🏢 Hong Kong University of Science and Technology

Vista: a novel driving world model achieving high-fidelity prediction and versatile controllability, outperforming state-of-the-art models in generalization and prediction accuracy.

VeXKD: A versatile framework boosts 3D perception by cleverly combining cross-modal fusion and knowledge distillation, improving single-modal student model accuracy without extra inference time.

UrbanKGent: A unified LLM agent framework revolutionizes urban knowledge graph construction, achieving significantly improved accuracy and efficiency.

UPS: A novel algorithm for lightweight single-image super-resolution, decoupling feature extraction and similarity modeling for enhanced efficiency and robustness.

Two-Stage Backdoor Defense (TSBD) unveils and mitigates backdoor vulnerabilities by cleverly unlearning weight changes and suppressing backdoor neuron activeness, significantly improving the robustnes…

Current backdoor defenses, while effective at reducing attack success rates, are vulnerable to rapid re-learning. This work unveils this superficial safety, proposes a novel attack, and introduces a p…

UltraPixel generates high-quality images at various resolutions (1K-6K) efficiently using cascade diffusion models, achieving state-of-the-art performance.

R2ET: training for robust ranking explanations by an effective regularizer.

ATProt: Adversarial training makes protein interface prediction robust to flexibility!

Object-centric occupancy completion boosts 3D object detection accuracy by using temporal information from long sequences to precisely reconstruct object shapes, particularly for incomplete or distant…

TIME-FFM: a Federated Foundation Model empowers time series forecasting using pre-trained Language Models, tackling data scarcity and privacy concerns for superior few-shot and zero-shot predictions.

This paper reveals fundamental limits of differential privacy in online learning, demonstrating a clear separation between pure, approximate, and non-private settings.

Adam’s implicit bias revealed: On separable data, Adam converges towards the maximum l∞-margin solution, a finding contrasting with gradient descent’s l2-margin preference. This polynomial-time conver…

TMEA: A novel approach significantly boosts multi-modal entity alignment accuracy by effectively handling uncertain correspondences between modalities, improving data integration for diverse knowledge…

SubgDiff enhances molecular representation learning by incorporating substructural information into a diffusion model framework, achieving superior performance in molecular force predictions.

SpikeSlicer: An adaptive event stream slicer using a spiking neural network (SNN) to efficiently split events for improved downstream processing in object tracking and recognition.

SpGesture: A source-free domain-adaptive SEMG gesture recognition system using a novel Spiking Jaccard Attentive Neural Network achieves real-time performance with high accuracy.

Language model editing’s limitations exposed: Scaling current methods leads to knowledge loss and compromised safety, urging research into more robust techniques.

RouterDC: A query-based router trained via dual contrastive learning assembles multiple LLMs, significantly outperforming individual LLMs and existing routing methods on both in- and out-of-distributi…

RL-GPT seamlessly integrates Large Language Models (LLMs) and Reinforcement Learning (RL) to create highly efficient agents mastering complex tasks in open-world environments.