🏢 Peking University

Zero-shot Event-Intensity Asymmetric Stereo (ZEST) uses visual prompting and monocular cues to achieve robust 3D perception without event-specific training, outperforming existing methods.

LLM-based agents are vulnerable to diverse backdoor attacks that manipulate their reasoning and outputs, highlighting the urgent need for targeted defenses.

VLMimic: Vision-Language Models enable robots to master intricate actions using only a few human video demonstrations, surpassing existing methods by a significant margin.

Visual Autoregressive Modeling (VAR) revolutionizes image generation by using a coarse-to-fine ’next-scale prediction’, outperforming diffusion models and exhibiting scaling laws similar to LLMs.

VideoTetris: a novel framework enabling compositional text-to-video generation by precisely following complex textual semantics through spatio-temporal compositional diffusion, achieving impressive qu…

Training-free video motion control achieved via novel Motion Feature (MOFT) extraction from existing video diffusion models, offering architecture-agnostic insights and high performance.

ConBench: Unveiling Inconsistency in Large Vision-Language Models

EVE, a groundbreaking encoder-free vision-language model, rivals encoder-based counterparts using a fraction of the data and resources, demonstrating efficient, transparent training for pure decoder-o…

This work systematically investigates the approximation properties of Transformer networks for sequence modeling, revealing the distinct roles of key components (self-attention, positional encoding, f…

U-DiT: Revolutionizing diffusion transformers with a U-Net design and token downsampling for superior image generation and drastically reduced computation cost.

VL-SAM: Training-free open-ended object detection & segmentation using attention maps as prompts, surpassing previous methods on LVIS and CODA datasets.

Hierarchical Representation Learning Framework (HRLF) significantly improves Multimodal Sentiment Analysis (MSA) accuracy by effectively addressing incomplete data through fine-grained representation …

A new dual-learning model resolves the paradox of perceptual learning, showing how task-based and feature-based learning interact to produce both specific and transferable improvements in sensory perc…

Leveraging data heterogeneity, this study reveals that standard SGD implicitly learns invariant features across multiple environments, achieving robust generalization without explicit regularization.

RaTSG network tackles Temporal Sentence Grounding with Relevance Feedback (TSG-RF) by discerning query relevance at multiple granularities before selectively grounding segments.

This research introduces a template-free articulated Gaussian splatting method for real-time dynamic view synthesis, automatically discovering object skeletons from videos to enable reposing.

Shortcut back-propagation and an evolutionary training framework conquer gradient vanishing in spiking neural networks, drastically improving training and achieving state-of-the-art accuracy.

StreamFlow accelerates video optical flow estimation by 44% via a streamlined in-batch multi-frame pipeline and innovative spatiotemporal modeling, achieving state-of-the-art results.

Researchers achieve minimax optimal sample complexity bounds for distributional temporal difference learning, enhancing reinforcement learning algorithm efficiency.

Spiking Experts Mixture Mechanism (SEMM) boosts Spiking Transformers by integrating Mixture-of-Experts for efficient, sparse conditional computation, achieving significant performance improvements on …