Software design documentation is an important artifact for software developers to implement their products. However, keeping documentation up-to-date is challenging; software developers have to spend time frequently modifying, reviewing, and merging documentation changes written in Microsoft Word. To address the up-to-dateness issue, a software team in a company has introduced AsciiDoc, which is a lightweight markup language, for design documentation so that they can review and merge changes using the same tools as source code. This study investigates the impact of the introduction of AsciiDoc by conducting a developer survey and a repository mining. The survey results show that developers perceived that AsciiDoc enabled them to update their documents efficiently. The result of the repository mining suggests that their software development velocity has improved. Based on the observations, we believe that the introduction of AsciiDoc has improved the efficiency of documentation workflows.

Efficient management of spatial data is essential for applications such as navigation, urban planning, and location-based services. With the rapid growth of spatial data volume and complexity, traditional storage solutions like SSDs are increasingly inadequate in meeting the performance demands of such workloads. Persistent Memory (PMEM) emerges as a promising alternative, offering low latency, high throughput, and byte-addressable access, while retaining data even after power loss. These characteristics make PMEM particularly attractive for spatial indexing structures such as the R-Tree, which are central to spatial data management. However, existing R-Tree implementations on PMEM-such as FBR using mutex-based synchronization and PMR using PMwCAS-suffer from scalability bottlenecks in multi-threaded environments. Mutex-based synchronization introduces significant contention, while PMwCAS provides only partial lock-freedom and incurs non-negligible overhead. To address these limitations, we propose TSR (Test and Set based R-Tree), an effective synchronization method based on Test-and-Set (TAS) instructions. By applying TAS at the node level, TSR minimizes contention and enables higher concurrency during insert operations. We have conducted extensive experiments using real PMEM hardware and demonstrate that TSR outperforms both FBR and PMwCAS in terms of insertion throughput and scalability. At eight threads, TSR improves throughput by up to 66% compared to FBR and 44% over PMwCAS. Our contributions are threefold: (1) we demonstrate the performance advantage of PMEM over SSDs in spatial indexing workloads; (2) we propose a TAS-based synchronization method that achieves better thread scalability; and (3) we provide a comprehensive performance analysis of R-Tree implementations under different synchronization strategies on PMEM.

In recent years, soccer analysis based on players’ movements and decision-making has been actively studied. One type of soccer analysis, called Action Spotting, is a task that identifies the occurrence time and type of actions during a match. For this task, many methods using visual features extracted from broadcast video have been proposed. However, due to factors such as replays and camera angles, some actions are not captured in the video. Actions that occur during a match but are not shown in the video are defined as invisible actions, and the lower detection accuracy for these compared to visible actions remains a challenge. Existing studies have mainly attempted to improve the detection accuracy of the invisible actions by using audio, but they have not focused on the content of the commentary. To address this, we propose ASPERA (Action SPotting thrEe-modal Recognition Architecture), a multimodal soccer action recognition method that incorporates video, audio, and commentary, along with three extended models based on ASPERA: (i) ASPERA_srnd, which incorporates commentary text within a ±20-second window around each specified second; (ii) ASPERA_cln, which removes background information unrelated to the flow of the match from ASPERA_srnd; and (iii) ASPERA_MC, which applies a Markov head containing prior knowledge about the sequence of soccer actions to ASPERA_cln. As a result, these proposed models improved the accuracy of identifying the timing of actions. In particular, ASPERA_MC, which incorporates the Markov head, achieved the highest accuracy in recognizing actions that are invisible in the video.

In-context learning is a remarkable property of transformers and has recently received a lot of attention. A key component of transformers is the attention mechanism, which uses an attention matrix to encode relationships between words in a sentence and assign weights to them. This mechanism effectively captures language representations. However, the suitability of naive self-attention is suitable for in-context learning in general tasks is questionable, since the computation implemented by self-attention is somewhat restrictive in terms of matrix multiplication. In fact, as seen in several works, we may need appropriate input form designs when considering heuristic implementations of computational algorithms. In this paper, we present an extension of linear self-attention by introducing a bias matrix in addition to a weight matrix for the input. Despite its simplicity, the extended linear self-attention can output any constant matrix, input matrix and multiplication of two or three matrices in the input. Note that the second property implies that it can be a skip connection. Therefore, flexible matrix manipulations can be implemented by connecting the extended linear self-attention components. As an example of implementation using the extended linear self-attention, we show a heuristic construction of a batch-type gradient descent for ridge regression with reasonable input forms.

A knowledge graph (KG) is structured information about the real world, which is formed by a collection of triplets of a head entity, a relation, and a tail entity. Although KGs can be automatically constructed from text or manually curated, they often suffer from misinformation, incompleteness, and noise. This study aims to improve the quality of a noisy KG by implementing a two-step process: (1) the detection of noisy (erroneous) triplets and (2) the refinement of these triplets by correcting the wrong entities. Leveraging recent advances in large language models (LLMs), we propose two methods for noise detection and refinement: LLM_sim and LLM_rule. LLM_sim uses LLMs to assess the semantic plausibility of triples by measuring their similarity to the existing triplets. LLM_rule automatically induces rules from an existing KG. These rules provide semantic constraints of head and tail entities for a given relation, and are used to identify noisy triplets and correct them. Results of experiments employing artificially constructed and authentic noisy KGs show that both LLM_sim and LLM_rule perform well for both noise detection and refinement. Furthermore, our methods are applied to the downstream task of Knowledge Graph Completion (KGC). The results demonstrate that the refinement by LLM_sim and LLM_rule yields substantial improvements in KGC performance.

Medical interviews are a core component of medical education, forming an essential part of the clinical skill that students must acquire. Simulated patients (SPs), who replicate the behavior of real patients, play a crucial role in the learning and examination of medical interviews. However, arranging for competent SPs who consistently perform according to detailed instructions and scenarios remains both labor-intensive and costly. To address this issue, many studies have proposed virtual simulated patients (VSPs) utilizing large language models (LLMs). Conventional VSPs, however, often generate undesirable excessive responses. Moreover, these methods typically rely on cloud-based LLMs, which raises significant concerns about data leakage and operational costs. To overcome these challenges, this study proposes an approach to develop VSPs equipped with mechanisms to suppress excessive responses by utilizing an open-source LLM. Our core mechanisms are self-refinement, based on an artificial intelligence agent approach, and question category-aware answering, which aligns the VSP’s responses with the appropriate granularity required for the category of the given questions. Through experiments with actual data, we demonstrate that the proposed method significantly reduces excessive responses.

Large-scale pre-trained vision-language models (VLMs) such as CLIP have shown remarkable performance on various downstream tasks. However, such a model often learns not only the information that is truly useful for classification, but also group attributes that are spuriously correlated with classes, leading to misclassification of an image into a wrong class with the same group attributes. The goal of this paper is to develop a method for tuning pre-trained visual language models to be robust to group attributes. Unlike most of the existing methods, our method satisfies four desirable properties: (i) knowledge-free: does not use any information of group attributes for training, (ii) linear: a lightweight method that trains only a single linear projection, (iii) calibration-based: does not change the original classifier at all, and (iv) zero-shot: does not utilize any training images. The negative effects of the group attributes can be canceled by projecting the classification space to the orthogonal complement of the subspace spanned by the group attributes. To achieve this, we propose Spurious Subspace Mining (SSM) to discover the subspace from a random set of text embeddings without any supervision. Experimental results on four standard benchmark datasets, Waterbirds, CelebA, PACS, and VLCS, show that the proposed method outperforms various existing methods and improves zero-shot baseline by 43.4% in worst-group accuracy.

High-resolution 3D point clouds are highly effective for detecting subtle structural anomalies in industrial inspection. However, their dense and irregular nature imposes significant challenges, including high computational cost, sensitivity to spatial misalignment, and difficulty in capturing localized structural differences. This paper introduces a registration-based anomaly detection framework that combines multi-prototype alignment with cluster-wise discrepancy analysis to enable precise 3D anomaly localization. Specifically, each test sample is first registered to multiple normal prototypes to enable direct structural comparison. To evaluate anomalies at a local level, clustering is performed over the point cloud, and similarity is computed between features from the test sample and the prototypes within each cluster. Rather than selecting cluster centroids randomly, a keypoint-guided strategy is employed, where geometrically informative points are chosen as centroids. This ensures that clusters are centered on feature-rich regions, enabling more meaningful and stable distance-based comparisons. Extensive experiments on the Real3D-AD benchmark demonstrate that the proposed method achieves state-of-the-art performance in both object-level and point-level anomaly detection, even using only raw features.

The Joint Video Experts Group (JVET) conducts exploration research on technologies for the next-generation video coding standard beyond VVC, using the Enhanced Compression Model (ECM) as reference software. ECM integrates both extended and newly developed coding tools. These include BV-based prediction tools like IntraTMP, as well as mode-derivation tools such as TIMD and TMRL, which utilize intra prediction mode lists. In the current ECM, intra prediction mode lists are constructed using modes of neighboring blocks, and IntraTMP is conventionally mapped to the Planar mode, limiting diversity. This paper proposes a unified mapping framework that derives the IPM of BV-based prediction tools—IntraTMP as well as IBC—by referencing the intra mode of the block indicated by the BV. In addition, a multiple-position mapping strategy is introduced for IntraTMP to exploit several reference positions within the reference block. Experimental results show that the proposed method achieves a 0.03% BD-rate reduction in the luma component, demonstrating a slight coding efficiency through enhanced intra prediction mode diversity.

In practical applications, there are a large number of small targets in infrared imaging data, and the existing detection model parameters are too large, making it difficult to achieve efficient detection of infrared imaging targets based on edge devices. In response to this issue, this paper proposes an infrared image object detection model based on Separable Wide Attention and Hierarchical Feature Fusion Mechanism. The proposed model embeds a Separable Wide Attention (SWA) module into a Small Detection Head (SDH) network to improve the recognition accuracy of small targets. In addition, the Hierarchical Feature Fusion Mechanism (HAF) module is integrated into the Neck. This design divides the input feature maps into multiple feature slices along the channel dimension and feeds them into different attention heads, thereby reducing the number of model parameters and ultimately enhancing the real-time performance of detection. The experimental results based on the extended M3FD dataset show that compared with existing mainstream models, the proposed model achieves an AP_small of 48.0, which is 10.1% higher than PPYOLOE+- m, and an FPS of 54.6, thus achieving efficient infrared target detection.

Semantic segmentation of outdoor point clouds plays a vital role in applications such as autonomous driving and urban scene understanding. However, the uneven density of point clouds manifests as sparsity in distant regions and incomplete geometric information of objects, while the proportions of different object classes remain significantly imbalanced. These challenges generally result in poor segmentation performance for distant and fine-grained objects. In addition, many existing methods are not well-suited for large-scale point clouds due to costly sampling strategies and computationally intensive pre- and post-processing steps. To address these issues, we propose MS-SADNet, an efficient and lightweight neural network architecture. To reduce the computational burden of large-scale point cloud processing, the network adopts random sampling to lower computational overhead, thereby avoiding the efficiency bottlenecks introduced by complex sampling strategies. A multi-scale local feature aggregation module is further incorporated to enlarge the receptive field of local features, compensating for information loss caused by uneven density and enhancing the representation of fine-grained objects. Moreover, we design an attention-based decoder that captures long-range dependencies during feature upsampling, enabling dynamic fusion of local details and global semantic information, and strengthening the features of rare classes and geometrically incomplete objects through inter-object relationships. In addition, learnable parameters are introduced into the skip connections to overcome the gradient propagation limitations of traditional skip connections for long-tail classes, thereby improving multi-level feature fusion. Experiments on the Toronto3D, SemanticKITTI, and Semantic3D datasets demonstrate that the proposed method achieves promising segmentation performance, with notable improvements in the accuracy of distant sparse regions and fine-grained objects.

Chinese spelling correction is a task that involves detecting and correcting errors in Chinese text. Most character errors arise from visual or phonetic similarities, which significantly increases the difficulty of correction. In recent years, many studies have employed BERT-based language models for error correction tasks. However, these models typically adopt an encoder-decoder architecture, where the correction process relies on analyzing the context of the input sentence to predict the output. This often biases the model toward contextual information, making it susceptible to being misled by erroneous context. As a result, the model may overcorrect originally correct sentences into incorrect ones. This issue is critical for error correction tasks, as such mistakes degrade model performance and generate misleading performance metrics. This paper proposes a collaborative error correction framework that integrates a detector. The framework uses the detector to identify error locations, thereby assisting the decoder in enhancing error recognition. Additionally, a post-output optimization strategy is introduced to mitigate overcorrection errors. Comprehensive experimental results on benchmark datasets demonstrate that the proposed method achieves promising performance in Chinese text error correction tasks, significantly improving model effectiveness.

Parallel end-to-end TTS models have made notable progress in speech naturalness but still suffer from limitations in fluency and computational efficiency. This letter presents GSViTS, a parallel TTS framework that achieves robust alignment and enchanced prosody by integrating a differentiable SoftDTW-based alignment module with uncertainty-aware duration prediction. In addition, a Dual-Path Gated Linear Attention (DP-GLA) mechanism is introduced to support efficient long-sequence modeling with reduced computational overhead. Experiments on LJSpeech and VCTK demonstrate 20% MOS improvement and 7% lower computational cost over baselines.

Multi-view soccer foul recognition aims to classify foul actions and their severity by leveraging multi-view video data. However, existing methods struggle with effective multi-view feature fusion and spatial-temporal modeling, limiting their ability to accurately interpret critical actions. To address these limitations, we propose the Spatial-Temporal Query Network (STQNet), a novel method for multi-view soccer foul recognition that improves both foul action classification and severity estimation. First, we adapt the existing Vision Transformer encoder for extracting multi-view spatial-temporal embeddings. Then, we introduce a dual-branch spatial-temporal query decoder that utilizes learnable action and severity queries to search for foul cues from the corresponding visual embeddings. Finally, dual classification heads are employed to predict foul action and severity. Experimental results on the SoccerNet-MVFouls dataset demonstrate that STQNet outperforms existing methods with superior performance.