VisualNeedle Benchmark Exposes MLLM Visual Search Shortcuts

Researchers have introduced VisualNeedle, a benchmark designed to test active visual search in multimodal large language models (MLLMs). The benchmark targets scenes where critical evidence is spatially constrained and information-dense, challenging models to rely on genuine visual processing rather than shortcuts. Prior studies identified three common shortcuts: linguistic priors and lexical cues in questions, coarse global semantics from visual encoders bypassing fine-grained details, and corruption of intermediate images barely affecting answers in some benchmarks. VisualNeedle aims to address these issues by requiring fine-grained perception beyond high resolution or large question pools. The work is published on arXiv under identifier 2605.26380.