Segmentation-Free Handwriting Recognition in Historical Handwritten Documents Using Large Vision-Language Models

The standard handwriting recognition pipeline segments a page into lines before recognising anything. Large vision-language models make it reasonable to ask whether that step is still needed.

We benchmark nine commercial and open-weight LVLMs across four scripts and three difficulty levels. Page-level input outperforms line segmentation on every handwritten script we tested, and at lower inference cost — for Cyrillic, Gemini 2.5 Pro’s character error rate drops from 22.6% to 3.9%. The exception is the most degraded historical collection (Bentham), where segmentation still helps.