Megadose The AI curve, compressed.

Know When to Zoom: How Small Is Too Small for Vision-Language Models?

· ArXiv · AI/CL/LG ·
Zooming recovered small-image details that services had shrunk away or rejected before the model could see them.

The study tested seven models from four vendors using printed codes and 601 questions about small objects in real photos. Splitting photos into up to nine tiles improved Claude and Gemini accuracy by 11.1–14.5 percentage points, while avoiding rejections lifted GPT-6.1 Sol from 50.7% to 83.7%. Letting models request a region to zoom into matched or approached tiling accuracy with about 40% of its image tokens. For Claude, tiling only after an uncertain reply came within 1.4 points of always tiling, using 62%–70% of the image tokens.

ArXiv · AI/CL/LG's note

score 5

Categories: Research