BaFCo: A Document Understanding Benchmark for Complex Bangla Form Comprehension

Document comprehension is a crucial task for Multimodal Large Language Models (MLLMs), yet it remains heavily skewed toward high-resource languages like English. Low-resource languages such as Bangla lack high-quality annotated data, which limits the real-world deployment of document understanding systems in public administration, banking, and education. To address this gap, the authors investigate how well current flagship MLLMs perform on Document Layout Analysis (DLA) and Key Information Extraction (KIE) using complex, multi-page Bangladeshi government forms.

The paper introduces BaFCo, a curated benchmark consisting of 200 multi-page complex Bangladeshi government forms spanning 15 administrative domains. To capture structural and contextual complexity, the authors define a fine-grained annotation schema comprising 26 types of form entities and a coarse set of 5 types, alongside labeled relationships between related fields. The dataset covers 16,382 entities and 8,771 relationships across 316 pages, categorized into easy, medium, and hard difficulty levels based on layout density, tables, and selection elements.

The authors evaluate several leading proprietary and open-source MLLMs—including models from the ChatGPT, Gemini, Claude, Qwen, and Kimi series—using zero-shot and Chain-of-Thought prompting under low and high reasoning setups. The results show that current flagship MLLMs face significant challenges when processing Bangla forms. While some models handle broader layout elements reasonably well, their accuracy drops sharply when tasked with localizing highly granular form entities and extracting fine-grained key-value pairs.

This work provides the first rigorous benchmark and evaluation framework for Bangla document understanding, establishing a critical foundation for advancing low-resource multimodal AI. By highlighting the limitations of current MLLMs on complex, real-world government forms, BaFCo paves the way for more inclusive, globally capable document intelligence systems.