Binary code analysis is essential for software security applications. While existing decompilation tools require manual adjustments and have limitations in inter-function analysis, Large Language Models (LLMs) like Meta's LLM Compiler FTD show promise for assembly-to-LLVM-IR conversion. However, current evaluations focus on syntactic similarity (BLEU scores) rather than structural preservation crucial for static analysis. This paper evaluates LLM Compiler FTD's applicability to static taint analysis from a structural preservation perspective. We define graph edit distance-based metrics to compare Control Flow Graphs (CFGs) and Data Dependence Graphs (DDGs) before and after decompilation, using the CASTLE dataset for systematic evaluation.

Results on 196 C programs reveal hierarchical preservation: CFG similarity (0.969), DDG similarity (0.863), and syntax (0.776). Notably, 79% of syntactically imperfect functions maintained perfect CFG preservation (average 0.956), demonstrating structural robustness. Major failure patterns include block/label inconsistencies (33.3%) and type mismatches (20.8%). Our findings demonstrate that LLM-based decompilation preserves essential structural properties for static analysis, with CFGs showing high resilience. This work extends decompilation evaluation from syntactic fidelity to structural preservation, establishing quantitative metrics for assessing decompilation quality in static analysis contexts.

Efficient anomaly detection and interpretable candidate root cause analysis are of great importance for ensuring the stable operation of cluster server systems. However, cluster server systems are usually characterized by large numbers of variables, complex correlations, and evident functional structures, and it remains challenging to effectively model variable relationships and explain anomaly propagation processes. To address this issue, this paper proposes a causal structure-driven unsupervised anomaly detection framework for cluster server anomaly detection. The proposed method adopts a block-wise graph search strategy to construct causal structures among system variables and utilizes the resulting causal matrix to guide variable-state representation learning and variable anomaly detection. Furthermore, the detected anomalous variables are mapped back onto the causal structures to construct anomaly propagation subgraphs for identifying candidate root cause variables with upstream structural features. Experimental results on a real-world cluster server dataset demonstrate that the proposed method achieves superior overall anomaly detection performance compared with existing methods and provides a structured basis for candidate root cause analysis and anomalous event interpretation.

Large language models (LLMs) have achieved competitive performance in machine translation (MT) via in-context learning (ICL), yet the selection of in-context examples remains an open question that significantly affects translation quality. Existing heuristic-based selection methods lack a targeted mechanism to address translation errors like omissions or mistranslations, which are particularly severe for China's minority languages due to their unique linguistic characteristics, limited resources, and distinct structural challenges. This paper proposes a self-reflection mechanism to optimize example selection for LLM-based MT with the explicit goal of correcting such errors, named Self-Reflection for In-Context Example Selection (SR-ICES). Our approach comprises four sequential stages: (1) an LLM generates an initial translation via a task-specific prompt; (2) the LLM conducts self-reflection to identify erroneous source segments using its internal linguistic knowledge; (3) these erroneous segments are used as a query to retrieve top-N lexically relevant source-target pairs from a parallel corpus; (4) finally, the retrieved examples are fed into the identical LLM again to generate refined translations. Comprehensive experiments on three China's minority language MT tasks, i.e. Uyghur-Chinese (Uy-Zh), Tibetan-Chinese (Ti-Zh), and Mongolian-Chinese (Mn-Zh), show that SR-ICES consistently outperform several competitive baselines, and verify the effectiveness of leveraging LLMs' self-reflection capabilities for in-context example selection, providing a scalable and language-agnostic solution to boost MT performance by closing the loop between translation generation and example selection.

Remote sensing segmentation underpins land classification and resource management. However, significant variation in object scales, high scene complexity, and blurred boundaries in remote sensing images make accurate segmentation particularly challenging. Traditional convolutional neural networks are limited in long-range modeling capability, while Transformer-based architectures often incur high computational costs, leading to a trade-off between prediction quality and inference speed. To address these issues, we propose an efficient dual-branch segmentation network for remote sensing images. The main branch extracts fine-grained local spatial features via a U-Net backbone, while an auxiliary branch based on the visual state space model captures long-range global context. To achieve adaptive and effective feature integration, a learnable gated fusion mechanism is designed to dynamically fuse multi-level features. Furthermore, a lightweight atrous spatial pyramid pooling module is introduced to enhance multi-scale object perception without introducing heavy computation. We conduct experimental validations on two publicly available remote sensing datasets, verifying that the proposed model achieves favorable segmentation accuracy and possesses great potential for practical deployment.

This study investigates the use of the Smooth Overlap of Atomic Positions (SOAP) descriptor as a preprocessing method for 3D point cloud data. SOAP captures local geometric features while being inherently invariant to rotation, translation, and mirror symmetries. This makes it particularly useful in applications where preserving physical symmetries is important and where traditional data augmentation techniques are typically required. By encoding symmetry directly into the representation, SOAP reduces the reliance on such augmentation. We apply SOAP preprocessing to the ShapeNet-Part dataset and compare its performance with PointNet and PointNet++, which are widely used deep learning models for point cloud classification. Our results show that, in most cases, the SOAP-based approach achieves similar or better accuracy on small training datasets than PointNet or PointNet++ while avoiding data augmentation. These findings highlight the potential of symmetry-aware descriptors like SOAP for efficient and accurate 3D point cloud analysis. We further propose a hybrid approach in which pre-computed SOAP spectra are injected into the intermediate feature representations of PointNet. The hybrid models achieve the highest accuracy at full data utilization, particularly for geometrically challenging categories, demonstrating the complementary nature of descriptor-based and learned representations.

Remote sensing image dehazing is challenging due to complex scenes and uneven haze, while spatial-based methods struggle to recover high-frequency details. To address these issues, we propose a dual-domain collaborative dehazing network that integrates spatial-domain representation with frequency-domain enhancement. Specifically, the Discrete Wavelet Transform is employed to decompose features into low-frequency and high-frequency components, enabling targeted restoration. A Spatial Processing Module is designed to capture large-scale contextual dependencies, while a High-Frequency Restoration Module focuses on recovering texture and edge information in remote sensing imagery. Furthermore, a Cross-Domain Interaction Module is introduced to facilitate effective information exchange between spatial and frequency domains. Extensive experiments demonstrate that the proposed method achieves superior performance in terms of structural preservation and detail recovery.

Dancing to music with lyrics is a popular form of expression. While it is generally accepted that there are relationships between lyrics and dance motions, previous studies have not explored these relationships. A major challenge is that the relationships between lyrics and dance motions are not constant throughout a song but are instead localized to specific parts. To address this challenge, we hypothesize that lyrics and dance motions that co-occur across multiple songs are related. Based on this hypothesis, we propose a novel data-driven method to detect the parts of songs where meaningful relationships between lyrics and dance motions exist. We use clustering to transform lyrics and dance motions into symbols, enabling the calculation of co-occurrence frequencies and detection of significant correlations. The effectiveness of our method is validated by a dataset of time-synchronized lyrics and dance motions, which showed high correlation values for emotionally salient lyrics such as “love”, which is expressed in heart-shaped motions. Furthermore, using our relationship detection method, we propose a method for retrieving dance motions from lyrics that outperforms previous text-to-motion retrieval methods, which focus on prose and non-dance motions.

Medical image fusion has recently emerged as a promising research direction because of its potential to support clinical diagnosis. Generating high-quality, clinically meaningful fused images facilitates the integration of complementary information from multiple imaging modalities, thereby potentially supporting clinical interpretation. Despite recent advancements, existing methods often suffer from the loss of salient features and suboptimal preservation of structural details, particularly at edges and corners. To address these challenges, we propose RCDFusion, a framework for multimodal medical image fusion. The proposed architecture leverages reconfigured residual blocks, inspired by ResNet, to extract hierarchical modality-specific features from each input modality. Features from different modalities are then jointly refined through a co-attention mechanism and fused via dynamic convolution, helping preserve critical information from each source image. Experiments were conducted on two benchmark datasets, PET-MRI and SPECT-MRI, and performance was evaluated using several objective metrics. Results on the two evaluated datasets indicate that RCDFusion performs strongly relative to the evaluated baselines.

Software libraries are increasingly being utilized to enhance software development efficiency and quality. Like most software, the libraries might also lose the backward compatibility in version update to add new features and fix bugs for maintenance. The backward incompatibility would cause errors in client software that uses the library. Hence, the library project properly informs the new library call statement formats to the clients in the update. However, since the syntax of the call statements that suffer from backward incompatibility varies among clients, it is difficult for library developers to comprehensively grasp all syntaxes used by clients. Then, the library developers may not always specify the correct syntax of call statement formats in the release document for new version. Previous studies proposed methods whether or not the released library lose the backward compatibility based on test execution results. This approach has an issue that clients with insufficient tests may not easily recognize errors caused by backward incompatibility. This study proposes a static analysis-based identification of clients that suffer any issues due to updating to a library version that loses backward compatibility. Specifically, similar to a previous study, our approach intentionally updates the library versions used by clients and then identifies clients whose tests fail after the library version update. For these clients, this study generates patterns using regular expressions to identify the import statement formats and function call statement formats common across clients. Using these patterns, we identify clients that suffer from backward incompatibility among those clients whose tests succeed both before and after the update. As a case study, we target clients using any of 8 library versions known to lose backward compatibility upon version updates from previous versions. We then investigate clients whose tests succeeded after the library update. Consequently, even among clients passing tests after updating to versions that lose backward compatibility, we found 15% to 79% of them used both import statement and call statement syntax affected by the backward incompatibility.

Recent advances in extended reality (XR) and immersive remote operation systems have increased the demand for low-latency and high-quality video streaming. Spherical displays provide wide field-of-view visualization that enhances spatial awareness. However, they inherently introduce geometric distortion that must be corrected in real time. Moreover, transmitting raw high-resolution video streams imposes significant bandwidth requirements, making perceptually efficient compression essential. This paper presents a low-latency FPGA-based streaming architecture for immersive remote operation using a spherical display. The proposed system integrates real-time distortion correction, gaze-based compression driven by eye-tracking information, and buffer-efficient memory management within a single FPGA pipeline. These components reduce end-to-end latency and resource utilization. By exploiting the geometric symmetry of the spherical display for distortion correction and selectively downsampling regions outside the user's gaze, the proposed system reduces on-chip buffer size by 52% and video bandwidth by 92.6% compared with raw Full-HD streaming. Implementation on a Zynq UltraScale+ FPGA demonstrates an endto- end latency of 79.8-96.5 ms from gaze acquisition to spherical-display projection. User studies further confirm that the proposed gaze-based compression has no measurable impact on human reaction time to peripheral visual stimuli. These results indicate that substantial bandwidth reduction can be achieved without degrading perceptual responsiveness in immersive XR streaming scenarios.

Elastic Coarse-grained reconfigurable arrays (CGRAs) are one approach to sustain processor throughput while maintaining power efficiency. They are programmable hardware platforms comprising processing elements and switching blocks. On elastic CGRAs, computational throughput may degrade when the mapped dataflow graphs (DFGs) contain reconvergent and imbalanced paths. However, the impact of throughput degradation and architectural extension to this issue has not been addressed. In this paper, we quantitatively investigate the impact of throughput degradation. We also propose a CGRA architectural extension that places a FIFO buffer in the switching block and an associated design flow to tackle this issue. We applied the proposed architecture extension to RIKEN CGRA and evaluated the impact of FIFO depth, placement mode, and number of FIFO buffers on computational throughput, power consumption, and circuit area. Using four applications with varying degrees of complexity, the proposed extension achieved higher performance per watt than the baseline under nearly all conditions.

We propose a deep learning algorithm for single image reflection removal. The proposed network leverages tree-structured attention modules to capture multi-scale structural coherence within transmission and reflection layers and facilitate layer separation. Feature refinement modules further exploit the inter-layer independence to suppress reflection artifacts in intermediate features. Comparative experiments demonstrate that our method achieves state-of-the-art performance with lower computational complexity.

Cloud computing enables multiple users to efficiently share limited computing resources through virtualization. With the increasing popularity and performance of GPUs, technologies such as vGPU, which virtualizes GPUs for multiple users, and passthrough, which directly assigns PCI devices to virtual machines, have been introduced. Although numerous attack models and vulnerability studies have been proposed since the emergence of side-channel attacks, GPU-virtualized heterogeneous computing systems have received little attention. In this paper, we propose an attack in which two virtual machines in different domains establish a covert channel by exploiting the CPU and the integrated GPU, respectively, in such a heterogeneous computing environment. Specifically, we use the Last Level Cache (LLC), which is shared by the Intel integrated GPU (iGPU) and the CPU, as the communication medium. Our experimental results show that the proposed covert channel achieves a bandwidth of approximately 1.9 Kbps with a bit error rate of 17.22%.

This paper presents Two-Stage LKPLO, a novel multi-stage outlier detection framework that overcomes the coexisting limitations of conventional projection-based methods: their reliance on a fixed statistical metric and their assumption of a single data structure. Our framework uniquely synthesizes three key concepts: (1) a generalized loss-based outlyingness measure (PLO) that replaces the fixed metric with flexible, adaptive loss functions like our proposed SVM-like loss; (2) a global kernel PCA stage to linearize non-linear data structures; and (3) a subsequent local clustering stage to handle multi-modal distributions. Comprehensive 5-fold cross-validation experiments on 10 benchmark datasets, with automated hyperparameter optimization, demonstrate that Two-Stage LKPLO achieves state-of-the-art performance. It significantly outperforms strong baselines on datasets with challenging structures where existing methods fail, most notably on multi-cluster data (Optdigits) and complex, high-dimensional data (Arrhythmia). Furthermore, an ablation study empirically confirms that the synergistic combination of both the kernelization and localization stages is indispensable for its superior performance. This work contributes a powerful new tool for a significant class of outlier detection problems and underscores the importance of hybrid, multi-stage architectures.

Integrating continuous valence scalars into high-dimensional semantic spaces presents a fundamental representation mismatch, where low-dimensional intensity signals are often marginalized during fusion. We propose Trajectory-Aware Valence Projection (TAVP), a framework that reformulates valence not as a scalar feature, but as a continuous geodesic trajectory traversing the hyperspherical manifold of a frozen text encoder. By employing Spherical Linear Interpolation (SLERP) between semantic anchors, TAVP ensures that valence intensity is projected into a representation that is both topologically continuous and semantically aligned, avoiding the feature norm collapse associated with standard linear interpolation. Experiments on MER2025 and IEMOCAP demonstrate that this geometric alignment yields substantial performance gains, achieving absolute accuracy improvements of +9.25 pp and +2.72 pp, respectively. Our analysis confirms that explicitly modeling the semantic trajectory provides a robust, interpretable prior for fine-grained emotion recognition.

Weakly Supervised Semantic Segmentation (WSSS) aims to generate pixel-wise predictions using only image-level labels. However, conventional pipelines still suffer from inherent limitations: Class Activation Maps (CAMs) only highlight the most discriminative regions of objects rather than full extent, multi-stage refinement methods demand extensive computational resources, and conventional single-stage models relying on plain or pyramid architectures inherently struggle to expand Effective Receptive Fields (ERFs) and lack explicit top-down semantic guidance. To tackle these challenges, we propose Class-aware Overview-first Look-closely-next (CaLo), a single-stage WSSS framework that fundamentally adapts a top-down architecture for WSSS by explicitly injecting semantic guidance into local feature refinement. Inspired by OverLoCK's overview-focus paradigm, CaLo adopts an SR-MCT-style Vision Transformer as Base-Net to produce mid-level tokens and multi-class tokens, and applies a Deep Stage Decomposition with two lightweight branches: Overview-Net that rapidly yields a coarse context prior, and Focus-Net that progressively refines features. The core of our novel top-down design is Class-aware Gated Aggregator (CAGA) that realizes context-mixing dynamic kernels, where multi-class tokens compute class-to-feature affinities and dynamically gate token mixing, thereby expanding the ERFs and providing explicit top-down guidance to recover object extents without heavy computation. Trained from scratch on PASCAL VOC 2012 and MS COCO 2014 using only image-level labels, CaLo delivers strong accuracy and efficiency, surpassing the prior state-of-the-art single-stage methods. Our results demonstrate that coupling overview-driven global priors with explicit class-aware dynamic token aggregation is a simple yet powerful recipe for scalable single-stage WSSS.

This study presents the development of a Moodle plugin that integrates a web-based handwriting input interface with handwriting-to-text recognition, supporting handwritten input in Japanese, English, and mathematical and chemical formulas. Although computer-based learning systems are widely used in education, they predominantly emphasize multiple-choice or short-answer question types that require conventional keyboard and mouse input. Conversely, structured responses, such as mathematical and chemical expressions, often require specialized input methods, such as LaTeX and Equation Editor, which are challenging for many learners to use. To address these limitations, we developed a Moodle plugin that provides an intuitive and cognitively efficient handwriting-based input interface. The plugin enables learners to produce handwritten responses naturally while preserving raw handwriting data and recognition texts to support transparent scoring. This study bridges the gap between traditional pencil-and-paper assessments and contemporary digital learning environments. It also enhances the authenticity and interactivity of assessments conducted within Moodle. In particular, this paper describes the architecture of the plugin, implementation, and integration challenges encountered within the Moodle system. Usability testing results indicate the plugin can enhance learner engagement and assessment accuracy, representing a step toward personalized and interactive digital education.

Multi-shot trajectory planning remains a critical challenge in autonomous driving and robotics, particularly when handling discontinuous motion states. To address this, we propose YUKUIv2, a novel Y-shaped trajectory planning module. Our approach introduces a key-point driven diffusion mechanism, explicitly predicting discrete transition points to guide the continuous generation of multi-shot trajectories. Evaluated on an CARLA parking dataset, the proposed planner effectively bridges motion state awareness with trajectory diffusion. Results demonstrate that YUKUIv2 maintains high planning accuracy while significantly improving generalization capabilities across complex environmental configurations.

Multi-head attention is a central component of modern transformer-based large language models (LLMs), allowing models to capture diverse patterns via multiple parallel attention heads. Empirical studies show that many heads are redundant, and pruning them can reduce computational and memory costs without substantially affecting output quality. Prior work has also suggested that head pruning may be universal, implying that the same heads can be pruned for different input datasets. Based on this assumption, much of the work on head pruning either considers only a single dataset or employs a fixed pruning strategy across different datasets. In this letter, we systematically evaluate this assumption across multiple datasets and models. Our results reveal that which attention heads to prune is largely dataset-specific rather than universal (i.e., dataset-agnostic), indicating that pruning strategies must be dynamically adapted to different datasets.

Large-scale simulations periodically enter an output phase to save restart images or dump multidimensional arrays for post-processing. In these phases, multi-dimensional arrays are distributed across MPI processes in a scattered manner, so that each process owns many small, non-contiguous regions of the global data. Although each rank holds its data contiguously in memory, this layout maps to disjoint offsets in the shared file. When encoded using MPI file views, the resulting I/O is decomposed into millions of tiny, non-contiguous write requests. Such fine-grained writes incur high per-request overhead, amplify metadata traffic, and ultimately prevent applications from saturating the available bandwidth of modern NVMe devices and parallel file systems.

We present BBView, a lightweight Open MPI plug-in that addresses this inefficiency by buffering MPI file views at the process level. Instead of immediately scattering a rank's contiguous in-memory data across the many disjoint file offsets prescribed by its view, BBView defers that scatter. Each output phase first performs a bulk write, writing the data exactly as it resides in memory, as a single contiguous stream to a node-local file. When the process later closes the file or installs a new view, a background daemon reconstructs the original file view and scatters the buffered data to the parallel file system. This design removes fine-grained writes from the application's critical path while ensuring that the final shared file layout is identical to what the application specified.

BBView is seamlessly compatible with applications built on HDF5 or PnetCDF and functions as an effective I/O acceleration mechanism without requiring source-code changes. An evaluation on 128 nodes demonstrates up to 74.2× higher throughput than UnifyFS on local NVMe SSDs, achieved solely by increasing the effective write granularity through bulk writes.

Sparse Matrix-Matrix Multiplication (SpMM) plays a central role in Graph Neural Networks (GNNs), yet it remains a significant performance bottleneck. Operations on highly sparse adjacency matrices suffer from poor cache locality and irregular memory access, which limits both performance and energy efficiency. Consequently, dataflow-centric optimizations that reduce redundant transfers and improve on-chip reuse are required. To address this issue, we propose IMAX-SpMM, a dataflow-optimized SpMM kernel designed primarily to improve energy efficiency by reducing DMA traffic and increasing on-chip data reuse. We also report speedups as supporting evidence that these optimizations shorten execution time, thereby contributing to the energy-efficiency gains. At the software level, IMAX-SpMM minimizes padding for each dataset. At the hardware level, it exploits the fast compilation capability of a coarse-grained linear array (CGLA) to perform dataset-specific dynamic optimization. These optimizations enable IMAX-SpMM to achieve a data reuse rate of 60.1% for Direct Memory Access (DMA)-loaded data, significantly improving memory transfer efficiency. Additionally, eliminating inter-thread data dependencies among the logical threads executed on each PE increases pipeline density and shortens computation time. Experimental evaluation on Graph Convolutional Network (GCN) demonstrates up to 3.07× speedup over an i9-10940X and 1.64× over an RTX3090. In terms of energy efficiency, the proposed method achieves up to 246.60× improvement compared to an i9-10940X and 25.60× compared to an RTX3090.

Micro-expressions are subtle, involuntary facial movements that reveal a person's underlying emotional state. Their automatic recognition has attracted growing attention in the deep learning community, yet remains challenging due to their subtle appearance and the scarcity of large-scale annotated datasets. This paper proposes FREED-UNet, a class-conditional diffusion model for synthesizing high-fidelity apex-frame images to support micro-expression recognition. The model integrates FreeU-based skip-connection reweighting and conditional self-attention to enhance both structural integrity and fine-grained facial details. To ensure validity, we introduce a multi-stage structural filtering pipeline that combines MediaPipe detection and BiSeNet facial part segmentation. Experiments on the CASME II and newly collected TUMME datasets demonstrate that incorporating FREED-UNet-generated images into training improves the recognition accuracy of SE-DenseNet-cc from 86.74% to 90.76%. These findings highlight the effectiveness of diffusion-based data augmentation in advancing micro-expression recognition. We also report FID/KID to quantify visual fidelity.

In the manufacturing of reconstituted stem silk, inconsistent cutting widths often occur, adversely affecting the final product's physical and chemical properties. Most traditional width measurement methods are time-consuming, labor-intensive, and suffer from limited estimation accuracy. To address these challenges, this study proposes a YOLOv7-BiFormer-NWD-based detection framework incorporating the BiFormer dynamic sparse attention mechanism and the Normalized Wasserstein Distance (NWD) loss function to enhance the detection of small and irregular targets. The model was evaluated on a custom-built dataset. It achieved a mAP@[0.5:0.95] of 75.65%, significantly surpassing the baseline YOLOv7 (67.01%) and achieving the best mAP@[0.5:0.95] among the evaluated models under our protocol (see Table 4). For width estimation, the Suzuki Contour Algorithm (SZA)-based maximum inscribed circle method that fits within the contour to estimate the stem width was employed, demonstrating high consistency with ISO 20193:2019 manual results, with a mean absolute error of 0.011 mm (maximum 0.02 mm) and relative errors < 1.3% in the ISO comparison, with measurement time reduced to 0.15 minutes per sample. The proposed system also showed excellent repeatability and reproducibility (Sr < 0.031 mm, SR < 0.052 mm), validating its effectiveness and suitability for high-precision industrial quality inspection.

In this study, we propose an overlapped wavelet diffusion framework for Low-Light Image Enhancement (LLIE), which incorporates two complementary components to achieve blocking artifact-free and detail-preserving enhancement. Although recent diffusion-based LLIE methods have demonstrated remarkable performance compared with traditional approaches, DiffLL still suffers from blocking artifacts caused by the Haar Wavelet Transform (WT) and blurred edges or over-smoothed textures due to the limitations of its High-Frequency Restoration Module (HFRM). To overcome these issues, we introduce an Overlapped WT (OWT) that incorporates correlations across neighboring regions, thereby structurally preventing blocking artifacts. Furthermore, we integrate a low-frequency-guided High-Frequency Enhance Block (HFEBlock) to strengthen detail recovery, yielding sharper edges and more reliable textures. Extensive experiments on the LOLv1 and LOLv2-real datasets demonstrate that our framework, termed “OWDiff,” consistently outperforms existing LLIE methods both qualitatively and quantitatively, achieving superior visual quality while maintaining computational efficiency. OWDiff effectively addresses the structural limitations of the Haar WT and the HFRM, achieving an average PSNR gain of 0.58 dB, along with a 1.64% relative improvement in SSIM and a 5.9% relative reduction in LPIPS, compared to DiffLL across both the LOLv1 and LOLv2-real datasets.

Emotion recognition in conversation (ERC) aims to identify the affective state of each utterance, such as Joy, Surprise, Happy, and Neutral. In practice, a speaker's affective tendency often persists across multiple consecutive utterances, that is, there exists emotional inertia in dialogue. However, such inertia has not been considered in the dominant ERC paradigm, which may impede further advancements in this field. To address this issue, we propose a chunk-based bi-scale decoder for ERC that decomposes the decoder state into two components, updating them at different timescales to simulate affective continuity throughout the conversation. Specifically, our method predicts an emotion chunk state to model this persistence, upon which multiple utterance context timescale states are generated. The proposed approach outperforms all baseline models on both benchmarks. In the RoBERTa-based setting, it yields F1-score gains of around 0.6 percentage points on the MELD dataset and 0.7 to 1.9 percentage points on the EmoryNLP dataset over the second-best model (+COSMIC), demonstrating its effectiveness.

We propose a format-independent visual obfuscation approach for JPEG images based on Random Intra Prediction Error-propagation (RIPE), including its extended version (XRIPE: eXtended RIPE), which is integrated into the JPEG compression process. Conventional format-compliant encryption schemes introduce randomness into prediction modes or prediction errors of standard codecs such as JPEG, Advanced Video Coding (AVC), and High Efficiency Video Coding (HEVC), which often results in format-dependent visual diffusion patterns originating from codec-specific prediction structures. To overcome this limitation, the proposed RIPE operates in the quantized discrete cosine transform (QDCT) domain of JPEG and predicts DC components using randomly weighted other DC ones from neighboring blocks, thereby inducing controlled error-propagation. This design produces format-independent visual obfuscation while effectively suppressing the degradation in coding efficiency. Experimental results demonstrate that both the proposed RIPE and XRIPE modules successfully reproduce the AVC-like avalanche effects in JPEG images, while incurring only a small bitrate increase of approximately 3.09 % and 2.73 % in terms of Bøntegaard-delta rate.

Extended reality applications over shared wireless networks face heterogeneous bandwidth and latency demands that current per-user approaches cannot adequately address. This paper proposes Enhanced Bandwidth Trading for Extended Reality (E-BITEX), a collaborative bandwidth-sharing framework that extends Bandwidth Trading for Extended Reality (BITEX) with shared bottleneck detection (SBD) and event-driven resource redistribution for multiuser XR deployments. E-BITEX integrates one-way delay measurement to achieve sub-millisecond precision, SBD-based user grouping leveraging skewness and median absolute deviation metrics, and cross-layer coordination spanning application and transport layers. The framework operates entirely through centralized server-side logic, enabling practical deployment across heterogeneous XR platforms. Comprehensive ns-3 evaluation demonstrates consistent Quality of Experience (QoE) improvements of 15.4%-24.0% over BITEX-V2 and 27.1%-36.9% over TCP-BBR across baseline, background traffic, and moving bottleneck scenarios, while maintaining fairness (Jain Index 0.9939).

Multimodal relation extraction (MRE) aims to jointly identify relationships between entities from text and images and plays a crucial role in knowledge graph construction. However, existing methods face two key challenges: (i) inconsistent image quality and semantic misalignment in readily available datasets, particularly from social media, and (ii) the limitations of traditional unidirectional attention mechanisms in capturing fine-grained semantic associations, often leading to information loss and noise. To address these issues, we propose LLM-VGA (Large Language Model-augmented Visual Generation with hierarchical Alignment), a novel framework that integrates an LLM-guided diffusion model with a hierarchical bidirectional cross-modal attention mechanism. Specifically, an adapter injects the semantic reasoning capabilities of LLM into a diffusion model to generate high-quality, text-consistent images, thereby constructing pseudo-aligned multimodal data. Furthermore, a hierarchical bidirectional attention module captures interactions between text and both generated and source images at multiple granularities, alleviating information imbalance and improving alignment. Experimental results on the MNRE and MORE datasets show that LLM-VGA achieves F1 scores of 90.29% and 73.02%, respectively, significantly surpassing state-of-the-art baselines and confirming its effectiveness in enhancing multimodal relation extraction.

Video artificial intelligence (AI) inference, especially object detection from 8K ultra high-definition video, on edge devices is becoming increasingly necessary due to the miniaturisation of 8K cameras and the growing focus on security and privacy protection. To detect objects in real-time in 8K video, both large and small objects need to be detected with limited computational resources on edge devices because the size of objects in distant and near views differs greatly within the same video frame. In the conventional methods for high-definition videos, an object detection model is applied to the regions where objects are expected to exist using density estimation or cluster detection models. The conventional object detection methods, however, have two issues. The first is that the object sizes are not considered when cropping regions from images. The second is that, depending on the scene, the number of cropped regions can be large, making them difficult to process in real time on edge devices. To address these issues, we propose a novel object detection method that can be performed in real-time on edge devices using a density estimation model trained with size-aware ground truths and a technique that reduces computational complexity per frame while maintaining the detection accuracy. We evaluated our proposed method with an 8K video dataset and the Jetson AGX Orin, a typical edge device. In the evaluation with an 8K video dataset, our method achieved of mean average precision (mAP) 46.9%, 1.68 times higher than that of the conventional method (28.0%), while requiring comparable computational complexity for object detection. Our method also achieved a processing speed of nearly 4 fps, which is required for various applications on edge devices, with only about 2% lower accuracy than when rectangle regions were not selected on the Jetson AGX Orin.

We revisit guided-filter-based joint upsampling (GIJU) from both empirical and theoretical perspectives. First, we present a systematic empirical finding that GIJU consistently achieves its highest accuracy with the smallest tested filter radius across the evaluated tasks and scales. To explain this, we provide a theoretical analysis that exposes a previously implicit approximation error in the GIJU pipeline. We formalize this error as a local covariance term between the guide image and the coefficient field, whose magnitude is bounded by local variances, thus clarifying why reducing the effective support improves accuracy. Third, motivated by this insight, we introduce Gaussian-weighted GIJU. By replacing the standard box filter with Gaussian weights, our method further reduces the effective filter support, directly mitigating the identified covariance error. Extensive experiments on 100 images for colorization and L₀ smoothing, as well as on 100 image pairs for depth upsampling, demonstrate that our method consistently outperforms conventional GIJU and a recent filter-based baseline, Detail-preserving Joint Image Upsampling (DPJIU), particularly around edges. Diagnostic analyses, including sensitivity maps and synthetic stepedge experiments, validate the connection between our theory and its practical benefits. Our work provides both a deeper understanding of GIJU and a practical, high-performance, training-free joint upsampler.

End-to-end learning methods can model complex spatial and temporal relationships and achieve good rate-distortion performance. However, many existing methods still rely on single-scale or coarse features for motion estimation. These features cannot capture fine motion details well. As a result, high-frequency information is often lost during motion compensation, which reduces the quality of the reconstructed video. In this work, we propose an end-to-end video compression approach that operates in the deep feature space. The core component is a Multi-scale Progressive Fusion (MPF) module. It extracts features at different scales and fuses them in a progressive manner. This strategy improves motion estimation accuracy while alleviating high-frequency information loss during compression. We further introduce a Global Attention Prediction Enhancement (GAPE) module. By combining high-frequency features with channel attention, this module refines motion compensation and enhances the quality of reconstructed frames. Experimental results show that the proposed method performs better than the baseline models with respect to PSNR and MS-SSIM.

In nighttime road scenarios, object detection tasks often face challenges due to weak illumination and reduced image contrast, resulting in degraded detection performance. To address this issue, this study proposes the Dual-backbone Brightness-aware Salience DETR (DBS-DETR) detection model based on the Salience DETR architecture. The model constructs a dual-branch feature extraction network integrating ConvNeXt and Swin Transformer, and incorporates a brightness-aware component into the saliency supervision module to mitigate the adverse effects of overexposed or low-illumination regions on detection results. Experimental results demonstrate that this method achieves 54.6% mean average precision (mAP) and 80.0% recall, significantly outperforming existing baselines and other comparison models.

This paper used a dataset from the Simple Network Management Protocol (SNMP) to gather Management Information Base (MIB) data for detecting network anomaly attacks. The MIB data, originally in a tree structure, is transformed for deep learning training with Convolutional Neural Networks (CNNs). Various transformation methods are explored, including DFS (Depth-First Search) and BFS (Breadth-First Search), aswell as conversion into two-dimensional or three-dimensional arrays. The tree structure is converted into a two-dimensional array based on the position and a three-dimensional array based on the Object Identifier (OID) values. The goal is to leverage the power of CNNs to identify hidden patterns in the MIB tree structure. Once these hidden features are identified, the model's accuracy can be improved from 92% to 99%. This demonstrates the potential of CNNs in handling the hierarchical structure of MIB OIDs.

We study a simple distributed bit climber algorithm for many-objective optimization of binary problems. This algorithm decomposes the many-objective problem into a minimum number of single-objective problems, specified by the original evaluation functions and one additional scalarizing function that computes the solution hypervolume. Random bit climbers optimize separately the single-objective functions until they reach a local optimum and restart their search from a population of non-dominated solutions collected from the solutions generated by all climbers and bounded according to a dense set of weights to enhance diversity. In this work, we study the climbing behavior of the simple distributed bit climber evaluating the method on subclasses of epistatic problems using MNK-landscapes and compare with other well established multi-objective evolutionary algorithms, showing that the proposed climber is superior.

Robotic collision detection accelerators suffer from Orientation Aliasing and Granularity Mismatch with multi-sphere models. We propose S-COPU, a sphere-level speculative scheduling architecture using a “generate-then-predict” flow. S-COPU aligns prediction granularity with geometric representation, reducing geometric queries by 25% and end-to-end latency by 14.9% relative toCOPUin high-clutter environments without increasing memory requirements.

Cross-category anomaly detection for 3D point clouds aims to determine whether an unseen object belongs to a target category using only a few normal examples. Most existing methods rely on category-specific training, which limits their flexibility in few-shot scenarios. In this paper, we propose DMP-3DAD, a training-free framework for cross-category 3D anomaly detection based on multi-view realistic depth map projection. Specifically, by converting point clouds into a fixed set of realistic depth images, our method leverages a frozen CLIP visual encoder to extract multi-view representations and performs anomaly detection via weighted feature similarity, which does not require any fine-tuning or category-dependent adaptation. Extensive experiments on the Shape Net-Part dataset demonstrate that DMP-3DAD achieves state-of-the-art performance under few-shot setting. The results show that the proposed approach provides a simple yet effective solution for practical cross-category 3D anomaly detection.

Single image dehazing is a key but really tough task in computer vision. Especially in dense and uneven haze scenes, existing methods usually can't recover complete structures, clear details, and normal colors. To solve these problems, we propose a two-stage patch-conditioned diffusion model (TS-PCDM). It first builds the global image structure for a stable basic result, then refines local textures and edges under this guidance. We also use overlapping patch inference and EMA to reduce artifacts and make the model more stable. Experiments on three real-world haze datasets show our model clearly outperforms current top methods in PSNR and SSIM. It removes haze effectively, restores rich details without color shifts, and proves to be much better than existing approaches.

The Agile-Chip platform enables the rapid and low-cost development of ASICs by implementing the CPU, SRAM, I/O, and Gate Array in the lower metal layers of an LSI, and forming the upper metal layers using maskless lithography. Currently, a prototype called Agile-1L is available, which uses only the fifth metal layer for routing in a 0.18 μm CMOS process with 1P6M (1-Poly, 6-Metal). By utilizing Minimal Fab, the routing of the fifth metal layer can be completed in approximately 30 minutes. A method using both the fifth and sixth metal layers (Agile-2L) is also technically feasible. This approach competes with FPGAs—particularly one-time programmable Anti-Fuse FPGAs—in terms of enabling rapid chip prototyping. This study compares the performance and power consumption of Agile-1L and Agile-2L with those of conventional ASICs and three types of FPGAs: SRAM-based, flash-based, and Anti-Fuse-based. Although the currently available Agile-1L, which uses only a single routing layer, performs significantly worse than conventional ASICs in terms of performance and power, it is comparable to various types of FPGAs. Agile-2L, on the other hand, demonstrates characteristics closer to those of conventional ASICs and shows sufficient competitiveness when compared to existing FPGAs.

Past methods for generating personas have primarily relied on manual assumptions, which have made the automated identification of system-level pain points challenging and have often limited the scope to actual inquirers who provided feedback. To address these limitations, we introduce a framework that combines user logs and inquiry data to generate personas in a more automated manner by integrating behavioral patterns and pain points. The framework consists of three phases: clustering users with k-means, analyzing those clusters with PrefixSpan, random forests, and BiLSTM models to identify pain points for both actual and potential inquirers (that is, actual inquirers—users who submitted feedback—and potential inquirers—users who did not reach out to developers regarding issues), and developing personas through large language models. A case study conducted on Japanese B2B SaaS software, involving 574 users and 3.5 million logs, showed that the generated personas were typically easy to conceptualize, and the identified pain points were compelling and actionable, underscoring their significance in guiding software enhancements.

Structure from Motion (SfM), which reconstructs 3D data using a monocular camera, is known to be susceptible to scale drift. To address this issue, previous approaches corrected scale using the camera height estimated from a ground plane fitted to road surface textures, or by utilizing the prior distribution of vehicle sizes detected through object detection. In this study, we propose a novel scale correction method based on multi-class object detection, without relying on prior approaches. Our approach estimates scale variations by tracking changes in the sizes of multiple object classes and aligning them to a dominant class. This method is ideally suited for scenes where the dominant object class appears sparsely across images. Through quantitative evaluations using in-vehicle video dataset, we demonstrate the effectiveness of our method in estimating scale drift and mitigating accumulated errors.

Behavioral cloning is a widely used approach for training text-based game agents. However, directly imitating low-level movement commands (e.g., move east) introduces ambiguous supervision, since the same movement command can lead to entirely different targets. This ambiguity lowers training efficiency and limits the agent's performance. To address this issue, we reformulate sequences of movement commands as high-level item navigation commands (e.g., navigate to red onion), which explicitly specify the intended goal. A parameter-free navigator module is then introduced to execute these commands by managing pathfinding and item tracking, allowing the agent to focus on higher-level decision-making. Experiments on the First Text World Problems (FTWP) dataset demonstrate that our method achieves a task completion rate of 96.8%, establishing a new state of the art and surpassing the previous best result of 91.2%.

This paper evaluates eBPF overheads in ultra-low-latency storage systems. The results show that verification and runtime costs are generally small, but become significant with near-DRAM devices and dynamic, multi-tenant environments. These findings confirm the feasibility of eBPF for programmable storage while emphasizing its implications in storage-specific and ultra-low-latency contexts.

Dataset distillation (DD) aims to generate a compact yet informative dataset that achieves performance comparable to the original dataset, thereby reducing demands on storage and computational resources. Although diffusion models have made significant progress in dataset distillation, the generated surrogate datasets often contain samples with label inconsistencies or insufficient structural detail, leading to suboptimal downstream performance. To address these issues, we propose a detector-guided dataset distillation framework that explicitly leverages a pre-trained detector to identify and refine anomalous synthetic samples, thereby ensuring label consistency and improving image quality. Specifically, a detector model trained on the original dataset is employed to identify anomalous images exhibiting label mismatches or low classification confidence. For each defective image, multiple candidates are generated using a pre-trained diffusion model conditioned on the corresponding image prototype and label. The optimal candidate is then selected by jointly considering the detector's confidence score and dissimilarity to existing qualified synthetic samples, thereby ensuring both label accuracy and intra-class diversity. Experimental results demonstrate that our method can synthesize high-quality representative images with richer details, achieving state-of-the-art performance on the validation set.

A dynamic 3D mesh is a key component in Virtual Reality applications. However, this type of content demands a significant processing resource for real-time rendering. To reduce processing requirements while preserving the user experience, adjusting the level of detail of 3D meshes based on viewing distance has been proposed. In this paper, we conduct an extensive subjective quality evaluation to investigate the effects of the level of detail and viewing distance on user perception of dynamic 3D meshes in a VR environment. Our evaluation results in a subjective dataset containing user ratings of 320 test stimuli generated from eight dynamic 3D meshes. Result analysis shows that it is possible to remove half of a mesh's faces without causing noticeable degradation in user Quality of Experience (QoE). An evaluation of popular objective quality metrics reveals that both 2D-based and 3D-based metrics have low correlation with subjective scores. Based on the subjective dataset, we develop a novel QoE prediction model that can accurately predict the MOS of a dynamic 3D mesh at a given level of detail and viewing distance. In addition, a QoE-aware resource allocation framework is proposed and evaluated under different resource constraints, showing significant improvement in the total QoE compared to conventional methods.

The generation of trees with a specified tree edit distance has significant applications across various fields, including computational biology, structured data analysis, and image processing. Recently, generative networks have been increasingly employed to synthesize new data that closely resembles the original datasets. However, the appropriate size and depth of generative networks required to generate data with a specified tree edit distance remain unclear. In this paper, we theoretically establish the existence and construction of generative networks capable of producing trees similar to a given tree with respect to the tree edit distance. Specifically, for a given rooted, ordered, and vertex-labeled tree T of size n + 1 with labels from an alphabet Σ, and a non-negative integer d, we prove that any rooted, ordered, and vertex-labeled tree over Σ with tree edit distance at most d from T can be generated using an appropriate random input sequence to a ReLU-based generative network of size O(n³) and constant depth. The proposed networks were implemented and evaluated for generating trees with up to 21 nodes. Due to their deterministic architecture, the networks successfully generated all valid trees within the specified tree edit distance. In contrast, state-of-the-art graph generative models GraphRNN and GraphGDP, which rely on non-deterministic mechanisms, produced significantly fewer valid trees, achieving validation rates of only up to 35% and 48%, respectively. These findings provide a theoretical foundation towards construction of compact generative models and open new directions for exact and valid tree-structured data generation. An implementation of the proposed networks is available at https://github.com/MGANN-KU/TreeGen_ReLUNetworks.

Background: A workforce shortage in information technology (IT) has become a pressing issue. To increase the number of women IT professionals, it is crucial to expand the available career options for women students. A previous study demonstrated that gender stereotype is a threat to the performance of women in mathematics tests. This phenomenon is referred to as stereotype threat. Aim: We focused on gender stereotype threat to support women IT students in achieving their best performance. Method: Through a subjective experiment, we examined the relationship between a stereotypical cue and performance in computer programming. Additionally, we investigated the impact of a stereotype threat-removal strategy on programming performance. Result: When analyzing the time required to develop a program, neither the cue nor the strategy showed any significant effect. This result might be influenced by the intrinsic nature of computer programming and the effects of gamification.

With the expanding constellation scale and growing demands for Quality of Service (QoS) assurance, satellite networks are evolving toward multi-layer architectures. Traditional single-layer designs fail to ensure link stability and service continuity for emerging applications. This paper proposes a cross-layer communication design for large-scale Low Earth Orbit (LEO) satellite networks oriented toward QoS optimization. Specifically, we establish a dynamic topology model for large-scale LEO satellite networks and define the multitask scenario. The routing problem is formulated as an Integer Linear Program (ILP) to maximum the average link utilization ratio under unsplittable flow constraints. We propose the Deep Deterministic Policy Gradient-Based QoS Routing Algorithm (DDPG-QoSRA) to ensure both path reliability and the transmission performance of critical services. Finally, the simulations show the effectiveness and superiority of the proposed algorithm.

Existing parking space line detection systems face persistent challenges under complex environmental conditions, particularly lowlight conditions, occlusions, and textured background interference, which lead to compromised accuracy, excessive computational overhead and high latency. To address these limitations, this study proposes PSL-YOLO11, an enhanced lightweight detection framework based on the YOLO11 architecture. The innovation lies in a hybrid dynamic attention mechanism that synergistically combines the Convolutional Block Attention Module (CBAM) and the Global Attention Mechanism (GAM) for superior environmental adaptability. The framework implements multi-scale feature fusion with channel expansion for discriminative feature learning, and adopts GhostNet-based feature compression to reduce computational complexity. Additionally, a key point-driven geometric abstraction algorithm is designed to convert parking lines into quantifiable topological representations. Experimental results on the HNUST-IR dataset demonstrate that PSL-YOLO11 achieves 93.8% mAP and 96.15 FPS, with 22% fewer parameters than YOLO11 while maintaining real-time performance.

This paper proposes a method for utilizing large language model (LLM) inference results in the state space of deep reinforcement learning (DRL)-based GUI black-box testing for web applications. Previous research on mobile applications has shown that using HTML element interaction information in training improves the exploration performance of DRL agents. However, in web applications, obtaining interaction information can be difficult depending on the browser. Therefore, this study proposes a method that uses an LLM to infer clickability of elements, that is, one example of interaction information, from HTML and utilizes these results as part of the DRL state. Through experiments on a simple web application, we confirmed that using LLM-inferred values as state information improved learning efficiency and accuracy. Similar results were also demonstrated in experiments targeting a complex web application. The findings of this research enable browser-independent automation of GUI black-box testing and present a novel approach to utilizing LLMs for obtaining DRL states.

Multimodal Named Entity Recognition (MNER) aims to improve entity prediction by fusing textual and visual information. Most current MNER methods face two main issues: 1) they do not adequately appreciate the necessity of incorporating external knowledge into the model and the presence of knowledge redundancy; 2) during cross-modal fusion, they fail to achieve text-guided multimodal integration, resulting in an excessive introduction of image noise. To address these issues, we propose a novel framework PKTF, which primarily consists of two stages: the prior assisted knowledge generation stage and the multimodal named entity recognition stage. In the first stage, we use Intern VL2-8B to generate prior assisted knowledge, aiming to provide additional contextual information for the original text. In the second stage, we design a Text-Max-Directed Fusion Module (TMDF). Specifically, we use gates to modulate the max pooling attention scores guided by the text, in order to obtain text-guided saliency attention scores. We can use this score to maximize the extraction of information favorable to MNER from the image features, while ensuring the dominance of the text. Experimental results show that our method is competitive compared to existing models, and it achieves F1-scores of 75.43% on the Twitter-2015 dataset and 88.74% on the Twitter-2017 dataset, respectively.