Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TableGPT2: A Large Multimodal Model with Tabular Data Integration

Aofeng Su, Aowen Wang, Chao Ye, Chen Zhou, Ga Zhang, Gang Chen, Guangcheng Zhu, Haobo Wang, Haokai Xu, Hao Chen, Haoze Li, Haoxuan Lan, Jiaming Tian, Jing Yuan, Junbo Zhao, Junlin Zhou, Kaizhe Shou, Liangyu Zha, Lin Long, Liyao Li, Pengzuo Wu, Qi Zhang, Qingyi Huang, Saisai Yang, Tao Zhang, Wentao Ye, Wufang Zhu, Xiaomeng Hu, Xijun Gu, Xinjie Sun, Xiang Li, Yuhang Yang, Zhiqing Xiao

TL;DR

TableGPT2 targets the core difficulty of integrating structured tabular data with large language models for business intelligence tasks. It introduces a dedicated Semantic Table Encoder with 2D bi-directional attention and a Q-Former-style adapter to capture table semantics and align with textual inputs, combined with continual pretraining, supervised fine-tuning, and an agent framework for production-style tool use. The model achieves substantial gains across a broad suite of table-centric benchmarks and introduces RealTabBench to reflect real-world irregular and ambiguous tables, demonstrating strong generalization alongside robust coding and reasoning in BI contexts. The work also outlines an enterprise-oriented agent workflow with RAG and a secure code sandbox, highlighting practical considerations for deployment and future directions toward multi-agent and domain-specific adaptations.

Abstract

The emergence of models like GPTs, Claude, LLaMA, and Qwen has reshaped AI applications, presenting vast new opportunities across industries. Yet, the integration of tabular data remains notably underdeveloped, despite its foundational role in numerous real-world domains. This gap is critical for three main reasons. First, database or data warehouse data integration is essential for advanced applications; second, the vast and largely untapped resource of tabular data offers immense potential for analysis; and third, the business intelligence domain specifically demands adaptable, precise solutions that many current LLMs may struggle to provide. In response, we introduce TableGPT2, a model rigorously pre-trained and fine-tuned with over 593.8K tables and 2.36M high-quality query-table-output tuples, a scale of table-related data unprecedented in prior research. This extensive training enables TableGPT2 to excel in table-centric tasks while maintaining strong general language and coding abilities. One of TableGPT2's key innovations is its novel table encoder, specifically designed to capture schema-level and cell-level information. This encoder strengthens the model's ability to handle ambiguous queries, missing column names, and irregular tables commonly encountered in real-world applications. Similar to visual language models, this pioneering approach integrates with the decoder to form a robust large multimodal model. We believe the results are compelling: over 23 benchmarking metrics, TableGPT2 achieves an average performance improvement of 35.20% in the 7B model and 49.32% in the 72B model over prior benchmark-neutral LLMs, with robust general-purpose capabilities intact.

TableGPT2: A Large Multimodal Model with Tabular Data Integration

TL;DR

TableGPT2 targets the core difficulty of integrating structured tabular data with large language models for business intelligence tasks. It introduces a dedicated Semantic Table Encoder with 2D bi-directional attention and a Q-Former-style adapter to capture table semantics and align with textual inputs, combined with continual pretraining, supervised fine-tuning, and an agent framework for production-style tool use. The model achieves substantial gains across a broad suite of table-centric benchmarks and introduces RealTabBench to reflect real-world irregular and ambiguous tables, demonstrating strong generalization alongside robust coding and reasoning in BI contexts. The work also outlines an enterprise-oriented agent workflow with RAG and a secure code sandbox, highlighting practical considerations for deployment and future directions toward multi-agent and domain-specific adaptations.

Abstract

The emergence of models like GPTs, Claude, LLaMA, and Qwen has reshaped AI applications, presenting vast new opportunities across industries. Yet, the integration of tabular data remains notably underdeveloped, despite its foundational role in numerous real-world domains. This gap is critical for three main reasons. First, database or data warehouse data integration is essential for advanced applications; second, the vast and largely untapped resource of tabular data offers immense potential for analysis; and third, the business intelligence domain specifically demands adaptable, precise solutions that many current LLMs may struggle to provide. In response, we introduce TableGPT2, a model rigorously pre-trained and fine-tuned with over 593.8K tables and 2.36M high-quality query-table-output tuples, a scale of table-related data unprecedented in prior research. This extensive training enables TableGPT2 to excel in table-centric tasks while maintaining strong general language and coding abilities. One of TableGPT2's key innovations is its novel table encoder, specifically designed to capture schema-level and cell-level information. This encoder strengthens the model's ability to handle ambiguous queries, missing column names, and irregular tables commonly encountered in real-world applications. Similar to visual language models, this pioneering approach integrates with the decoder to form a robust large multimodal model. We believe the results are compelling: over 23 benchmarking metrics, TableGPT2 achieves an average performance improvement of 35.20% in the 7B model and 49.32% in the 72B model over prior benchmark-neutral LLMs, with robust general-purpose capabilities intact.

Paper Structure

This paper contains 72 sections, 4 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Motivation behind TableGPT2
  3. A system perspective.
  4. A data perspective.
  5. A BI perspective.
  6. Upon the Previous Version of TableGPT
  7. General Introduction of TableGPT2
  8. Continual pretraining.
  9. Supervised fine-tuning.
  10. Data augmentation for tabular data.
  11. Semantic table encoder.
  12. Agent framework.
  13. Evaluation.
  14. Disclaimer of Open Source
  15. Continual Pretraining from Qwen
  16. ...and 57 more sections

Figures (4)

  • Figure 1: Overall framework of TableGPT2.
  • Figure 2: Overall design of the Semantic Table Encoder.
  • Figure 3: The workflow of the complete agent process. Input queries are initially prepared and processed through a prompt engineering module. With the assistance of a retrieval-augmented generation (RAG) module from an external knowledge base, they are fed into the main model. TableGPT2 then collaborates with a visual language model (VLM) to generate tool calls, code, or other related actions. By observing intermediate results, the execution may iterate as necessary, leveraging the agent's reflection ability. This iterative process ultimately leads to the final output $y$ through seamless interaction between agents and tools.
  • Figure 4: Qualitative cases of TableGPT2.