🏢 Zhejiang University

Ex-MCR: Efficiently build unified multi-modal representations by extending, not connecting, pre-trained spaces, achieving superior performance with less paired data and training.

eFreeSplat: a novel, epipolar-free 3D Gaussian splatting model for generalizable novel view synthesis, surpassing state-of-the-art methods by achieving superior geometry reconstruction and novel view …

RDSS: a novel sample selection method for semi-supervised learning, boosts model accuracy by minimizing a-MMD, striking a balance between sample representativeness and diversity.

Boosting LLM trustworthiness, researchers introduce Sparse Activation Control, a training-free method that concurrently enhances safety, factuality, and bias mitigation by selectively controlling atte…

LLMs struggle with complex, long-form text. This paper introduces ‘context structurization,’ transforming unstructured text into a structured format to enhance LLM comprehension. Experiments across …

Vision-Augmented Prompting (VAP) boosts LLM reasoning by automatically generating images from textual problem descriptions, incorporating visual-spatial clues to significantly improve accuracy across …

Dual-Perspective Activation (DPA) efficiently denoises ANN channels by jointly using forward and backward propagation criteria, improving sparsity and accuracy.

DRIP: A novel image matting method using pre-trained latent diffusion models achieves state-of-the-art performance by jointly predicting foreground and alpha values, significantly improving accuracy a…

DreamMesh4D: Generating high-fidelity dynamic 3D meshes from monocular video using a novel Gaussian-mesh hybrid representation and adaptive hybrid skinning.

DMNet: A novel self-comparison driven model significantly improves subject-independent seizure detection from intracranial EEG, outperforming existing methods.

DiffPano generates scalable, consistent, and diverse panoramic images from text descriptions and camera poses using a novel spherical epipolar-aware diffusion model.

Large language models struggle to generalize knowledge when facing seemingly simple reversals, a phenomenon termed the ‘reversal curse.’ This study reveals that this limitation is strongly linked to t…

DECRL: A novel deep learning approach for temporal knowledge graph representation learning, capturing high-order correlation evolution and outperforming existing methods.

Attackers can steal massive private data from federated learning diffusion models using multiple Trojans and an advanced attack, AdaSCP, which circumvents existing defenses.

Data-faithful feature attribution tackles misinterpretations from unobservable confounders by using instrumental variables to train confounder-free models, leading to more robust and accurate feature …

LeapAD, a novel autonomous driving paradigm, uses a dual-process architecture mirroring human cognition to achieve continuous learning and improved adaptability. Employing a VLM for efficient scene u…

CGFormer: a novel voxel transformer boosting semantic scene completion accuracy by using context-aware queries and 3D deformable attention, outperforming existing methods on SemanticKITTI and SSCBench…

Con4m, a novel consistency learning framework, leverages contextual information to effectively classify segmented time series with inconsistent boundary labels and varying durations of classes, signif…

Chat-Scene: Bridging 3D scenes and LLMs using object identifiers for efficient, object-level interaction and improved scene comprehension.

New efficient attack reveals GNN model training data properties.