What's this about?
Multimodal models increasingly draw sketches, annotations, and intermediate images while reasoning. This benchmark tests whether they actually rely on those visual states or just decorate their answers with them — with coverage designed so tasks can't be solved from text alone.
What exactly is an "intermediate visual state"?
It's the model's visual scratch work — cropping an image to zoom in, drawing boxes, generating a diagram mid-solution — analogous to a human sketching on paper while thinking. The open question: is it load-bearing reasoning or theater? Prior benchmarks couldn't tell because their tasks leaked into text.
Why it matters
"Thinking with images" is a headline feature of current frontier models; a benchmark that can catch fake visual reasoning keeps that claim honest (22 upvotes, top paper of the day).