Preprint finds rotated views improve AI reading of large grid diagrams

Researchers report that reading a rotated copy of a diagram with crop checks improved an AI model's exact reading of large grids to 86/96, compared with 57/96 when it simply repeated the upright reading. The preprint separates errors in reading diagrams from errors in counting their structures. With code and an exact transcription, the primary model counted correctly even on 10x10 grids, making visual reading the bottleneck. In chain-of-thought tests, most Gemini models read small drawings correctly but then failed at counting; answer size mattered more than the counting problem's complexity class.