
In Wireless Rechargeable Sensor Networks (WRSNs), multi-UAV systems provide an effective solution for wireless energy replenishment and real-time data collection. However, existing reinforcement learning-based UAV coordination methods often suffer from entangled state representations, unstable cooperative learning, and poor scalability under dynamic network conditions, making it difficult to jointly optimize information freshness and energy efficiency. To address these challenges, this paper proposes a Disentangled Multi-Agent Learning framework for Age of Information (AoI)-oriented UAV coordination in WRSNs. The framework factorizes UAV observations into semantically independent spatial, energy-related, and temporal subspaces, and integrates the resulting representations into a Multi-Agent Double Deep Q-Network (MA-DDQN) for cooperative trajectory optimization and adaptive task scheduling. A weighted K-Medoids clustering strategy is employed to determine UAV hovering points by jointly considering sensor distribution, residual energy, and task urgency, while an AoI-driven feedback mechanism updates clustering weights online to improve long-term coordination efficiency. In addition, a Privacy–Accuracy Adaptive Controller (PATC) introduces adaptive state perturbation during centralized synchronization to enhance privacy preservation and robustness. Extensive simulations demonstrate that the proposed framework outperforms representative heuristic and reinforcement learning baselines in AoI, energy efficiency, task completion ratio, and convergence stability while maintaining reliable decision quality, with average AoI and total energy consumption reduced by 11.3% and 4.7%, respectively, and task completion ratio improved by 3.2% relative to I-DDQN.
Few-shot knowledge graph completion (FKGC) has emerged as a crucial research direction for real-world applications, aiming to predict new facts in knowledge graphs (KGs) under the limited instances per relation. However, existing FKGC approaches generally overlook the deeper semantic interactions between the support set and the task-specific relations, leading to representations that are not sufficiently discriminative, especially when multiple entity pairs share the common relation. To address this issue, we propose a novel model, Fusing Topological and Semantic information (FTS), which instantiates a sequential topology-to-semantics modeling pipeline. Specifically, a topology learner adopts an adaptive gating mechanism, a topology-aware attention mechanism, and a relation-adaptive attention mechanism to gain the topology-aware entity embedding by aggregating relations and entities from local graph neighborhoods. Furthermore, a topology-augmented semantics learner applies a semantic-aware attention mechanism and a relation-adaptive attention mechanism to dynamically integrate topological features with semantic features to gain the semantics-aware entity representation, yielding entity pair representations. An adaptive matching processor is then utilized to effectively compare query triples with support instances for completion. Experiments on two datasets demonstrate that FTS outperforms several baselines, achieving competitive performance in FKGC. The code is publicly available at: https://github.com/jiawei1224/few-shot-knowledge-graph-completion
Foundation models are reshaping computer vision by replacing task-specific pipelines with adaptable models pretrained mainly with self-supervised learning on large-scale data. This review surveys 84 foundation models identified through a structured search and analyzes their use across natural images, medicine, and remote sensing, including both two-dimensional and three-dimensional settings. The article proposes a unified categorization based on training data modality, learning objective, architectural design, and application domain, covering single-modality vision models, multi-modality vision models, vision-language models, and three-dimensional vision models. Across these groups, the review identifies recurring design patterns linking modality, architecture, pretraining strategy, and target downstream application. The study also examines evaluation practice and finds substantial fragmentation in benchmark choice, reported metrics, and experimental protocols, with the strongest inconsistency in medical models and the lowest in natural imaging models. Based on these findings, the review provides practical design and evaluation guidelines and discusses key barriers to real-world adoption, including efficiency, interpretability, and deployment complexity.
A considerable challenge of dividing data for multi-view clustering tasks is exploiting multiple-view information in unison. To sufficiently leverage cross-view information, tensor-based representation learning approaches attract much attention, which represents the data in a tensor and encourages essential representation learning. Many existing methods are based on self-expressive tensor learning to calculate essential affinity matrix. They do, however, may encounter instability due to outliers or noise contaminations. To that end, this paper utilizes the dual high-order relations to establish the reliable multi-view integrating and further proposes Enhanced Tensorial t-product Representation for Multi-View Clustering (ETRMVC) to effectively represent multi-view data for accurate data partition. First, ETRMVC fully captures cross-view high-order correlation via tensorial t-product representation. Second, we incorporate high-order proximity to the regularization term, which can explore and correct within-view high-order relations. ETRMVC effectively realizes the combination of cross-view and within-view dual high-order relations in the adaptive learning model. Extensive empirical studies on six benchmark datasets well demonstrate that ETRMVC achieves superiority over other state-of-the-art approaches.
Multiphysics systems are commonly governed by coupled partial differential equations (PDEs), in which interactions among physical processes can complicate stable and interpretable modeling. This paper presents AdaMPSR, an adaptive multiphysics symbolic regression framework that discovers physically consistent, closed-form approximations for coupled PDE systems. A constraint-complexity scoring mechanism determines the priority with which variables are decoupled and is integrated with residual-driven symbolic optimization, thereby transforming the original coupled system into a sequence of tractable subproblems. To improve adaptability and convergence efficiency, a structure-entropy feedback mechanism adjusts the symbolic search strategy according to the structural diversity of the candidate population, which accelerates convergence and reduces the likelihood of becoming trapped in local optima. We also introduce two quantitative metrics, the Decoupling Discriminability Index (DDI) and Decoupling Efficiency (DE), to evaluate the decoupling process and convergence behavior. AdaMPSR is evaluated on several benchmark multiphysics problems and a solid oxide fuel cell (SOFC) case study. Experimental results show that AdaMPSR achieves competitive or superior performance across the evaluated cases in accuracy and convergence speed while retaining interpretable symbolic forms, providing an effective approach to symbolic modeling of coupled physical systems.