What's this about?
A benchmark testing whether vision systems can act like human vision — a closed loop where you look, form a hypothesis, and redirect your gaze — rather than judging a single snapshot.
What exactly is "active vision"?
Perception as a process: eyes move ~3 times a second, steered by what you currently believe. Most vision models get one image and answer; this exam scores the looking strategy itself (20↑).
Why is it worth noting?
Robots and GUI agents both need "where to look next," not just "what is this" — a capability current VLM benchmarks barely measure.