When Models Misread Ink: Auditing VLMs for Arabic Calligraphy Recognition
Researchers audited twelve Vision-Language Models (VLMs) on Arabic calligraphy recognition using numeric, visual-cue, and Diwani-first prompts. The main finding is that models suffer from default-label collapse, consistently misclassifying styles as Thuluth. Prompting helps unevenly; while visual cues can improve accuracy for some models, they often relocate the default error rather than fixing it. This highlights a critical vulnerability in current VLM robustness for fine-grained cultural tasks.