Performing node classification in text-attributed graphs (TAGs) has become a critical area of research. While two-stage methods effectively address scalability by decoupling feature extraction from GNN training, they often suffer from a “structural gap” where global topology is ignored during text encoding, and local aggregation leads to feature oversmoothing. To bridge this gap, we propose DCD-CAWA, a framework that integrates community structures into both feature learning and message passing. Specifically, in the first stage, we employ Decoupled Community Detection (DCD) to generate structural priors. These are integrated with text attributes through an auxiliary multi-task fine-tuning strategy, forcing the language model to capture the correlations between semantic content and community membership. In the second stage, we introduce the Community-Aware Weighted Aggregation (CAWA) module. The workflow of CAWA involves computing dynamic attention weights between nodes and their respective community prototypes, allowing the model to adaptively refine node representations by emphasizing global community contexts. This approach not only eliminates discrepancies between static topology and dynamic features but also effectively mitigates oversmoothing by preserving community-level distinctiveness. Experimental results across multiple datasets demonstrate that DCD-CAWA significantly outperforms state-of-the-art baselines.

One-to-one (O2O) label assignment strategy has demonstrated superiority in crowded scenarios. However, owing to the uneven distribution of objects, a uniform label assignment criterion only achieves suboptimal results in drone images with many crowded regions. Dense objects exist in aggregated regions and are often semantically ambiguous, making them easily confused with one another. Sparse objects are typically located in regions dominated by background and are fewer in number compared to dense objects. For these two different types of objects, the use of the uniform label assignment criterion, i.e., O2O, leads to suboptimal solutions, as the network tends to be biased toward objects with more samples. In this work, we propose a density-aware adaptive label assignment to optimize dense and sparse objects simultaneously for end-to-end dense detectors in drone images. Specifically, we introduce the spatial distribution of ground truth as the prior information to categorize objects into dense and sparse. We apply the one-to-one label assignment for dense objects to avoid duplicate predictions. For sparse objects, which are fewer in number, we design a decreasing label assignment strategy. This strategy provides more positive samples for representation learning in the early training stage and gradually reduces the number of positive samples as training progresses to remove duplicate predictions. Experiments based on four commonly used dense detectors (e.g., FCOS, RetinaNet, ATSS, GFL V1) on three major benchmarks, i.e., VisDrone, COCO and CrowdHuman, demonstrate the effectiveness of our method.

Conversational agents struggle to handle long conversations due to context window limitations. Therefore, memory systems are developed to leverage essential historical information. Existing memory systems typically follow a pipeline of offline memory construction and update, and online retrieval. Despite the flexible online phase, the offline phase remains fixed and task-independent. In this phase, memory construction operates under a predefined workflow and fails to emphasize task relevant information. Meanwhile, memory updates are guided by generic metrics rather than task specific supervision. This leads to a misalignment between offline memory preparation and task requirements, which undermines downstream task performance. To this end, we propose an Adversarial Memory Adaptation mechanism (AMA) that aligns memory construction and update with task objectives by simulating task execution. Specifically, first, a challenger agent generates question answer pairs based on the original dialogues. The constructed memory is then used to answer these questions, simulating downstream inference. Subsequently, an evaluator agent assesses the responses and performs error analysis. Finally, an adapter agent analyzes the error cases and performs dual level updates on both the construction strategy and the content. Through this process, the memory system receives task aware supervision signals in advance during the offline phase, enhancing its adaptability to downstream tasks. AMA can be integrated into various existing memory systems, and extensive experiments on long dialogue benchmark LoCoMo demonstrate its effectiveness.

Knowledge-based Visual Question Answering (KB-VQA) requires models to answer visual questions by reasoning over external knowledge beyond the given image. Existing approaches suffer from two main limitations. First, candidate knowledge is often retrieved in a single modality, either textual or visual, which prevents effective use of heterogeneous and complementary evidence. Second, current approaches typically employ fixed fusion weights, ignoring the varying importance of modalities for different queries. Consequently, irrelevant evidence is introduced while critical knowledge may be overlooked. To address these issues, we propose Dynamic Multi-Path Retrieval for KB-VQA (DMRAG). Our framework retrieves candidates through multiple retrieval paths that capture complementary visual and semantic cues. It then performs Question-Adaptive Gated Fusion (QGF) to balance contributions from different modalities according to the query’s information need. The fused candidates are further refined via multimodal rearrangement to support accurate answer generation. Experiments on the E-VQA and InfoSeek datasets show that DMRAG improves both retrieval recall and answer accuracy over prior methods, demonstrating its effectiveness for KB-VQA. Our code is available at https://github.com/qwqq335/DMRAG.

Image-based quantified haze visibility evaluation is crucial for both atmospheric science and computer vision. Existing methods typically rely on datasets with precisely annotated visibility labels. However, the scarcity of real-world labeled data significantly hinders their scalability and practical deployment. To address this, we introduce HZSC, a real-world haze visibility dataset collected from traffic surveillance cameras. It consists of 36 diverse outdoor scenes, each containing 20 images ranked by haze visibility levels. Meanwhile, we propose the haze visibility evaluator (HVE) for referenceless image haze visibility assessment. HVE comprises three key components: a depth feature extractor, a haze visibility feature extractor, and a feature fusion module. The fused features are fed into learned fog-free and foggy Multivariate Gaussian (MVG) models to generate haze visibility scores. For training, We design a scoring-ranking framework using pairwise supervision, adopting a weighted hinge ranking loss to improve discrimination of subtle visibility variations. Extensive experiments on both the public MRFID dataset and our collected HZSC dataset demonstrate that our HVE consistently outperforms existing state-of-the-art methods. Examples of our dataset can be available at https://github.com/ZanChen1/HLVRS.

This paper presents DeRA, a novel 1D video tokenizer that decouples the spatial-temporal representation learning in video tokenization to achieve better training efficiency and performance. Specifically, DeRA maintains a compact 1D latent space while factorizing video encoding into appearance and motion streams, which are aligned with pretrained vision foundation models to capture the spatial semantics and temporal dynamics in videos separately. To address the gradient conflicts introduced by the heterogeneous supervision, we further propose the Symmetric Alignment-Conflict Projection (SACP) module that proactively reformulates gradients by suppressing the components along conflicting directions. Extensive experiments demonstrate that DeRA outperforms LARP, the previous state-of-the-art video tokenizer by 25

Existing multimedia recommender systems provide users with suggestions of media by evaluating the similarities, such as games and movies. To enhance the semantics and explainability of embeddings, it is a consensus to apply additional information (e.g., interactions, contexts, popularity). However, without systematic consideration of representativeness and value, the utility and explainability of embedding drops drastically. Hence, we introduce RVRec, a plug-and-play model-agnostic embedding enhancement approach that can improve both personality and explainability of existing systems. Specifically, we propose a probability-based embedding optimization method that uses a contrastive loss based on negative 2-Wasserstein distance to learn to enhance the representativeness of the embeddings. In addtion, we introduce a reweighing method based on multivariate Shapley values strategy to evaluate and explore the value of interactions and embeddings. Extensive experiments on multiple backbone recommenders and real-world datasets show that RVRec can improve the personalization and explainability of existing recommenders, outperforming state-of-the-art baselines.

Diagnostic prediction and clinical reasoning are critical tasks in healthcare applications. While Large Language Models (LLMs) have shown strong capabilities in commonsense reasoning, they still struggle with diagnostic reasoning due to limited domain knowledge. Existing approaches often rely on internal model knowledge or static knowledge bases, resulting in knowledge insufficiency and limited adaptability, which hinder their capacity to perform diagnostic reasoning. Moreover, these methods focus solely on the accuracy of final predictions, overlooking alignment with standard clinical reasoning trajectories. To this end, we propose MultiDx, a two-stage diagnostic reasoning framework that performs differential diagnosis by analyzing evidence collected from multiple knowledge sources. Specifically, it first generates suspected diagnoses and reasoning paths by leveraging knowledge from web search, SOAP-formatted case, and clinical case database. Then it integrates multi-perspective evidence through matching, voting, and differential diagnosis to generate the final prediction. Extensive experiments on two public benchmarks demonstrate the effectiveness of our approach.

Low-quality detection aims to enhance detector performance when handling multi-degraded inputs, such as fog and low-light conditions. Relying on numerous images and annotations, significant progress has been made in addressing each type of degradation. However, these methods have high-quality requirements for data size and model design, ignoring the transferability and generalization of features across different degradation conditions. Inspired by prompt learning in reconstituting the relationship between foundation models and downstream tasks, we propose a prompt-based task-conditioning transformer framework for low-quality detection (LQPrompt). In particular, we present a module for updating prompts (PUM). By combining both high- and low-quality features, PUM adaptively retains tailored visual representations as task-relevant prompts using an attention-like mask. To facilitate effective interaction between prompts and the foundation model, we design a prompt aggregation strategy. These designs enable the detector to learn the correlations and differences between features under different degradation conditions, enhancing generalization and improving detection performance. Experimental results demonstrate that LQPrompt achieves SOTA performance with minimal parameter cost across various low-quality scenarios without corresponding annotations. Moreover, LQPrompt is also effective on real-world datasets such as RTTS and ExDark.

Large language models have demonstrated strong reasoning capabilities in general knowledge question answering. However, their ability to handle temporal information remains limited. To address this limitation, existing approaches often involve external tools or manual verification and are tailored to specific scenarios, leading to poor generalizability. Moreover, these methods apply a fixed pipeline to all questions, overlooking the fact that different types of temporal questions require distinct reasoning strategies, which leads to unnecessary processing for simple cases and inadequate reasoning for complex ones. To this end, we propose AdapTime, an adaptive temporal reasoning method that dynamically executes reasoning steps based on the input context. Specifically, it involves three temporal reasoning actions: reformulate, rewrite and review, with an LLM planner guiding the reasoning process. AdapTime integrates seamlessly with state-of-the-art LLMs and significantly enhances their temporal reasoning capabilities without relying on external support. Extensive experiments demonstrate the effectiveness of our approach.

Human pose estimation (HPE) is a fundamental problem in computer vision, aiming to locate anatomical keypoints of the human body in a given picture. Benefiting from recent progress in deep learning, dominant HPE methods can achieve more advanced performance. Unfortunately, these methods rely heavily on large-scale, high-quality datasets captured expensively, resulting in limited learning capabilities in data-constrained low-light situations. Existing methods enhance the model's ability for low-light scenarios by performing intermediate feature alignment between low-light image and its well-lit counterpart. However, these methods fall short in fully exploiting explicit semantic feature exploitation that is independent of lighting conditions, resulting in sub-optimal performance. In this paper, we propose a Progressive Decomposition-Modulation network (PDMNet) for human pose estimation in extremely low-light condition. In particular, PDMNet mainly consists of 1) a semantic-specific decomposition module (SDM) for decomposing reflectance component with rich semantic information, and 2) a semantic-specific modulation mechanism (SMM) that enables the reflectance component to modulate the representation learning of human body parts in a tailored manner. Two closely-related components cooperate with each other to achieve more effective content-specific feature learning in low-light conditions. We further equip them progressively into different scales to enhance the feature learning. Experimental results demonstrate the superiority of PDMNet over state-of-the-art models on publicly available datasets. Our code will be released soon.

Vision Transformer (ViT) on object re-identification (ReID) has attracted significant attention recently. However, ViT-based ReID substantially increases computational complexity, imposing significant burdens during training and inference. This paper presents an efficient ViT-based backbone for ReID tasks, called the Locally Enhanced Vision Transformer (LEViT). ViT models typically emphasize global relationship modeling, yet ReID tasks are more sensitive to local information. To address this gap, we propose a Locally Enhanced (LE) block to enhance local information by performing self-attention within local split windows. Since part-based models dominate ReID, calculating self-attention across all patches is computationally inefficient. We also replace the traditional Query-Key-Value projector with the Group Convolution (G-Conv) projector, enabling the model to capture local details. Furthermore, G-Conv is integrated into the channel MLP to strengthen local feature sensitivity. Using these components, we develop two LEViT variants: LEViT-S and LEViT-L. To our knowledge, LEViT is the first highly adaptable ViT backbone for ReID tasks. Experimental evaluations demonstrate the effectiveness in five ReID datasets and three deep metric learning datasets. Notably, LEViT-S outperforms TransReID while requiring less than 10% computational complexity.

Most modern Convolutional Neural Networks (CNNs) employ a multi-branch structure with various-sized convolutions to capture long- and short-range dependencies. However, these CNNs use large kernel convolutions (e.g., astonishingly 101 kernels) and specialized techniques (e.g., reparameterization and sparsity), increasing complexity in both training and inference stages. This paper focuses on designing an efficient CNN based on pure 3×3 convolutions without introducing complex operations and techniques. Specifically, we propose a Spatial Pyramid (SP) block, which consists of the Multi-branch Residual (MbR) module and the Gated-branch Residual (GbR) module. The MbR introduces multiscale pooling as the key component, thus capturing long-range visual cues through large down-sampling rates and shorter-range dependencies through low down-sampling rates while maintaining low computational complexity. Besides, the GbR uses one 3×3 convolution to refine dependencies along spatial and channel dimensions. Based on the SP block, we construct the Spatial Pyramid CNN (SPCNN), a model composed exclusively of Point-Wise Convolution and 3×3 Depth-Wise Convolution. Under comparable computational complexity, SPCNN significantly outperforms the state-of-the-art CNN PeLK (83.6% vs 82.6%) with only 3 × 3 kernels (compared to 101 × 101 kernels in PeLK). Besides, our SPCNN demonstrates comparability with state-of-the-art backbones in lightweight models, object detection, instance segmentation, and semantic segmentation. Moreover, evaluations of four image retrieval benchmarks also demonstrate the effectiveness. All codes are released at https://github.com/xiaolai-sqlai/SPCNN.

With the development of deep learning in recent years, the performance of object detection under conventional cameras has been significantly improved. Nevertheless, due to the distortion caused by the fisheye cameras, detecting objects in this scenario remains a significant challenge. The dominant approaches focus on modifying the shape of the bounding box to better align the boundaries of the distorted object. However, these methods neglect the learning of spatial distortion information, which prevents them from satisfactory results. In this paper, we propose a novel fisheye camera detection network to learn distortion features better, dubbed SDANet. SDANet is composed of a series of SDABlocks, which are designed to learn spatial distortion features. Each SDABlock consists of multiple convolution kernels of different sizes, and it can generate the most suitable kernel based on the current input's distortion characteristics. Moreover, to address the limitations of the scarcity and uneven spatial distribution of fisheye image datasets on performance improvement, we propose a dedicated data augmentation strategy called Prominent Fisheye Distortion Augmentation (PFDAug). PFDAug can further introduce distortions to fisheye images, effectively alleviating these problems. Experimental results on the CEPDOF, MW-R, HABBOF, LOAF, and FishEye8k fisheye image datasets demonstrate that our method achieves state-ofthe-art performance.

Due to factors such as dense placement, small size, and similar appearance, accurately detecting products on retail shelves remains a significant challenge. To address these challenges, most of the existing methods use advanced detectors directly or focus only on improving feature representation capability during feature extraction. However, these methods ignore the spatial characteristics of products on shelves, i.e., items in the same category or with similar attributes are often clustered and exhibit similar scales, resulting in sub-optimal performance. In this paper, we propose a novel product detector, called SAPDet, to better learn the spatial distribution properties of shelf items. In particular, a Hybrid Information Capture (HIC) module is used to integrate 1) the product distribution information obtained by clustering and 2) the product structure information obtained by high-frequency enhancement after FPN. We formulate them as visual tokens in the proposed two-stream attention mechanism, boosting the spatial representation of features. Following the HIC module, the Attention-based Iterative Regression (AIR) leverages a cross-attention mechanism to further integrate the aforementioned information into the feature map, refining boundary box predictions through precise offset corrections. These designs effectively facilitate the spatial characteristics learning of products and improve the performance of the detector in shelf scenarios. Experimental results demonstrate the superiority of SAPDet over state-of-the-art models on publicly available datasets. On the most challenging Unitail dataset, SAPDet gains 4.1% AP improvements. Our code will be released soon.

Recommender systems are increasingly reliant on rich and diverse user data, but growing privacy regulations and data silos severely limit centralized data collection. As a result, extending them to federated learning (FL) frameworks to address this challenge by enabling collaborative model training without sharing raw data has recently gained widespread attention. In this survey, we propose a novel perspective that formalizes FedRec as a problem of distributed knowledge fusion under data isolation, where raw interaction logs remain local and knowledge distilled from local data is integrated across distributed clients. This perspective reveals a critical insight that the questions of what to fuse and how to fuse have fundamentally different answers across scenarios, including collaborative, cross-domain, multi-modal, and LLM-based FedRec. We systematically analyze and summarize the knowledge types, fusion strategies, and technical bottlenecks for each scenario. Through this scenario-oriented perspective, we uncover open challenges invisible from a unified optimization perspective: the different causes of statistical heterogeneity, the profound impact of pseudo-interactions on cross-source alignment, and the inherent trade-off between fusion scope and model capacity. This paper aims to offer researchers and practitioners a unified, practically-oriented and scenario-oriented perspective for FedRec, and to suggest promising future research directions. To facilitate subsequent exploration, we maintain a continuously updated repository at: https://github.com/SmilesLab-XJTU/Survey-FedRec.