Visual Search

Visual search studies how an observer finds a target or useful evidence among competing visual information, from human eye movements to image retrieval and embodied agents. Current AI methods give multimodal models active tools such as cropping, zooming, parallel tile inspection, and Bayesian or uncertainty-guided region selection, while other approaches train search policies from trajectories, process rewards, or human annotations. Evaluation spans psychophysical tasks such as feature and conjunction search, GUI grounding on ScreenSpot-Pro, tool-using tasks such as VisualNeedle, and image-retrieval benchmarks. Models can improve accuracy and search efficiency, but reliable multi-hop evidence use remains difficult, and matching human outcomes does not necessarily mean matching human search processes.