Adversarial procurement in blockchains: a new arXiv analysis
A new preprint examines how blockchain protocols that outsource computational work could be vulnerable to adversarial procurement.
Research Digest
Preprints and papers, rewritten in plain English.
Mixed Reality Platform CAESAR Targets Cognitive Attacks in Search-and-Rescue
CAESAR applies mixed reality's embodied, physically-coupled training to cognitive-attack testing in safety-critical collaboration.
GTSM: Reconstructing 4D Thermal Fields with Gaussian Splatting
A new arXiv preprint proposes a Gaussian-splatting method with thermal state-space dynamics for reconstructing and predicting spatio-temporal temperature fields.
CodecMotion-OAD: Efficient Online Action Detection via Compressed Motion
A new method uses motion information already present in compressed video to speed up online action detection, avoiding expensive optical-flow computation.
Controllable Object Insertion for Detection Data Augmentation
A new arXiv paper proposes controllable object insertion to augment detection datasets while preserving category counts, annotations, and source content.
Benchmarking Privacy in Multi-User LLM Agent Systems
A new arXiv benchmark targets privacy risks in shared LLM agents, focusing on contextual integrity rather than only external output leakage.
Event Rate as a Feature in Graph Neural Networks
A new training strategy for event-based graph neural networks reframes event rate as a useful signal rather than a nuisance to be discarded.
MatchCityLoc Uses Surface Normals to Localize Drones in Cities
A new visual localization method matches surface normals from drone cameras to LoD city models to estimate pose where GNSS fails.
LVMem: Image-Augmented Memory for Long-Horizon Question Answering
A new arXiv paper proposes an image-augmented memory approach to help embodied agents answer questions about objects, events, and locations across long observation histories.
Infrared Polarimetric Fusion Targets Blur in Vehicle Detection
A new fusion approach for long-wave infrared polarimetric imagery aims to reduce polarization noise and better combine complementary information for vehicle detection.
Occlusion-Aware Densification for 3D Gaussian Splatting
A new approach to adaptive density control in 3D Gaussian Splatting focuses on occluded contributions rather than simple view counts.
Multimodal model links lung histology and spatial transcriptomics
A new multimodal large language model learns a shared representation from lung tissue images and spatial gene expression data.
Body-Driven Physics for Multi-Object Human-Object Interaction Synthesis
A new arXiv paper introduces BUMP, a method that treats human motion and object manipulation as distinct but unified physical causes to improve multi-object interaction synthesis.
Implicit Visual Thinking Aims to Improve Fine-Grained VLM Reasoning
A new arXiv preprint proposes a method to make vision-language models look more carefully before reasoning, targeting small image details.
LoRASpace: A Shared Substrate for Composing Multiple LoRAs
A new arXiv preprint proposes a pool-wide shared substrate to make combining multiple LoRA adapters easier.
TokenFlow: A Token-Based Flow Model for 3D Point Clouds
A new arXiv paper proposes replacing single-vector latent encodings with token-based rectified flow matching for finer 3D point cloud generation.
FocusOVD refines features for open-vocabulary small-object detection
A new method targets the persistent failure of open-vocabulary detectors on small objects by refining geometric and contextual features.
Saliency-Guided Augmentation for Low-Resource Handwriting Recognition
A new arXiv paper proposes constraining data augmentation to non-salient image regions so that critical character features are not erased.