BaFCo benchmark exposes MLLMs' struggles with Bangla forms

Document comprehension remains a tough challenge for Multimodal Large Language Models (MLLMs), especially for low-resource languages like Bangla. To address the lack of annotated data, researchers from multiple institutions introduced BaFCo, a benchmark dataset for Bangla form comprehension. BaFCo curates 200 multi-page complex government forms from sectors like agriculture, education, banking, and land management. The dataset features a fine-grained annotation schema with 26 entity types, plus a coarse set of 5 types, enabling detailed evaluation of Document Layout Analysis (DLA) and Key Information Extraction (KIE) tasks.

The team evaluated the latest MLLMs from the ChatGPT, Gemini, Claude, Qwen, and Kimi series using zero-shot and chain-of-thought prompts under both low and high reasoning setups. Results exposed clear limitations: all models struggled to accurately localize highly granular form entities in Bangla. The work, accepted at ECCV 2026, provides a crucial benchmark for advancing multilingual AI. The dataset and code are publicly available, offering a foundation for future research in low-resource language document understanding.