PLaCA: Learning Patch-Level Future Latent Change for Robot Actions
A new VLA approach predicts how local image patches will change, helping robots anticipate environment dynamics during manipulation.
Research Digest
Preprints and papers, rewritten in plain English.
Pairwise Comparison Framework Boosts Data Efficiency in Deep Neural Networks
A new arXiv preprint suggests reframing regression as pairwise comparisons could help deep networks learn from small samples.
StreamEval Benchmark Aims to Evaluate Interactive Streaming Video Models
A new benchmark proposes a unified way to evaluate streaming video models that support real-time interaction.
UniWALM: A Unified Model for Language, Video, and Action in Robots
A new arXiv preprint proposes a single model that learns video, action, and language together for embodied robot control.
Multi-Agent Video World Models Need Consistent Streams
A new arXiv paper highlights the challenge of generating agent-specific video streams that stay consistent in time and across agents.
Deep learning aligns historical maps to profile urban change
A new arXiv study uses deep learning to extract consistent, fine-grained urban transformation data from historical maps, filling a gap before satellite observation.
Reinforcement learning unifies metal artifact reduction and accelerated MRI
A new arXiv preprint proposes a metal-aware reinforcement learning approach that treats artifact reduction and accelerated MRI as a single problem.
Scene-R1 Aims to Make 3D Scene Reasoning Transparent for LLMs
A new approach, Scene-R1, uses video-grounded data with lightweight annotations to give large language models a visible reasoning path for 3D scene understanding.
TrajTok: Trajectory Tokens for Efficient Video Understanding
A new tokenizer replaces patch-based tokens with trajectory-based ones to cut redundancy and improve video model scalability.
Can LLM dialogues curb belief in unfolding conspiracy theories?
A new arXiv study tests whether conversational AI can reduce belief in conspiracy theories as they emerge after major events.
Frequency-Domain Method Aims to Make Audio-Visual Segmentation More Robust
A new arXiv preprint proposes a frequency-domain decomposition-and-recomposition approach to make audio-visual segmentation more robust.
WildPPG Dataset Targets Real-World Wearable Heart-Rate Accuracy
A new open dataset of long, continuous PPG recordings aims to improve heart-rate monitoring in wearable devices under real-world conditions.
Self-Supervised Pre-Training Tackles Sparse X-Ray Angiogram Labels
A new arXiv preprint proposes vascular anatomy-aware self-supervised pre-training to reduce reliance on annotated X-ray angiograms.
SpectralTrain framework targets hyperspectral image classification cost
A new framework aims to overcome the heavy data and training demands that limit practical deep learning for hyperspectral image classification.
Joint Open-Vocabulary Segmentation for Point Clouds and Panoramas
A new framework tackles semantic segmentation across 3D point clouds and panoramic images without relying on fixed label sets.
Anchoring attention for coherent inbetweening from sparse keyframes
A new arXiv paper proposes anchoring and rescaling attention to keep generative inbetweening semantically coherent when keyframes are sparse and motion large.
District Magnitude Shapes Proportionality Under STV, Study Finds
A new arXiv preprint examines how the number of seats per district affects partisan proportionality in multimember districts using Single Transferable Vote.
Study: Granting Personhood to Bears Safer Than to AI
A new arXiv paper contends that extending personhood to bears is less risky than doing so for artificial intelligence systems.