Visual Geometry Models / 2026
VGGT-Prime
Compute-adaptive attention for efficient visual geometry. VGGT-Prime routes attention heads between different computation modes to reduce the cost of reconstruction while maintaining competitive quality.
Personal research lab
Exploring how
machines see.
I'm Abteen Arab. My research interests are in vision, particularly the interplay between vision and other modalities.
01 / On the workbench
Selected work
Computational Medicine / 2026
AC-MIL
An anatomy-aware framework for predicting kidney function from whole-slide images. Separate glomerular and non-glomerular tissue streams are combined through bidirectional gating to predict estimated glomerular filtration rate (eGFR).
02 / Research directions
Research areas
Visual Geometry Models
I'm interested in how images reveal 3D geometry, from depth and scene structure to relationships across views. I want to understand how information from other modalities can complement that visual understanding.
World Geometry Models
I'm interested in how visual observations become representations of a world or environment that changes over time. I want to explore how vision and other modalities contribute to those representations.
Geometry Action Models
I'm interested in how visual and spatial understanding inform actions and respond to interaction. I want to explore the connections between what a system sees, how it moves, and the other signals it receives.
Computational Medicine
I'm interested in combining visual learning with clinical measurements and complementary medical data. I want to understand what images and other forms of medical data reveal when considered together.