VisualNeedle: Benchmarking Active Visual Search in Information-Dense Scenes
VisualNeedle is a new benchmark designed to test if multimodal models truly perform active visual search—using minute visual clues—instead of relying on linguistic shortcuts. The benchmark includes constraints ensuring answers cannot be found through just the question or a quick global view, forcing models to actively localize evidence in dense scenes. Results show current models struggle, confirming that active visual reasoning remains limited.