Online image-reading services may shrink or reject large photos before analyzing them. A tiny car, label, or sign can become too small to recognize, so a wrong answer may reflect lost image detail rather than poor reasoning.
The researchers tested several services with printed codes and questions about small objects in real photos. They compared original images with versions split into smaller sections, and with a two-step approach: the model identifies a region, then receives a closer crop.
Splitting images raised accuracy by 11.1 to 14.5 percentage points for Claude and Gemini. Targeted zoom often achieved similar results using fewer image tokens, the units services use to process image input. This suggests software should zoom when details matter and treat uncertain answers as a cue to inspect more closely.
