How Robust is OCR-Reasoning? Evaluating VLM OCR-Reasoning Under Visual Perturbations

arXiv·medium signal

This paper shows that vision-language models' strong OCR-reasoning scores on text-rich images degrade meaningfully under visual perturbations, exposing a reliability gap. Authors Yuxing Cheng, Yuan Wu, and Yi Chang stress-test robustness rather than clean-input accuracy. A practical caution for anyone deploying VLM-based document or OCR pipelines in production.