RL Sharpens VLMs' Attention but Not Their Perception
A new study finds that reinforcement learning improves vision-language agents by teaching them where to look, not how to see.
Research Digest
Preprints and papers, rewritten in plain English.
SituHOI dataset targets full-body human-object interaction in real scenes
A new arXiv preprint introduces SituHOI, a dataset designed to jointly capture whole-body motion, manipulated objects, and scene geometry for human-object interaction research.
Time-JEPA: Grounding Video World Models in Physical Time Series
A new arXiv paper proposes grounding video world models with physical time-series data to improve prediction of environment evolution.
Cutting Token Bloat in Visual Document Retrieval
A new arXiv paper targets the inefficiency of token-heavy vision-language representations in OCR-free document retrieval.
FrameScout selects query-relevant frames for long video understanding
A new method addresses the token bottleneck in VideoLLMs by scouting only the frames that matter for a given query.
Sparse-frame VideoQA answers flip when sampling timestamps shift
A new arXiv preprint finds that sparse-frame video-language models are unstable to small changes in which frames they sample, and raw agreement between runs is not enough to detect or fix the errors.
SPELUNKA Framework Uncovers Hidden Biases in Vision-Language Models
A new framework aims to identify unanticipated spurious associations in vision-language models, going beyond the removal of known biases.
Synthetic Hajj Crowd Images Could Ease Data Scarcity for Counting Models
A new method generates synthetic Hajj crowd images with reference-guided instruction, targeting the shortage of annotated domain-specific data.
New Decoder Adds Confidence and Evidence to Document Extraction
VERDICT, a decoder for multimodal document extraction, targets the lagging areas of confidence and traceable evidence.
Geo-vision: classifying vehicle-part damage with geometry and topology
A new arXiv preprint proposes a fixed representation that fuses appearance, geometry, orientation, and topological features for vehicle-part damage classification.
IMPULSE: Push objects first to improve V-JEPA 2 predictions
A new method proposes physically probing objects before predicting their dynamics in a latent world model's feature space.
Temporal Domain Shift Hampers AI Phytoplankton Event Detection
A new arXiv study highlights the gap between image-level AI accuracy and reliable phytoplankton event detection under changing environmental conditions.
SPIRIT: A Sparse, Physics-Informed Approach to Sharper SAR Imaging
A new inverse-rendering method aims to break the speed-resolution trade-off in synthetic aperture radar.
Adaptive AI aims to keep fetal brain MRI biometry reliable
A new adaptive AI system, BioTTA, targets the domain-shift problem that makes fetal brain MRI measurements unreliable across different clinical settings.
Adding Inner Structure to AI-Generated 3D Shapes
A new method constructs interior geometry for AI-generated 3D models, aiming to make them functional beyond just looking good.
Fine-Tuning One Skill in Multimodal AI Can Damage Others
A new paper highlights a post-training hazard unique to unified multimodal models: improving one capability may degrade another.
Adapting Vision-Language Models Without Labels When Vocabulary Is Incomplete
A new approach targets unsupervised adaptation of vision-language models when the model's label vocabulary does not cover the target domain.
STLQ: A New Quantization Method for Approximate Nearest Neighbor Search
An arXiv paper proposes STLQ, a spatio-topological linkage quantization approach for approximate nearest neighbor search.