Instagram Feed Ranking: Splitting Baseline and Incremental Engagement
A new arXiv paper proposes separating baseline from incremental engagement predictions to debias Instagram's feed ranker.
Research Digest
Preprints and papers, rewritten in plain English.
RAG's Retrieval, Not Reading, Limits Long-Context QA
A new arXiv paper argues that retrieval-augmented generation's similarity-driven, chunked pipeline—not the model's reading ability—is the main bottleneck for evidence-based long-context question answering.
ArtSeek: A Multimodal Framework for Deep Artwork Understanding
A new AI framework combines multimodal language models with retrieval to analyze artworks using visual and historical context.
AutoSim: A Self-Improving User Simulator for Conversational Search
A new arXiv paper introduces AutoSim, a user simulator designed to make large-scale conversational search evaluation more realistic and trustworthy.
Multimodal Reranking as a System One Decision
A new arXiv paper argues that multimodal reranking can be reframed as a general System One decision task rather than a dedicated retrieval model.
ICIBS: A New Approach to Ice Boundary Segmentation in Radar Data
A new method aims to improve how ground-penetrating radar images are segmented to better measure glacier ice thickness and volume.
Test-Time Adaptation Proposed for Point Cloud Denoising Under Unseen Noise
A new approach aims to keep point cloud denoising models accurate when test noise differs from training noise.
Decoupling Action from Egocentric Observation for World Simulation
A new approach separates actions from egocentric views to enable simulation in novel visual contexts.
PRAISE method grounds person re-ID geometry with query-conditioned subspaces
A new approach uses attribute-based query-conditioned subspaces to improve person re-identification under viewpoint and illumination changes.
Gaze-Aware Model Disambiguates Intent for Hand-Object Forecasting
A new arXiv paper proposes using gaze to disambiguate user intent when forecasting hand-object interactions in egocentric video.
New Platform Detects AI-Generated and Manipulated Media Across Formats
A proposed multimodal system aims to identify synthetic faces, forged videos, AI images, fake audio, and edited photos in one platform.
Unsupervised Hyperspectral Segmentation via Scoring Rules and Mumford-Shah
A new arXiv paper connects proper scoring rules and edge biases to Mumford-Shah models for robust pixel-level segmentation without labeled data.
Unified Audio-Visual Editing Model Aims to Keep Human Videos in Sync
A new arXiv paper argues that replacing separate audio and video pipelines with a unified end-to-end model could improve alignment of speech with lip movements, gestures, and expressions.
AdaptMIL aims to make pathology AI robust across hospitals
A new method targets the distribution shift that limits pathology foundation models in real-world clinical settings.
Efficient SAM3 Adaptation for Multi-Class Segmentation
A new arXiv paper proposes an efficient method to adapt SAM3 for multi-class semantic segmentation using dense competitive representations.
Full-Context Memory for Stable Long Video Generation
A new arXiv paper proposes retaining full historical context in long-horizon video generation, avoiding information loss from compression.
Digital Twin Generates Training Data for SMC Strip Inspection
A process-informed photorealistic digital twin could overcome the lack of labeled images for inline quality control in Sheet Molding Compound manufacturing.
Auditing Concept Bottleneck Models: Where Does CUB Accuracy Come From?
A new arXiv paper examines whether Concept Bottleneck Models' accuracy on CUB-200-2011 truly stems from their interpretable concepts or from other model components.