Three recent arXiv papers converge on the same question: how useful are vision-language models (VLMs) for OCR, especially on historical documents? All three acknowledge that VLMs have posted impressive or strong results, but they diverge on whether that translates into practical deployment.
One paper argues that large VLMs are too computationally expensive and too dependent on large-scale pretraining for historical text recognition, and instead proposes adapting the traditional PP-OCRv6 pipeline as a "free lunch." A second paper pushes back on that pessimism, presenting HunyuanOCR-1.5, a lightweight OCR-specialized VLM designed to be faster and better while unifying document parsing, text spotting, information extraction, and related tasks.
A third paper adds a different caveat: on historical Uruguayan documents, VLMs can achieve low character error rates yet still hallucinate. That means standard accuracy metrics may not capture the reliability needed for archival digitization. Together, the papers suggest that the choice between traditional OCR and VLMs is not settled, and that cost, speed, and hallucination all need to be weighed.