on a simple two-mass-spring model. After we throw it on ground, it can stabilize itself and then jump once. The detailed mechanism design including the load holding and self-stabilization are presented. Jumping heights and distances with different robot weights are measured and compared with calculated values from the two-mass-spring model.