# GTR: Graph-Table-RAG for Cross-Table Question Answering

**Authors:** Jiaru Zou, Dongqi Fu, Sirui Chen, Xinrui He, Zihao Li, Yada Zhu, Jiawei Han, Jingrui He  
**arXiv:** [2504.01346](https://arxiv.org/abs/2504.01346)  
**Date:** 2025-04-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2504.01346v1/

## Key points

- The paper explores the utilization of tables as a source of knowledge beyond pure text
- Introduces GraphRAG to enhance Large Language Models (LLMs) for addressing complex queries involving information from diverse tables
- Presents MutliTableQA benchmark with 60k tables and 25k user queries sourced from real-world data
- Proposes the Graph-Table-RAG framework (GTR) which transforms table corpora into a heterogeneous graph structure
- GTR outperforms existing methods in cross-table question answering while maintaining high deployment efficiency

## Summary

, , , , 

The paper "GTR: Graph-Table-RAG for Cross-Table Question Answering" by Jiaru Zou, Dongqi Fu, Sirui Chen, Xinrui He, Zihao Li, Yada Zhu, Jiawei Han, and Jingrui He explores the utilization of tables as a source of knowledge beyond pure text. In real-world scenarios, user questions often span multiple tables, requiring advanced reasoning capabilities to retrieve accurate answers. The authors introduce GraphRAG as a method to enhance Large Language Models (LLMs) in addressing complex queries that involve information from diverse tables. To bridge the existing gap in available data for cross-table question answering, the authors present the MutliTableQA benchmark consisting of 60k tables and 25k user queries sourced from real-world data. Building upon this benchmark, the authors propose the Graph-Table-RAG framework (GTR), which transforms table corpora into a heterogeneous graph structure. GTR employs a hierarchical retrieval process to extract relevant tables and incorporates graph-aware prompting to facilitate tabular reasoning for downstream LLMs. Through extensive experiments, it is demonstrated that GTR outperforms existing methods in cross-table question answering while maintaining high deployment efficiency. This highlights the practical applicability of GTR in real-world scenarios where users need to extract information from multiple tables to effectively answer complex queries. The study showcases how organizing external knowledge in a structured graph format can significantly enhance LLMs' ability to reason across diverse sources of information for improved question-answering performance.