Agentic Data Intelligence for General Tabular Modeling

Tabular data is one of the most common forms for organizing real-world information, supporting diverse tasks such as prediction, reasoning, querying, and generation. Despite rapid progress in tabular learning, most existing methods remain specialized for isolated problem formulations. Recent advances in large language models and agentic systems have shown increasing potential for general-purpose problem solving, offering an opportunity for unified tabular modeling. However, applying such models to tables remains challenging, as language models are often insensitive to numerical and structural information, while tabular tasks are highly heterogeneous in objectives, inputs, and evaluation protocols. In this work, we propose a general tabular agent for agentic data intelligence, {\sc TabAgent}, aiming to solve diverse table-centric tasks within a unified framework. The agent integrates rich tool-use interfaces, structured task-solving skills, and a reliable execution harness, allowing it to select appropriate actions and reasoning strategies for different tabular scenarios. Beyond standard table processing, we further endow the agent with specialized capabilities, including multimodal table understanding, table tabulation, and implicit prediction through problem decomposition. Experiments across diverse tabular tasks demonstrate that our method effectively handles heterogeneous task requirements and moves toward general tabular modeling.