Get Started
Home
Topics
Search
Library
Research questionHow can multimodal models interpret handwritten Bengali land records with unfamiliar base-16 fractions and answer structured legal questions?These authoritative land records combine handwriting, Bengali text, and specialized base-16 fractional glyphs that mainstream OCR and tokenizers do not support. Legal questions also require correctly linking information across structured fields rather than merely recognizing isolated text.
AI
Computer Vision
Evaluation & Benchmarks
Multimodal Models
Natural Language Processing
Research Paper
Latest papersRecent research connected to this question, newest first.KhatianDoc: A Human-Verified Benchmark Diagnosing Multimodal LLM Failure on Bengali Legal Land RecordsKhatianDoc provides evidence from 107 real RS Khatian records collected at the Vumi Office in Munshiganj, Bangladesh, with 1,634 legal question-answer pairs. Its four tasks cover symbol recognition, base-16-to-decimal conversion, field extraction, and legal question answering; human transcriptions were verified by a land-law practitioner. The reported results come from a fixed zero-shot evaluation of six open and closed multimodal models ranging from 8B to over 72B parameters, using a redacted image release.research paper · Sep 3, 2026
Related questions
How can we evaluate large language models’ understanding of Bangla idioms in a low-resource setting?How can multimodal models integrate evidence across deeply interleaved text and images?How can multimodal models rely on images or audio rather than language shortcuts?How can multimodal models reason about fine-grained interpersonal relationships from conversational and visual cues?