Khondo: A Multimodal Benchmark for Document Packet Splitting of Bangla Forms

Khondo introduces what the authors describe as the first benchmark for splitting document packets built from Bangladeshi government forms. The bilingual Bangla-English, vision-native dataset covers 14 administrative domains and five concatenation schemes, with annotations for document boundaries, domain types, and page order. Zero-shot multimodal large language models can cluster pages by source document reasonably well, but struggle to reconstruct the original order after shuffling. Controlled analyses indicate that explicit ordering instructions are necessary but insufficient, while English packets are ordered more reliably than Bangla packets.