Mamba-based AGGR Network for Undersampled MRI Reconstruction
A new architecture combines Mamba with attention-guided gated residual refinement to recover high-quality images from partial k-space data, potentially shortening scan times.
Research Digest
Preprints and papers, rewritten in plain English.
Deep Learning Assesses Road Lane Quality from Dashcam Video
A new arXiv study proposes using deep learning to automatically evaluate lane marking condition from egocentric dashcam footage, potentially replacing manual inspection.
VisualEvolve lets vision-language models refine their own visual tools
A new framework addresses the limits of fixed visual operations in VLMs by enabling self-evolution of visual reasoning skills.
Brain2Vec: A self-supervised foundation model for brain MRI
A new 3D foundation model aims to make deep learning on structural brain MRI more efficient by learning from unlabeled data.
Task-Driven Camera Design via Differentiable Coded Apertures and 3D Gaussians
A new method jointly optimizes camera aperture and perception task performance using differentiable 3D scene representations, moving beyond hand-crafted camera design.
PRIMFLOW: Few-Step Distillation via Trajectory Primitives
A new arXiv paper proposes learning trajectory primitives to improve few-step distillation by capturing global updates from dense local supervision.
SenseAlign uses language to keep vision and touch models robust at test time
A new tactile-visual-language model aligns vision and touch at test time to handle real-world shifts in lighting, contact, and hardware.
GazeLens uses gaze cues to cut video tokens for egocentric QA
A new method for small vision-language models reduces visual tokens in egocentric video by following the wearer's gaze.
CoGaussian: Collaborative 3D Scene Understanding with Open-Vocabulary Gaussians
A new method lets multiple agents share open-vocabulary 3D scene understanding beyond their individual fields of view, potentially improving transportation safety and efficiency.
Residual-Space Adaptation Enables Continual Concept Learning in Diffusion Models
A new approach lets text-to-image models learn new user concepts sequentially without retraining from scratch or forgetting earlier ones.
ScreenCalib: Using a Programmable Display for LiDAR-Camera Calibration
A new method called ScreenCalib could simplify LiDAR-camera extrinsic calibration by using an ordinary display instead of specialised targets.
Hidden-State Iterations in Autoregressive LMs: A Theoretical Look
A new arXiv note examines the promise and pitfalls of letting language models think in latent space without emitting tokens.
PACE: A Training-Free Method for Timing Responses in Streaming Video
A new arXiv paper proposes PACE, a training-free framework that helps streaming video models decide when to answer, not just what to say.
MindView Benchmark Targets Multi-View Spatial Intelligence
A new benchmark, MindView, aims to evaluate how well models generate spatially consistent views from partial multi-view observations using language guidance.
Bayesian method tracks scene changes from streaming images
A new algorithm updates change beliefs in real time as images arrive, using generalized Bayesian evidence and 3D Gaussian splatting edits.
Detecting Copyright Leakage in AI-Generated Videos
A new arXiv paper tackles the growing risk that video generation models will reproduce copyrighted visual content without permission.
AEON Unifies Video and 3D World Models for Consistent Simulation
A new arXiv paper proposes a single framework that bridges the gap between realistic video generation and 3D-consistent world modeling.
LoWAgent benchmark tests web agents in low-resource languages
A new benchmark, LoWAgent, aims to fill the gap in evaluating web agents on multilingual and low-resource language websites.