KoalaGS: Complexity-Aware Continuous LoD for Gaussian Splatting
A new method aims to keep 3D Gaussian Splatting scenes effective when compute budgets shrink, using a continuous level of detail approach.
Research Digest
Preprints and papers, rewritten in plain English.
Co-Generating 3D Motion and Video for Arbitrary Rigs
A new paper proposes linking video generation with 3D joint data so realistic motion can be reused to animate specific 3D assets.
LocaLeJEPA targets representation collapse in JEPA models
A new arXiv paper proposes LocaLeJEPA, which builds on LeJEPA to keep joint-embedding predictive architectures from collapsing by enforcing maximum entropy under geometric priors.
GALA: Geometry-Aware Language Model for Controllable Object Arrangement
A new arXiv preprint proposes a geometry-aware language model that reasons about support and containment constraints for realistic 3D scene creation.
Learned Compression for 3D Gaussian Splatting Using Video Codecs
A new paper describes a learned compression method that maps 3D Gaussian Splatting attributes to 2D for standard video codecs.
New arXiv paper tests associative creativity in humans and multimodal LLMs
A new preprint proposes using difficult-to-name abstract visuals to compare associative creativity between humans and multimodal large language models.
GPT-5.1 Feedback on Handwritten Math: Exam Impact Studied
A new arXiv preprint tests whether AI-generated feedback on handwritten university math work improves later exam scores.
CARI4D Reconstructs 4D Human-Object Interactions from Single RGB Video
A new arXiv paper proposes a category-agnostic method to reconstruct 4D human-object interaction from a single RGB camera, potentially enabling richer understanding for robotics and gaming.
MICCAI Atlas Maps 28 Years of Medical Imaging Research
A new arXiv analysis examines how the MICCAI conference has shaped medical image computing through representation, collaboration, and inclusion.
How data-source choices skew urban park accessibility findings
A new arXiv preprint shows that the selection of data sources can substantially shape urban park accessibility analyses, complicating comparisons across studies.
Structured spatial observation helps agents edit infinite canvases
A new arXiv paper proposes CanvasObs, a structured spatial observation method that gives editing agents information beyond pixels to handle offscreen or covered targets.
LSGQuant: Quantizing One-Step Diffusion for Video Super-Resolution
A new quantization method targets the heavy model size and compute cost of one-step diffusion Transformers in real-world video super-resolution.
Study: AI Abstention Alters Advice Acceptance
New research examines how AI systems that sometimes withhold advice affect whether humans accept their recommendations.
Piecewise Sparse Attention Takes On the Diffusion Transformer Bottleneck
A revised preprint argues that PISA, a piecewise sparse attention pattern, is a wiser alternative to block sparse attention for speeding up diffusion transformers in video and image generation.
Thought-Level Progress Tracking for Psychosocial Goals
A new arXiv preprint examines how tracking progress at the level of thoughts could support psychosocial goals such as coping with anxiety and personal growth.
Digital Nudges Can Work Against Each Other in Layered Designs
New research shows that combining digital nudges may weaken their individual effects in sequential decision-making.
Study Probes Bias in LLM Learning-Style Inference from Student Dialogues
A new arXiv study evaluates GPT-4o and Gemini 2.0 Flash for bias and reliability when inferring student learning styles from dialogue.
Agentic Tooling Aims to Automate Video Editing Across Complexity Levels
A new arXiv paper proposes agentic tooling that unifies automated video editing across tasks of varying complexity and creativity.