Vision
The Internet revolutionized the virtual world. Strong visual AI is the missing piece for the physical one.
We study the effectiveness of grounded reasoning strategies. We introduce BrickVision, a challenging benchmark designed to measure a model's capability to reason with undescribable visual references.
Read the postA counting benchmark that tests whether a model actually counts the objects, or reports a substitute. Built in partnership with Mercor.
Read the post250 expert-level questions about three-dimensional scientific artifacts, and what happens when you give a model the tools to inspect them. Built with UniPat AI.
Read the postVisual thinking that stays grounded. Glimpse looks closely, questions its first impression, and returns to the visual evidence before it answers.
Read the postA benchmark measuring the implicit visual reasoning and intuitive understanding we draw from a scene. Built in partnership with Scale AI.
Read the post