🏢 Hong Kong University of Science and Technology

Reverse Transition Kernel (RTK) framework accelerates diffusion inference by enabling balanced subproblem decomposition, achieving superior convergence rates with RTK-MALA and RTK-ULD algorithms.

RestoreAgent, an AI-powered image restoration agent, autonomously identifies and corrects multiple image degradations, exceeding human expert performance.

QGFN boosts Generative Flow Networks (GFNs) by cleverly combining their sampling policy with an action-value estimate, creating controllable and efficient generation of high-reward samples.

Offline RL struggles with OOD action overestimation. QDQ tackles this by penalizing uncertain Q-values using a consistency model, enhancing offline RL performance.

Phased Consistency Models (PCMs) revolutionize diffusion model generation by overcoming LCM limitations, achieving superior speed and quality in image and video generation.

Performative control, where control policies change system dynamics, is analyzed; offering sufficient conditions for unique solutions, and proposing a convergent algorithm for achieving them.

SSCNN, a novel decomposition-based model, achieves superior long-term time series forecasting accuracy using 99% fewer parameters than existing methods, proving that bigger isn’t always better.

LiDAR Translator (LiT) unifies diverse LiDAR data through a novel data-driven translation framework, enabling zero-shot and multi-domain joint learning, thus improving autonomous driving systems.

LISA, a layerwise importance sampling method, dramatically improves memory-efficient large language model fine-tuning, outperforming existing methods while using less GPU memory.

Actionable AI agents are trained efficiently via a novel framework, VPDD, which uses discrete diffusion to pre-train on massive human videos, and fine-tunes on limited robot data for superior multi-ta…

LaSe-E2V: Language-guided semantic-aware event-to-video reconstruction uses text descriptions to improve video quality and consistency.

Kaleidoscope: Learnable Masks for Heterogeneous MARL achieves high sample efficiency and policy diversity by using learnable masks for adaptive partial parameter sharing.

Boosting Neural ODE training, Temporal Adaptive Batch Normalization (TA-BN) resolves traditional Batch Normalization’s limitations by providing a continuous-time counterpart, enabling deeper networks …

This paper significantly improves Bayes regret bounds for hierarchical Bayesian bandit algorithms, achieving logarithmic regret in finite action settings and enhanced bounds in multi-task linear and c…

HOPE: a novel shape matching method achieving both accuracy and smoothness by aligning different k-hop neighborhoods and refining maps via local map distortion.

HAWK: a novel framework leveraging interactive VLMs and motion modality achieves state-of-the-art performance in open-world video anomaly understanding, generating descriptions and answering questions…

GVKF: A novel method achieves highly efficient and accurate 3D surface reconstruction in open scenes by integrating fast 3D Gaussian splatting with continuous scene representation using kernel regres…

GITA, a novel framework, integrates visual graphs into language models for superior vision-language graph reasoning, outperforming existing LLMs and introducing the first vision-language dataset, GVLQ…

GIC: Novel hybrid framework leverages 3D Gaussian representation for accurate physical property estimation from visual observations, achieving state-of-the-art performance.

Boost pathology model accuracy with Concept Anchor-guided Task-specific Feature Enhancement (CATE)! This adaptable paradigm enhances feature extraction for specific tasks using task-relevant concepts,…