In online medical recommendation systems, modeling heterogeneous interactions between patient needs and medical resources is still challenging under sparse and few-shot scenarios. An Iterative Intra–Inter-Node Contrastive Learning (I3CL) framework is proposed, which alternately updates an intra-node contrastive module for attribute-level discrimination and an inter-node contrastive module for cross-graph alignment, forming a closed-loop optimization process. The BioBERT encoder is frozen to retain medical knowledge and mitigate overfitting. Experiments on two datasets show that I3CL improves Recall, AUC, and F1-score by 5.64
Multi-modal document question-answering aims to search multi-modal evidences distributed across multiple pages for answer generation. Existing methods can be categorized into Large Vision-Language Model (LVLM)-based and Retrieval-Augmented Generation (RAG)-based methods. However, the former were susceptible to hallucinations, while the latter struggled for inter-modal disconnection and cross-page fragmentation. To address these challenges, a novel multi-modal RAG model, named MHier-RAG, was proposed, leveraging multi-modal cues across multi-pages to reason accurate answers. A hierarchical indexing method with the integration of flattened in-page and topological cross-page chunks was designed to jointly establish in-page associations and cross-page dependencies. By means of joint similarity evaluation and large language model (LLM)-based re-ranking, a multi-granularity retriever, including page-level parent page retrieval and document-level summary retrieval, was proposed to encourage multi-modal connection and long-distance reasoning. Experimental results performed on MMLongBench-Doc and LongDocURL demonstrated the superiority of MHier-RAG in understanding modality-rich and multi-page documents. Our code is available at https://github.com/Gzy1112/MHier-RAG .
Intra-modality cross-dataset medical image segmentation faces two key challenges: distribution shift and boundary uncertainty. Specifically, due to device-inherent noise, contrast fluctuations, and geometric deformations of lesions, a significant marginal distribution shift occurs between the source and target domains, which severely limits the generalization capability of existing methods. To address these issues, this paper proposes a lightweight Intrinsic Feature Consistency Learning Network (IFCLNet). Specifically, we propose a Domain-Agnostic Semantic Routing (DASR) module, which leverages a soft-thresholding mechanism to actively filter domain-specific physical noise to purify semantics. Building upon this, a Dynamic Scale-Aware Aggregation (DSAA) module is proposed, wherein convolutional weights are dynamically generated based on feature statistics, enabling structure-adaptive alignment across multi-scale lesion patterns. Furthermore, an Intrinsic Perturbation Consistency (IPC) module is proposed as the core regularization component, which incorporates multi-view perturbations and enforces consistency in the deep feature space, thereby mitigating boundary ambiguity and promoting the learning of anatomically meaningful boundaries. Experimental results demonstrate that, compared with state-of-the-art methods, IFCLNet achieves superior performance across nine datasets with reduced computational cost.
Text-to-SQL systems have transitioned from simple machine translation models to complex reasoning frameworks as database schemas grow in scale and ambiguity. Despite the impressive capabilities of Large Language Models, one-shot generation often fails to produce correct SQL in real-world scenarios where schemas are wide, values are noisy, and user intent is under-specified. This survey introduces the Agentic-SQL Taxonomy, an autonomy-based classification that reevaluates existing methods through the lens of inference complexity, categorizing research into single-turn generation, iterative refinement, and multi-agent collaboration. Unlike prior surveys that organize the field by training paradigm — In-Context Learning vs. Supervised Fine-Tuning — we focus on the structure of inference itself, arguing that the level of autonomy a system exercises during query generation is the primary driver of modern performance. We analyze how these pipelines bridge the performance gap on challenging benchmarks: leading models still produce incorrect execution in nearly 40
Vision-Language Models (VLMs) achieve strong performance on many benchmarks, yet a basic reliability question remains underexplored: when visual evidence conflicts with commonsense, do models follow what is shown or what commonsense suggests? In such conflict cases, models may fail by reverting to the commonsense answer instead of following the visual evidence. We term this phenomenon Commonsense-Driven Hallucination (CDH). To evaluate it, we introduce CDH-Bench, a benchmark designed to create explicit visual evidence–commonsense conflicts. CDH-Bench covers three dimensions: counting anomalies, relational anomalies, and attribute anomalies. We evaluate frontier VLMs under binary Question Answering (QA) and multiple-choice QA, and report metrics including Counterfactual Accuracy (CF-Acc), Commonsense Accuracy (CS-Acc), Counterfactual Accuracy Drop (CFAD), Commonsense Collapse Rate (CCR), and Relative Prior Dependency (RPD). Results show that even strong VLMs exhibit Commonsense-Driven Hallucination (CDH): under visual evidence–commonsense conflict, models frequently revert to the commonsense alternative rather than follow the visual evidence. CDH-Bench provides a controlled diagnostic of visual fidelity under visual evidence–commonsense conflict. Data are available at and .
A cross-site scripting (XSS) attack is one of the major web threats and security issues in web applications. Black-box web scanners are widely used to detect XSS vulnerabilities, but their effectiveness is often limited by a rigid testing strategy. This paper designs a scheme, named Automated XSS Vulnerability Detection (AUTOXSS), which trains an intelligent agent based on the Advantage Actor-Critic (A2C) reinforcement learning (RL) algorithm. This agent can automatically generate XSS attack payloads and perform vulnerability detection on target applications. AUTOXSS proposes a novel attack vector combination strategy and incorporates Web Application Firewall (WAF) feedback mechanisms to optimize the capability of the A2C algorithm, thereby improving the efficiency of the generation of effective attack vectors. Experimental results demonstrate that AUTOXSS outperforms comparable scanning tools across four benchmark tests, successfully detecting 286 XSS vulnerabilities with zero false positives while maintaining the lowest request frequency. The proposed methodology effectively enhances the overall efficiency of XSS vulnerability detection.
One-shot neural architecture search (NAS) is an efficient and widely adopted method for powerful networks. However, it suffers from inconsistency, where sub-networks inheriting supernet weights exhibit different performance rankings compared to networks trained independently from scratch. To address this issue, we present a theoretical analysis showing that the discrepancy in performance ranking between sub-networks and their optimal counterparts is strongly correlated with the gradient norm and weight variance of the supernet. Based on these insights, we propose two regularization techniques and introduce Oneshotimizer, a novel method designed to enhance the consistency and effectiveness of one-shot NAS. Experiments demonstrate that Oneshotimizer significantly improves both ranking consistency and search performance. Specifically, it achieves a 14.8 over SPOS on NAS-Bench-201, and a state-of-the-art 77.8
Underwater biological target detection is a critical technology for enabling the automated harvesting capabilities of underwater robots [1]. To address the challenges in detecting benthic organisms in marine pastures [2], including small target sizes, mutual occlusion between targets, and misdetections or missed detections caused by low-quality marine images that make it difficult to distinguish benthic organisms from the ocean background, this study proposes ReFM-YOLO, a detection algorithm for benthic organisms in marine pastures based on an improved YOLO11n model. Firstly, to enhance the model’s ability to detect small target organisms, a novel Re-Calibration Feature Pyramid Network (Re-CalibrationFPN) is designed in the neck network; Secondly, to enhance the model’s ability to extract features from targets of different scales, a novel shared-parameter convolution module, FPSConv, is proposed to replace the original SPPF module. The FPSConv module utilizes the shared-parameter convolution operations for feature extraction, enabling the capture of finer-grained features while reducing the model’s parameter count; Finally, to enhance the model’s ability to detect occluded organisms, an occlusion-aware attention mechanism is integrated into the head, and a novel detection head, MultiSEAM-Head, is proposed. Experiments were conducted on the DUO dataset to validate the performance of the improved model. The results demonstrate that the improved model exhibits better detection performance. Compared to YOLO11n, the proposed algorithm achieves a 2.3
Particle swarm optimization (PSO) is widely utilized in various optimization problems. The traditional PSO optimization algorithm relies on inertial, cognitive, and social factors to update the positions and velocities of the particles. It is difficult to produce an effective downward direction due to the curse of dimensionality for large-scale optimization problems. The local and global optimal solutions are unchanging for a long time, and then the particle's velocity tends to vanishing or exploding. Meanwhile, zero-order optimization is a straightforward, effective, and robust optimization algorithm that demonstrates an apparent convergence bound even in high-dimensional optimization. However, this kind of pointto-point optimization search algorithm tends to exhibits premature convergence at suboptimal equilibria. Therefore, we propose a particle swarm optimization based on zero-order optimization, namely PSO-ZO, to solve large-scale optimization problems. It dynamically adjusts the search stride and updates the direction of particle evolution by integrating the zero-order optimization sampling method to enhance the search efficiency. Experiments conducted in 11 benchmark functions demonstrate the efficacy of our algorithm in solving large-scale optimization problems. Additionally, this paper calculates the cosine similarity between the gradient direction obtained by PSO-ZO and the gradient direction of the objective functions to prove the algorithm's feasibility.
Graph co-clustering aims to simultaneously group two types of nodes by uncovering patterns within bipartite graphs. While existing co-clustering methods are effective for static bipartite graphs, they are ill-equipped to handle temporal bipartite graphs, where time-ordered edge sequences occur at irregular intervals. This leads to a failure in capturing essential temporal dependencies. To address this, we propose a novel Co-Clustering algorithm for Temporal Bipartite Graphs (ccTBG), designed to uncover temporal patterns and identify co-clusters within these graphs. ccTBG employs distinct generative mixture models for the two types of nodes to model cluster membership and leverages the Hawkes process to capture the temporal patterns of co-clusters from observed edge sequences. Additionally, we develop a variational algorithm for efficient inference and parameter estimation. Extensive experiments on synthetic datasets validate the effectiveness of ccTBG in co-clustering relative to existing methods, while analysis of real-world datasets reveals the discovered co-cluster structures.
In recent years, contrastive learning has emerged as a powerful self-supervised approach for training large models. A key challenge in achieving high-performance contrastive models is the effective construction of high-quality positive sample pairs. Existing methods typically rely on RandomCrop of the same sample, often overlooking the semantic content and fine-grained details of the image that could be beneficial for model training. In this paper, we introduce a novel FilteringCrop technique to improve positive sample pair construction. We begin by developing a target rough localization algorithm, leveraging prior knowledge of the image’s semantic content, to facilitate coarse localization through image filtering prior to training. Next, we present a cropping algorithm designed to generate positive pairs that aid in more efficient model training. Additionally, we propose the Target Center Momentum Optimization algorithm to continuously refine the cropping center throughout the contrastive learning process. Our method demonstrates a 0.4 https://github.com/PPChuan/FilteringCrop .
Network anomaly detection in the Internet of Vehicles (IoV) struggles with non-IID data caused by dynamic environments, hardware heterogeneity, and communication noise. Traditional methods relying on feature-label correlations suffer from noise-induced spurious associations and performance degradation under distribution shifts. We propose CDS-IoV-NAD, a causal deep stable learning framework with two innovations: A dynamic bit-width binary encoding (DBBE) mechanism that losslessly maps continuous features to 0,1 space through robust Z-score normalization and differentiable encoding, enabling causal intervention compatibility. A dual-regularized causal weight learning model that amplifies causal features via propensity score weighting while suppressing non-causal pseudo-correlations using L2 and stability regularization. Experiments on CIC-IoV24 and UNSW-NB15 show CDS-IoV-NAD achieves > 5
Cage damage in rolling bearings, due to its characteristics of abrupt faults and rapid degradation, is highly to result in bad accidents. This paper tries to solve the key challenge in incipient cage fault detection: how to capture the subtle abnormal changes of system states as early as possible in a noisy environment. A dynamic model-assisted multi-domain rule adaptation anomaly detection method (MD-RAAD) is proposed for extracting sensitive features. The core of this method is to integrate normal-state simulation data generated by the dynamic model of the target bearing into the anomaly detection transfer learning framework. First, a parameter inversion model is developed to correct the dynamic model so as to obtain simulation data that is more consistent with the actual data distribution. Second, a novel multi-domain adversarial neural network with rule adaptation mechanism is proposed to integrate auxiliary bearing data, target bearing data, and simulation data. Guided by the "noise-free" simulation data, the network adaptively extracts domain-invariant features that are more sensitive to incipient faults. The explicit rule adaptation constraint enables effective transfer at the detection rule level, which will enhance the discriminative capability and robustness of feature representations. Results on public bearing datasets demonstrate that the proposed method can find incipient fault occurrence in an earlier location with a lower false alarm rate.
Retrieval-Augmented Generation (RAG) is a pioneering approach that greatly improves the factual accuracy and contextual relevance of model outputs by seamlessly incorporating background knowledge into the generation process. The efficacy of RAG hinges on the selection of pertinent evidence that not only exhibits semantic alignment with the query but also demonstrably contributes to the downstream task. However, a significant limitation of existing retrieval methods lies in their tendency to conflate relevance with utility to generate accurate and coherent answers. This conflation often leads to the retrieval of information that, while topically related, fails to provide substantial benefit for answer generation. In this work, we address this critical challenge within the context of visual document question answering (VDQA), a task demanding precise integration of visual and textual information from complex document layouts. Through a rigorous analysis of the interplay between relevance and utility, we demonstrate that while relevant information exhibits strong semantic alignment with queries, it frequently falls short of providing the necessary cues for generating accurate and contextually coherent answers in VDQA. Specifically, we observe that highly relevant document segments may contain redundant or distracting information, hindering the model's ability to extract the essential visual and textual cues required for accurate question answering. Based on these findings, we propose a novel utilitydriven retrieval method that explicitly prioritizes evidence based on its predicted contribution to answer quality. Our approach incorporates a learned utility scoring mechanism that assesses the potential of retrieved document segments to facilitate accurate generation.
Graph-structured Deep Hierarchical Reinforcement Learning frameworks have made notable progress by modeling states, actions, and subtasks as graph nodes and edges, enabling rich semantic reasoning and capturing task dependencies. However, existing methods often suffer from inefficient experience replay and inaccurate subgoal relabeling, especially in sparse-reward and highdimensional environments. This paper proposes a hybrid prioritized sampling mechanism based on state-goal distance and a temperature-regulated weighting strategy, along with an attention-driven subgoal refinement module integrated with a dual Q-network. Experiments across multiple benchmarks show that DAG-HRL achieves faster convergence and higher task success rates than prior methods.
This paper investigates prognostics and health management (PHM) for data communication in industrial environments, focusing on serial communication devices. The article first provides a comprehensive review of health status prediction methods in circuit-related fields and summarizes the existing issues, such as outdated technology and the challenges in adapting those methods to real-world industrial environments. To address these gaps, we propose our targeted analysis and resolution strategies toward the specific domain. By constructing a dataset based on real-world communication voltage and bit error rate (BER) data, we establish a quantitative health status metric. Subsequently, an improved STMDLinear prediction method based on feature decomposition and fusion is introduced. In this model, we perform wavelet decomposition, reconstruction, and compression on the input data to enhance key information while preserving as much of the original data as possible and reducing storage requirements. Then the remainder vector and trend vector are processed through a TCN module and an attention module. These modules further extract features and fuse them at the end of our model, ultimately producing the predicted BER results. Finally, the network is applied to the real-world communication dataset acquired by us. The proposed model achieves significant performance improvements over existing advanced time series prediction networks, with over 20
Robust fall detection using acoustic signals remains a significant challenge, primarily due to background noise and non-line-of-sight (nLoS) propagation in dynamic indoor environments. This paper presents AP-Fall, a novel system that combines both audible and inaudible acoustic signals for environment-adaptive fall detection. We propose a decision fusion framework that integrates the audible and inaudible feature streams based on the signal-to-noise ratio (SNR). Specifically, we apply a joint optimization method with the minimum mutual information (MMI) criterion to determine a set of optimal SNR partition boundaries and corresponding optimal stream weights. Experimental evaluations show that AP-Fall achieves an accuracy of over 93
Spectral clustering, a fundamental unsupervised learning algorithm, is extensively applied in machine learning and data science. With advancements in fairness research, fair spectral clustering has gained prominence, which mainly focuses on group fairness and individual fairness to reduce the decision bias on sensitive attributes. Existing algorithms typically rectify inequalities for specific individuals or groups through resource reallocation, but often face challenges with disproportionately large or small cluster sizes. This paper introduces GameTheoretic Optimization for Scale Fair Spectral Clustering (GTSC) to enhance fairness in spectral clustering with unbalanced cluster sizes. The algorithm models data points as strategic agents in a game and uses a dynamic competition mechanism and Gini coefficient adjustment to balance the cluster size. By introducing fairness constraints in the objective function, the penalty for cluster size differences is dynamically adjusted. In addition, a dynamic threshold merging strategy is used to coordinate the optimization of cluster quality and scale fairness. Experimental results demonstrate that GTSC outperforms existing scale fairness algorithms across various datasets, achieving clustering outcomes that meet the expected criteria.
This paper presents an optimization design method for heliostat fields by combining physical modeling with intelligent optimization. A generalized mirror surface diagram is derived using the cone-ray model, and optical efficiency and shadow-blocking areas are calculated via the micro-element integration method and quadrat counting principle. A Particle Swarm Optimization (PSO) algorithm is constructed, treating each heliostat as a particle and guiding the search in a twodimensional distribution space based on historical best fitness values. To enhance the model's flexibility, three additional parameters-length, width, and installation height-are introduced, expanding the solution space to five dimensions. The improved PSO model enables dynamic global optimization of both positional and dimensional parameters, resulting in a heliostat field layout that meets the required power output with higher overall efficiency.
Accurate fault detection and root cause location are crucial to ensuring the efficiency and reliability of power distribution networks. The widespread deployment of sensors, such as Phasor Measurement Units (PMUs), has substantially increased multivariate time-series data, necessitating advanced methods to capture complex spatiotemporal dependencies of power distribution systems. However, existing approaches often overlook the simultaneous modeling of spatial and temporal dependencies, leading to increased false alarms and misdiagnoses. We propose Multivariate Spatial-Temporal Graph Convolutional Informer (MSTGCI), designed for long-term fault detection in complex power distribution networks, incorporating a novel Graph Convolutional Informer (GCformer) to better capture spatial and long-term temporal dependencies. Additionally, the correlation relationship learned by Multivariate Time-series Graph Learning (MTGL) and a Variational Autoencoder (VAE)-based fault scoring component are used to identify fault detection and precise root cause localization. Experimental evaluation in public datasets demonstrates that MST-GCI significantly outperforms eight stateof-the-art baseline models with respect to different types of fault detection accuracy. Further experiments show that MST-GCI effectively localizes root causes buses in complex power distribution network.
