Vision-language models often read Latin text confidently and misread mixed-script Japanese. The gap shows up in specific, testable places.
Text in the wild
Vertical signage, stylized fonts, furigana and mixed scripts challenge both detection and transcription. Real scene text with human transcription is the reference.
Cultural grounding
Dishes, garments, festivals and objects need local names and context. A model that calls a regional dish 'noodle soup' has not failed on vision, it has failed on knowledge.
Documents and comics
Right-to-left, panel-based reading order in manga and dense table layouts in business documents require layout-aware annotations.
Evaluation
Build a held-out evaluation with native-written questions and rubrics. Do not reuse images from your training set.
General information, not legal advice.