13 upvotes
VisSearch-Bench challenges agents with multi-step visual web hunts
What is VisSearch-Bench: Evaluating Long Horizon Visual Search in Multimodal Agents?
VisSearch-Bench is a benchmark that measures how multimodal agents perform long-horizon visual search across the web. Instead of stopping after spotting an initial object, the tasks require agents to stitch together clues from multiple images, objects and search steps.
Who it is for
Researchers building visual assistants can use it to expose reasoning gaps and improve long-range multimodal performance.