🎯 What it does: This paper proposes an Uploadable Multi-source Few-shot Domain Adaptation (UMFDA) framework and designs a Visual Perception Multi-modal Prompt Tuning (VAMP) scheme to achieve collaborative transfer across multiple source domains in low-computation, low-annotation environments on edge devices.
Vision-guided Text Mining for Unsupervised Cross-modal Hashing with Community Similarity Quantization
Haozhi Fan (University of Pennsylvania), Yuan Cao (Ocean University of China)
CodeObject DetectionRetrievalOptimizationVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes an unsupervised cross-modal hashing method VTM-UCH based on visually guided text mining, which enhances text semantics using CLIP and object detection, and optimizes hash distribution through community detection.
VisRec: A Semi-Supervised Approach to Visibility Data Reconstruction in Radio Astronomy
Ruoqi Wang (Hong Kong University of Science and Technology), Hejun Wu (Guangzhou University)
CodeRestorationConvolutional Neural NetworkSupervised Fine-TuningNeural Radiance FieldImagePhysics Related
🎯 What it does: We propose VisRec, a model-agnostic semi-supervised learning framework for the reconstruction of visibility data from sparse to dense in radio interferometry.
Pengcheng Zhang (Beihang University), Jin Zheng (Beihang University)
CodeRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a Visual Perturbation Network (ViPer) that improves the alignment of visual and linguistic features for the Text-Based Person Search (TBPS) task.
🎯 What it does: A framework called ResAct for residual action learning and observation difference learning in visual RL is proposed to simplify action learning.
🎯 What it does: This paper presents VOILA, a general CT image segmentation method based on voxel-language alignment, utilizing variable voxel sampling and contrastive learning to achieve multi-class segmentation.
VProChart: Answering Chart Question Through Visual Perception Alignment Agent and Programmatic Solution Reasoning
Muye Huang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
CodeTransformerLarge Language ModelAgentic AIImageText
🎯 What it does: The VProChart framework is proposed, which combines a lightweight visual alignment agent (VPAgent) with LLM-based programmatic solution reasoning for chart question-answering tasks.
VQA4CIR: Boosting Composed Image Retrieval with Visual Question Answering
Chun-Mei Feng (Institute of High Performance Computing Agency for Science Technology and Research), Yong Liu (Harbin Institute of Technology)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a post-processing framework called VQA4CIR, which first generates QA pairs related to relative descriptions using LLaMA, and then performs visual question answering (VQA) with Fine-tuned LLaVA to self-validate the consistency of the retrieval results with the descriptions, and accordingly re-ranks the retrieval list to improve the performance of synthesized image retrieval.
VTG-LLM: Integrating Timestamp Knowledge into Video LLMs for Enhanced Video Temporal Grounding
Yongxin Guo (Chinese University of Hong Kong), Kevin Zhao (Tencent)
CodeRecognitionRetrievalOptimizationTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: This paper proposes VTG-LLM, which significantly improves the zero-shot performance of video large language models in video temporal localization tasks by integrating timestamp knowledge.
Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models
Susmit Agrawal (Indian Institute of Technology Hyderabad), Vineeth N. Balasubramanian (Indian Institute of Technology Hyderabad)
CodeClassificationExplainability and InterpretabilityTransformerImageMultimodality
🎯 What it does: This paper studies the maintenance and expansion of concept and category relationships in the context of incremental learning, and proposes a MuCIL model that achieves multimodal concept embedding without increasing parameters.
Wasserstein Distance Constraint and Parameter Sparsification for Batched and Iterative Knowledge Editing
Shanbao Qiao (Jeonbuk National University), Seung-Hoon Na (Jeonbuk National University)
CodeTransformerLarge Language ModelText
🎯 What it does: This study investigates the issue of parameter distribution drift leading to model performance collapse during batched iterative editing on large language models, and proposes two improvement strategies: Wasserstein distance constraint and parameter sparsification, to maintain the stability of model parameter distribution and enhance editing effectiveness.
🎯 What it does: This paper proposes a t-distribution embedding method based on Wasserstein distance, called WatE, which uses GNN to learn the mean and covariance of the node embedding distribution for each graph, visualizing the graph as an ellipse, thus balancing graph-level clustering and node-level structural information.
🎯 What it does: A multi-frequency attention network based on wavelet transform, WFANet, is proposed for the fusion of high-resolution multispectral images;
Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration
Siyang Feng (Guilin University of Electronic Technology), Xipeng Pan (Guilin University of Electronic Technology)
CodeSegmentationConvolutional Neural NetworkImage
🎯 What it does: A weakly supervised gland segmentation method is proposed, achieving high-precision segmentation through improved CAM generation and pseudo-label filtering.
🎯 What it does: A new weighted embedding method (WEMBED) is proposed, which assigns weights to each node and uses weighted Euclidean distance to approximate hyperbolic geometry, generating low-dimensional graph embeddings.
What Is a Good Question? Assessing Question Quality via Meta-Fact Checking
Bo Zhang (Nanjing Normal University), Junsheng Zhou (Beihang University)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the Meta-Fact Checking (MFC) method, which interacts with large language models (LLMs) and knowledge graphs (KGs) to obtain complete knowledge, enabling automatic quality assessment of knowledge-based questions and improving LLM performance in multi-hop reasoning tasks.
CodeComputational EfficiencyGraph Neural NetworkLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This study investigates the issue of visual token redundancy in multimodal large language models (MLLMs) and proposes a training-independent graph-structured visual token pruning method called G-Prune, which can retain important tokens for both foreground and background while significantly reducing computational load.
🎯 What it does: Compared the learning efficiency, progressive performance, and computational cost of State-to-Visual DAgger and Visual RL across 16 different tasks (from ManiSkill, DMControl, Adroit), systematically evaluating the strengths and weaknesses of the two paradigms and providing practical recommendations.
🎯 What it does: This paper proposes the Witness Graph Topological Layer (WGTL), which introduces persistent homology and witness complexes to provide a defense mechanism against adversarial attacks for Graph Neural Networks (GNNs); WGTL enhances robustness against interference in graph node classification tasks.
🎯 What it does: This study investigates the impact of interpolation-based data augmentation (Mixup and Fair Mixup) on the multicalibration fairness and accuracy of binary classification models in multi-group small sample scenarios, comparing it with traditional post-processing multicalibration methods.
Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity
Zhufeng Li (Technical University of Munich), Niki Kilbertus (Max Planck Institute for Biology)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelBiomedical Data
🎯 What it does: A Transformer framework based on whole-genome sequences is proposed for predicting habitat specificity in microorganisms and explaining gene interactions.
🎯 What it does: A self-supervised graph contrastive learning framework EPAGCL based on Error Propagation Rate (EPR) is proposed, which can selectively add and delete edges during graph view generation, thereby maintaining a low error propagation rate of the graph.
🎯 What it does: A large-scale AI-generated image detection dataset, WildFake, has been constructed and made publicly available, providing generated images across multiple categories, architectures, weights, times, and versions, and evaluating the generalization and robustness of detectors based on this dataset.
xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition
Artyom Stitsyuk (Korea Advanced Institute of Science and Technology), Jaesik Choi (Korea Advanced Institute of Science and Technology)
CodeOptimizationConvolutional Neural NetworkRecurrent Neural NetworkTime Series
🎯 What it does: A dual-stream (MLP+CNN) structure called xPatch is proposed, which achieves long sequence time series prediction through exponential moving average decomposition combined with channel independence and slicing techniques.
🎯 What it does: The Adaptive Stop Denoising Network (ASDN) is proposed, which prevents over-smoothing by adaptively stopping the denoising of cleaned points, thereby improving the quality of point cloud denoising.
Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling
Xinyue Fang (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
CodeGenerationAnomaly DetectionGraph Neural NetworkLarge Language ModelTextGraph
🎯 What it does: This paper proposes a zero-resource hallucination detection method for long text generation, utilizing a graph structure to perform consistency comparisons on extracted knowledge triples, and strengthening detection through three reverse verification tasks.
🎯 What it does: A zero-shot conditional sampling method without training is proposed, utilizing a pre-trained unconditional score diffusion model and soft constraints to sample under arbitrary logical constraints.
🎯 What it does: Proposed the MMIO industrial defect open dataset and the RTVP zero-shot detection method based on expert guidance, sparse modeling, and cross-modal interaction.
Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models
Yan Huang (South China University of Technology), Yong Xu (South China University of Technology)
CodeRestorationDiffusion modelImage
🎯 What it does: A zero-shot low-light image enhancement framework is proposed, utilizing a pre-trained latent diffusion model to achieve low-light enhancement without the need for specialized training.
🎯 What it does: A zero-shot image denoising method ZS-N2M is proposed, which can train and achieve high-quality denoising results using only a single noisy image;
🎯 What it does: This paper proposes a zero-shot scene change detection method that does not require training, utilizing the pre-trained segmentation model SAM and the tracking model DEVA to achieve change recognition across temporal images.
🎯 What it does: A zero-shot video restoration and enhancement framework is proposed, utilizing a pre-trained image diffusion model and incorporating spatiotemporal attention, temporal consistency guidance, spatial-temporal noise sharing, and early stopping sampling.
🎯 What it does: This paper proposes ZeroMamba, a zero-shot learning framework based on Vision Mamba, which integrates three main modules: semantic-aware local projection, global representation learning, and semantic fusion, enhancing visual-semantic interaction.
Zeroth-Order Methods for Nonconvex Stochastic Problems with Decision-Dependent Distributions
Yuya Hikima (University of Tokyo), Akiko Takeda (University of Tokyo)
CodeOptimizationGaussian SplattingTabular
🎯 What it does: Two zeroth-order methods for decision-related stochastic non-convex optimization problems are proposed, along with convergence and sample complexity analysis.
CodeObject DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: A framework for zero-shot remote sensing instance segmentation, ZoRI, has been developed, utilizing the CLIP model to achieve segmentation of unseen categories.