HTAN: A Hierarchical Table-Aware Network for Complex Table Visual Understanding

This paper proposes HTAN, a hierarchical tableaware network grounded in multimodal large language models (MLLMs), for joint visual-textual understanding of complex tables. Our method enhances MLLMs with three key components: (1) hierarchical table encoding via multilevel (row/column/table) attention to preserve structural semantics, (2) robust table detection combining geometric clustering with OCR-free visual features, and (3) memoryefficient multimodal fusion through gradient-aware tensor decomposition. HTAN jointly optimizes visual and textual modalities within the MLLM framework, enabling seamless integration of layout-aware features.