BlindGrid: Diagnosing Efficient Visual Evidence Acquisition and Binding under Local Observation

When people answer a question about a cluttered scene, they explore it piece by piece, gathering evidence before committing to an answer. Multimodal large language models (MLLMs), by contrast, are usually evaluated after receiving the full image at once, which hides the cost of acquiring that visual evidence in the first place. We introduce BlindGrid, a blind-grid diagnostic protocol that reframes multimodal reasoning as efficient visual evidence acquisition and binding under local observation. An image is divided into patches, and an MLLM observes one patch at a time while deciding whether to move, remember partial evidence, or stop and answer. This setting isolates two coupled abilities: actively acquiring relevant evidence and binding partial observations into a stable answer. Across 1500 attempts on two high-resolution artworks, we show that engineered prompts improve answer rate and reduce exploration steps compared with a minimal baseline, yet final accuracy remains bottlenecked by evidence use rather than prompt format alone. We further identify a visual-granularity trade-off: smaller local views can enlarge the search space and raise cross-patch integration cost, outweighing their sharper detail. Refined questions, judge-based audits, and trajectory analysis reveal evidence-use failures rather than simple recognition errors. MLLMs often reach relevant patches but fail to count structures across views, bind local attributes, preserve object relations, or attach a queried object to the correct person. These results establish BlindGrid as an audit-friendly diagnostic benchmark for measuring how MLLMs allocate observation budget and use partial visual evidence efficiently.