New DISCO benchmark reveals OCR and vision-language models have distinct strengths in document intelligence tasks

DISCO evaluates OCR pipelines and VLMs separately on text extraction and question answering across diverse document types including handwritten text, multilingual scripts, medical forms, and multi-page documents

OCR pipelines outperform VLMs on handwriting recognition and long documents where explicit text grounding supports text-heavy reasoning

Vision-language models show better performance on multilingual text and visually rich document layouts

Performance varies substantially across different tasks and document characteristics, highlighting the need for complexity-aware model selection