MATable: Multi-Agent Table Understanding over Image and Structured Representations

• 1、School of Electronic Engineering, Beijing University of Posts and Telecommunications, Beijing,100876

Abstract：Understanding tables presented as images is an emerging challenge in table understanding, commonly encountered in reports, dashboards, academic papers, and scanned documents. Compared to text-based tables, image tables exhibit diverse visual layouts, implicit structural cues, and heterogeneous representations, posing difficulties for reliable information extraction and structural reasoning. To address these challenges, we propose MATable, a unified multi-agent multimodal table understanding framework. MATable first transforms table images into task-relevant structured representations through a query-guided pipeline, decoupling visual perception from symbolic reasoning. It then employs a demand-driven dual-stream reasoning strategy: a lightweight language model serves as the cognitive core for comprehensive table analysis, while a visual perceptor is activated only when the structured evidence is insufficient, ensuring both efficiency and robustness. Coordinated via a multi-agent system, MATable achieves interpretable, stable, and high-accuracy table understanding. Extensive experiments on multimodal table understanding benchmarks demonstrate that MATable significantly outperforms both general-purpose and table-specialized models across reasoning-intensive and structure-focused tasks, validating the effectiveness of its modular and adaptive design.

keywords： Artificial Intelligence Multimodal Table Understanding Multimodal Question Answering Multi-Agent System

点击查看论文中文信息

点击收起