ReTrace: Visual Backtracking to Cut Multimodal LLM Token Overhead
A new arXiv paper proposes ReTrace, a visual backtracking method that may reduce the computational cost of multimodal large language models without relying solely on token pruning.
Research Digest
Preprints and papers, rewritten in plain English.
Cross-polarization pairing improves few-shot rock thin-section classification
A new method leverages paired cross-polarized images to reduce the need for costly petrographic labeling in few-shot classification.
UniAVC: A Unified Model for Audio-Video Generation and Editing
A single framework aims to handle joint synthesis, cross-modal generation, and editing of audio and video while reducing reliance on paired data.
SwitchSplat Models Discrete Object States in Casual 3D Captures
A new method extends 3D Gaussian Splatting to handle scenes where objects change state during casual capture.
Image Editing with Reusable Layers via Edit-Prior Distillation
A new arXiv paper proposes distilling compositional edit priors to turn image edits into reusable layers, improving precision and reuse.
Propagating Uncertainty in 3D Morphable Models
A new arXiv paper studies how parameter-space variance in 3D Morphable Models maps to per-vertex variance, moving beyond visualization.
Active 3D Exploration Could Sharpen Vision-Language Spatial Reasoning
A new arXiv paper argues that passive, view-centric perception limits VLMs' 3D reasoning and proposes active exploration over 3D worlds instead.
MOMO: Modeling Human Motion from Moving Sounds
A new arXiv preprint proposes MOMO, a model that links the evolving spatial relationship between sound sources and listeners to human motion responses.
Spectral-Aware Diffusion Method Targets Multispectral UAV Tracking
A new arXiv preprint proposes a spectral-aware enhancement for diffusion-based multi-object tracking in multispectral UAV imagery, though the abstract is truncated.
Geometry-Aware Interpolation Targets Spatial Variation in Image Recovery
A new arXiv preprint proposes a geometry-aware interpolation strategy to improve reconstruction in inverse problems where spatial variations and complex structures hinder accuracy.
Knowledge Diversion: Pre-training Image Models for Downstream Tasks
A proposed 'knowledge diversion' paradigm rethinks pre-training for decomposable image generation models by factoring in downstream requirements.
Beyond geometric closure: measuring mechanical consistency in visual dynamics
A new framework assesses whether predicted motion obeys dynamics, not just visual accuracy.
Multi-exit residual ViT targets small-data image classification
A proposed architecture combines multi-exit residual connections with dynamic self-ensembling to help Vision Transformers learn from small datasets.
Compact Reusable Memory Lets MLLMs Query Long Videos Once
A new arXiv paper proposes a compact reusable memory that lets multimodal models answer many queries about long videos without reprocessing them.
New Dataset Aims to Fix Video Super-Resolution Training Gaps
A broad-coverage training dataset addresses semantic imbalance and diversity gaps in open-domain video super-resolution.
Scene-Adaptive VLA: Dynamic Layer Routing for Efficient Autonomous Driving
A new method dynamically adjusts computation in vision-language-action models to reduce the onboard resource strain of autonomous driving.
PACE-Mem: Better Evidence Retrieval for Multimodal Agents
A new arXiv preprint proposes a prefix-aware approach to retrieving complementary evidence from long-term memory in multimodal agents.
RiSE: A New Approach to Localize Text-Centric Image Forgeries
A proposed method called RiSE targets the challenge of generalizable localization of AI-edited text in images, which is critical for financial and legal records.