TabAgent: A Multi-Agent Table Extraction Framework for Unstructured Documents

With the increasing amount of unstructured documents in various domains, extracting structured data from such sources has become critical for efficient data management and advanced analytics. However, existing methods for structured table extraction face challenges: (1) the complexity and implicitness of semantic context and patterns in unstructured documents hinder accurate information extraction, and (2) limited adaptability to evolving user intents and requirements for strict schema alignment and structural constraints. To address these limitations, we propose TabAgent, a novel multi-agent collaborative framework for structured table extraction from unstructured documents. TabAgent integrates four specialized agents, Schema Agent, Extraction Agent, Semantic Agent, and Validation Agent with a shared memory repository to iteratively refine extraction results. By leveraging collaborative reasoning and iterative self-correcting loops, TabAgent enables accurate, adaptive, and robust table extraction across diverse document domains and user instructions. Extensive experiments on four datasets demonstrate that TabAgent consistently outperforms several baselines, including pure LLM extractors and LLM-based systems, highlighting the effectiveness of this collaborative framework. Our work represents one of the first multi-agent frameworks for structured table extraction, offering an applicable solution for real-world applications.