Tag
This paper reveals that Vision-Language Models often rewrite rather than faithfully transcribe text when encountering perturbations like typos or visual artifacts, introducing the FaithC4 benchmark to evaluate this behavior across multiple models and languages.