Image-tabular multimodal learning is receiving increasing attention in visual classification tasks. However, heterogeneous tabular data suffers from semantic space incompatibility, and existing methods often struggle to achieve effective alignment for image-tabular cross-modal learning. To address these challenges, we propose a multi-metric optimal transport-based visual-tabular feature alignment network for multimodal classification. Firstly, to tackle the semantic incompatibility between categorical and continuous features in tabular data, a Transformer-based tabular encoder is designed; this module employs a self-attention mechanism to capture complex dependencies among features, providing accurate structured semantic features for cross-modal alignment. Subsequently, a multi-metric optimal transport alignment module is designed, which establishes a structured mapping relationship between image channels and tabular attributes through optimal transport theory, thereby achieving effective alignment for image-tabular cross-modal learning. Finally, a contrastive learning alignment module is designed, which leverages the global information from the feature extraction modules to enhance the consistency of cross-modal global features and learn modality-invariant feature representations. Experimental results demonstrate that our proposed method achieves performance improvements on both medical image and natural image datasets. The method achieves an Accuracy of 0.9884 on the Multi-brain dataset, and an Accuracy of 0.9726 and 0.9296 on the SUN and DVM datasets, respectively.

Abstract: Small object detection in remote sensing images has long faced three core challenges due to characteristics such as large target scale (up to 2–3 orders of magnitude), high background complexity, and dense target distribution: the loss of high-frequency details in multi-scale feature fusion, feature confusion and missed detection in dense scenes, and long-tail category and regional biases induced by imbalanced data distribution. To address these issues, this paper proposes a small object detection framework tailored for multi-scale remote sensing images, with synergistic innovations across three dimensions. First, a Cross-Scale Frequency-Guided Feature Reconstruction Module (CFFRM) is introduced, which decouples multi-scale features into high- and low-frequency components via twodimensional Discrete Cosine Transform (DCT) and incorporates a deformable attention mechanism to achieve cross-scale high-frequency interaction, thereby dynamically reconstructing detail-enhanced feature representations and effectively mitigating high-frequency attenuation in deep networks. Second, a Dual-Pathway Semantic-Detail Synergistic Enhancement Network (DPSDNet) is designed. Inspired by the ventral/dorsal pathway mechanisms of biological vision, a Transformer encoder is employed as the semantic pathway to capture global context, while a CNN encoder serves as the detail pathway to preserve spatial edge and texture information; bidirectional complementarity is achieved through cross-pathway cross-attention interaction. Third, an Adaptive Dense Region Perception and Balanced Learning Framework (ADRP-BLF) is constructed, comprising a density heatmap-guided adaptive cropping mechanism, a blur-background suppression enhancement module, and a classadaptive focal loss function, which systematically addresses spatial distribution non-uniformity and long-tail bias. Extensive experiments on four benchmark datasets(MS COCO, PASCAL VOC, DIOR, and DOTA) demonstrate that the proposed method outperforms current state-of-the-art methods in both general and remote sensing scenarios: on the COCO dataset, mAP@50 reaches 67.1%, mAP@50:95 reaches 44.8%, and small object AP (mAP_S) reaches 27.6%; on the DIOR and DOTA datasets, mAP@50 reaches 74.2% and 76.1%, respectively, significantly outperforming existing mainstream detection methods. Visual analysis further validates the superior detection capability of the proposed method in challenging scenarios such as dense small objects, multi-scale scenes, and long-tail categories.

“Poetry in pictures and pictures in poetry” elucidates a naturally direct correlation between the classical Chinese poetry and vision, making the image and poetry retrieval crucial. Most existing works employ the photograph as the query, but obtaining a suitable photograph brings additional difficulties. In contrast, the free-hand sketch has served as a more convenient tool to depict human perception since ancient times. In this paper, we introduce a new task of Sketch-Based Poetry Retrieval. The task is challenging due to the following factors: (i) the significant differences between sketches and images, as well as between poetry and modern Chinese; (ii) the high time cost of collecting the parallel cross-modal data for the traditional supervised learning. To address these challenges, we construct a sketch-and-poetry pre-training model based on unsupervised vision-and-language learning named SKP-CLIP. Specifically, we utilize a multi-modal knowledge graph for poetry to bridge the semantic gap and then learn the modal alignment through sketch and its caption as well as poetry and its corresponding image entities instead of sketch-poetry pairs. Furthermore, we semi-automatically assemble a dataset for evaluation. Experiments confirm the effectiveness of our method and establish benchmarks for this new task.

Attributed graph clustering has attracted increasing attention for its ability to reveal latent community structures. It aims to learn a rich node representation, whose quality is closely related to the clustering performance. Recent contrastive learning approaches have shown promise by aligning similar nodes and separating dissimilar ones. However, they generally suffer from two limitations: (1) They fail to mine hard samples from a clustering perspec tive by viewing all samples equally. (2) They focus solely on contrast between local node-level representations, leading the model to concentrate on single-scale information while ignoring global semantics of the graph. To ad dress these issues, we propose Multi-scale Contrastive learning with Cluster-aware Hard sample mining (MSCH) for attributed graph clustering. MSCH is beneficial for learning discriminative node representations by mining cluster-aware hard samples, and capturing both local and global information through multi-scale contrastive learning, thereby better capturing complex topological structures. Specifically, MSCH leverages high-confidence cluster pseudo-labels to guide the selection of positive and negative samples, aligning node representation learn ing with cluster structure. Hard samples are then identified based on cluster difficulty through a weight regulation function. Meanwhile, a multi-scale contrastive learning is proposed to combine graph-to-node and node-to-node level contrasts, which learns complementary information between local node and global graph representations. The proposed MSCH method has been evaluated through extensive experiments on six datasets. The results demonstrate the superiority of MSCH for clustering nodes that lie on ambiguous boundaries, with an average ACC gain of 1.2% over the best existing method.

As a core task in intelligent education scenarios, Public Speaking Anxiety Detection (PSAD) aims to effectively predict learners' speaking anxiety states using multimodal data. However, in practical educational applications, unavoidable noise interference or data collection constraints often lead to uncertain modality absence, which significantly degrades model performance. Against this backdrop, incomplete multimodal learning faces the dual challenges of insufficient feature learning and difficulties in distribution adaptation. Current approaches often overemphasize modality-invariant features, which not only restricts the exploration of modality-specific characteristics but also overlooks the dynamic changes in modality contributions across different samples. Moreover, dynamic modality missingness may further cause significant source-target distribution shifts. Mainstream adaptation methods fail to adequately address the uncertainty inherent in missing data. To address these issues, we propose a general framework named Counterfactual-enhanced Knowledge distillation and Uncertainty-aware contrastive Adaptation (CKUA), which aims to strengthen the robustness of incomplete multimodal learning in PSAD. Specifically, we construct a counterfactual knowledge distillation framework that integrates modality contribution estimation into the distillation process, thereby enhancing the efficient transfer of high-contribution semantics. Furthermore, uncertainty-aware contrastive adaptation quantifies uncertainty in multimodal representations and reinforces consistency constraints via uncertainty-based noise injection, thereby mitigating source-target distribution shifts. Extensive experiments on multimodal benchmark datasets demonstrate that CKUA significantly outperforms state-of-the-art methods1.

Fake news detection is crucial for safeguarding information security and maintaining social stability. Existing studies use graph neural networks to capture news propagation features to boost fake news detection. However, constrained by the pairwise nature of conventional graphs, they decompose propagation into a set of individual-level behaviors. In practice, news diffusion is a collective propagation process rather than a simple aggregation of individual behaviors. Real and fake news exhibit distinct collective propagation behavior patterns, which can provide valuable cues for fake news detection. To exploit such collective-level characteristics, we propose LeHT, an LLM-enhanced hypergraph transformer framework for fake news detection via collective-level propagation analysis. LeHT leverages a large language model as a context provider to generate context-enhanced semantic representations, alleviating semantic sparsity in short news posts and enabling better understanding of semantics that may trigger collective reactions. Meanwhile, we construct propagation hypergraphs to explicitly model collective propagation behaviors, where each news item is represented as a hyperedge connecting all participating users. Based on the propagation hypergraph and the context-enhanced semantic representations, we design a hypergraph transformer network with a three-level attention mechanism to capture complex higher-order correlations in collective propagation behaviors and learn discriminative collective propagation behavior pattern representations for fake news detection. Experiments on three public datasets - Twitter (1152 news items), PHEME (1705 news items), and DRWeibo (6037 news items) - demonstrate that LeHT achieves ACC/F1 scores of 0.923/0.921, 0.926/0.924, and 0.916/0.913, respectively.

In practical applications of deep learning-based Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems, new target categories emerge continuously. This requires the systems to learn incrementally—acquiring new knowledge while retaining previously learned information. To mitigate catastrophic forgetting in Class-Incremental Learning (CIL), this paper proposes a CIL method for SAR ATR named Multi-center Prototype Feature Distribution Reconstruction (MPFR). It has two core components. First, a Multi-scale Hybrid Attention feature extractor is designed. Trained via a feature space optimization strategy, it fuses and extracts discriminative features from both SAR amplitude images and Attribute Scattering Center data, while preserving feature space capacity for new classes. Second, each class is represented by multiple prototypes to capture complex feature distributions. Old class knowledge is retained by modeling their feature distributions through parameterized Gaussian diffusion, alleviating feature confusion in incremental phases. Experiments on public SAR datasets show MPFR achieves superior performance compared to existing approaches, including recent SAR-specific CIL methods. Ablation studies validate each component’s contribution, confirming MPFR’s effectiveness in addressing CIL for SAR ATR without storing historical raw data.

Multi-view clustering integrates complementary signals from multiple heterogeneous views to reveal hidden grouping structures. It has become increasingly relevant in large-scale practical scenarios, such as recommendation systems and multimedia processing. Although considerable progress has been achieved, many deep multiview clustering approaches still struggle to effectively capture the underlying data structure: low-rank properties inherent in multiview data are often overlooked in representation learning, cross-view interactions are often constrained by inflexible alignment schemes that restrict view diversity, and scalability is frequently achieved through ad hoc approximations or loosely coordinated components. In this work, we present LADA-Cluster, a low-rank-aware deep clustering framework designed for large-scale multiview data. The proposed approach integrates low-rank structure directly into representation learning by projecting view-specific embeddings into a shared low-dimensional subspace while preserving view-dependent information through adaptive alignment. In addition, a structure-aware hierarchical anchor graph fusion mechanism is developed to support efficient and scalable relational modeling. Evaluations conducted on several large-scale real-world multiview datasets indicate that LADA-Cluster delivers consistently strong clustering performance, achieves stronger robustness under noisy and incomplete views, and maintains favorable scalability as the data scale increases, highlighting the effectiveness of low-rank-aware modeling for deep multiview clustering.

ABSTRACT Most existing Transformer‐based visual object tracking methods rely exclusively on the feature map from the last encoder layer for object prediction, thereby overlooking the rich information contained in shallow and intermediate layer feature maps. This limitation reduces the representational capacity of the model. Moreover, current multi‐modal tracking frameworks typically construct multi‐modal features through simple concatenation, which fails to adequately account for the differential contributions of individual modalities to the final prediction task. As a result, these approaches exhibit an insufficient ability to express key features within the multi‐modal representation. To address the aforementioned issues, this paper proposes a multi‐modal channel attention tracking algorithm, where a multi‐modal channel attention block is incorporated for the purpose of enhancing the representation ability of the key features within the multi‐modal features. Specifically, the multi‐modal channel attention block first aggregates multi‐modal information from the multi‐layer feature maps of the encoder through cross layer cascading and then applies channel attention mechanism to dynamically calibrate the channel weights in the generated multi‐modal features, thereby enhancing the representation of key features. In addition, this article proposes a new regression loss function to improve localisation accuracy. Finally, abundant experiments conducted on five benchmarks including GOT‐10K, TrackingNet, TNL2K, VisEvent and RGBT234 have verified the effectiveness of our theory.

Temporal Knowledge Graph Completion (TKGC) aims to infer missing facts by modeling historical events and latent temporal dependencies in Temporal Knowledge Graphs (TKGs). Recently, TKGC methods that integrate graph embeddings into Large Language Models (LLMs) have shown great promise by leveraging the structural information of TKGs together with the powerful reasoning capabilities of LLMs. However, these embedding-based methods are limited by suboptimal graph representations due to noise and long-tail issues in real-world scenarios, and insufficient cross-modal alignment between graph and language, hindering LLMs' ability to fully capture the temporal and structural information of TKGs. To address these issues, we propose TGCA-LLM, a novel embedding-based framework for TKGC. Specifically, TGCA-LLM first employs time-aware contrastive learning to align fact texts with graph structures in the temporal dimension, generating robust graph embeddings and establishing initial cross-modal alignment. Then, through a two-stage tuning process, it enables LLMs to gradually acquire structural and temporal knowledge from graph embeddings while enhancing their cross-modal reasoning capabilities in TKGC. Extensive experiments on three widely used real-world benchmarks demonstrate that TGCA-LLM outperforms state-of-the-art (SOTA) baselines by at least 8.7% MRR, highlighting its effectiveness.

A long-lived LLM agent, such as OpenClaw, earns its value by acting on a user's preferences and constraints across sessions, not just the current request. Yet today's agents keep what a user volunteers but rarely ask for what stays unspoken, leaving a proactivity gap in long-lived LLM agents: an agent cannot act on a preference it never obtained. As users delegate more of their affairs to agents, the impact of this gap grows. We isolate one concrete, controllable slice of this gap as Ask-to-Remember (ATR): the agent decides whether to ask now for a reusable user preference that the current task does not need but a later session with the same user will. ATR is hard even to evaluate: the right question is underdetermined and its payoff deferred to tasks that may never arise. ATRBench, to the best of our knowledge the first ATR benchmark, makes it measurable by fixing each user's preferences as hidden ground truth, so success demands asking, not recall. Across eight frontier LLM agents, defaults fall at least 62 points below an oracle handed the relevant preference, and prompting closes little of it. Diagnostics identify acquisition as the bottleneck. ATRBench surfaces this proactivity gap in current agents and offers a diagnostic testbed for closing it.

Training performance of athletes relies on the training initiative, which in turn is related to the construction of training plans. To forecast training performance of athletes, based on the Event-group training theory, this paper proposed a novel forecast method through combining with the discrete Hopfield neural network and wavelet function. The critical principle is that using Event-group training theory to construct 13 training indicators for athletes. The training scores of athletes is calculated by the constructed 13 training indicators. According to the designed training indicators, the discrete Hopfield neural network with wavelet function is implemented. In process of model training, to complete quick convergence and to guarantee the stability of state update to our discrete Hopfield neural network, the designed wavelet function is taken. Following that, the proposed model is trained by using the experimental dataset, and using the trained model to forecast the training performance of athletes. Experimental results show that the proposed model not only accurately forecasts the training performance of athletes, but also outperformed the comparative models in forecast accuracy. Results also show that the running efficiency of the proposed model won major competitors. The value of the designed training indicators not only finds these major factors affecting training performance of athletes, but also assists coaches in scientifically specifying training plans and observing athletes' training performance.

Target template updating is a crucial strategy in visual object tracking for capturing target state changes in real-time. Current implicit template update methods based on a single token exhibit limited representational capacity and inefficient interaction, hindering the full utilization of fine-grained information between the template and the search sequence. To overcome these limitations, this paper proposes a visual tracking method based on multi-token encoding and token convolutional interaction (MCTrack). The method employs a multi-token encoder to decompose a single template token into multiple parallel, collaborating sub-tokens, enhancing the representation of multi-dimensional target features. Furthermore, a token convolutional interaction mechanism is introduced, where multiple tokens are served as convolutional kernels to perform local structure-aware matching with search features. Experimental results on multiple benchmark datasets such as LaSOT and GOT-10k demonstrate that MCTrack outperforms existing state-of-the-art methods in both tracking accuracy and robustness.

Chinese classical poetry, inheriting thousands of years of Chinese civilization, reflects the social ethos and cultural aspects of its times. In recent years, researchers have increasingly focused on using artificial intelligence to analyze Chinese classical poetry. Many of these studies rely on pre-trained language models. Unfortunately, Chinese classical poetry has a unique form and the direct use of a general pre-trained language model is ineffective. Its long time span, frequent changes in language meaning, and small amount of training data limit the development of pre-trained models for Chinese classical poetry. To address these challenges, we construct a dynamic pre-trained model for Chinese classical poetry, based on SikuBERT and using comparative learning and multi-task training strategy. During the training process, we search for hard negative and positive examples and use them for data augmentation. And we introduce sliding window to dynamically learn poetry information. Compared to the encoding provided by the baseline model, our model's encoding achieves better performance in downstream tasks classification, translation and poem-poet matching.

Video social relation recognition is a fundamental task in video understanding, which is dedicated to the construction of multi-modal knowledge graphs. Previous work mainly focuses on multi-modal fusion and the construction of special character graphs. However, they often treat the global frame sequence equally, ignoring the influence of key frame sequence on relation recognition. Specifically, the key frame sequence that significantly reflect character relationships in a video tends to be sparse and short. At the same time, the key frames have not only temporal but also strong causal relationship. Therefore, we propose a novel Video Local Causal Frame (VLCF) model to explore the causal relationship between frames. Inspired by Granger causality theory, we estimate inter-frame causal relationships by comparing the predicted result frames with and without masking the premise frame. We then construct global connections between video frames. Multiple local causal frame sequences and global frame sequences are extracted to capture the key information and global information in the video. Extensive experiments conducted on the ViSR dataset and the MovieGraphs dataset demonstrate that the proposed model achieves state-of-the-art performance.

To address the problem of maximizing desired opinions in social networks, we present the Limited Opinion Maximization with Dynamic Propagation Optimization framework, which is grounded in information entropy theory. Innovatively, we introduce the concept of node expression capacity, which quantifies the uncertainty of users’ expression intentions via entropy and effectively identifies the impact of silent nodes on the propagation process. Based on this, in terms of seed node selection, we develop the Limited Opinion Maximization algorithm for multi-stage seed selection, which dynamically optimizes the seed distribution among communities through a multi-stage seeding approach. In terms of node opinion changes, we establish the LODP dynamic opinion propagation model, reconstructing the node opinion update mechanism and explicitly modeling the entropy-increasing effect of silent nodes on the information propagation path. The experimental results on four datasets show that LOMDP outperforms six baseline algorithms. Our research effectively resolves the problem of maximizing desired opinions and offers insights into the dynamics of information propagation in social networks from the perspective of entropy and information theory.

Recommender systems (RS) serve as a fundamental tool for navigating the vast expanse of online information, with deep learning advancements playing an increasingly important role in improving ranking accuracy. Among these, graph neural networks (GNNs) excel at extracting higher-order structural information, while large language models (LLMs) are designed to process and comprehend natural language, making both approaches highly effective and widely adopted. Recent research has focused on graph foundation models (GFMs), which integrate the strengths of GNNs and LLMs to model complex RS problems more efficiently by leveraging the graph-based structure of user-item relationships alongside textual understanding. In this survey, we provide a comprehensive overview of GFM-based RS technologies by introducing a clear taxonomy of current approaches, diving into methodological details, and highlighting key challenges and future directions. By synthesizing recent advancements, we aim to offer valuable insights into the evolving landscape of GFM-based recommender systems.

This paper addresses the need for real-time monitoring and early warning of the psychological state of emergency personnel in the unique environment of the plateau. To address the challenges faced by existing methods in plateau hypobaric, hypoxic, and high-stress tasks, such as strong signal interference, large individual differences, and insufficient real-time performance, a real-time early warning model for psychological abnormalities based on multimodal physiological signals is proposed. First, a physiological signal database for plateau conditions is constructed, then, a spatiotemporal convolutional network incorporating an attention mechanism is designed, an adaptive threshold learning mechanism is introduced, finally, a lightweight edge computing deployment solution is constructed to achieve low-latency, real-time early warning. Experimental results demonstrate that the proposed model achieves an F1 score of 0.891 in the psychological abnormality recognition task and an average F1 score of 0.852 in the cross-individual generalization test. The edge inference latency fully meets the application requirements. Overall, the proposed model strikes a good balance between performance, real-time performance, and resource consumption, making it feasible for practical deployment.