Abstract
Beyond pure text, a substantial amount of knowledge is stored in tables. Inreal-world scenarios, user questions often require retrieving answers that aredistributed across multiple tables. GraphRAG has recently attracted muchattention for enhancing LLMs' reasoning capabilities by organizing externalknowledge to address ad-hoc and complex questions, exemplifying a promisingdirection for cross-table question answering. In this paper, to address thecurrent gap in available data, we first introduce a multi-table benchmark,MutliTableQA, comprising 60k tables and 25k user queries collected fromreal-world sources. Then, we propose the first Graph-Table-RAG framework,namely GTR, which reorganizes table corpora into a heterogeneous graph, employsa hierarchical coarse-to-fine retrieval process to extract the most relevanttables, and integrates graph-aware prompting for downstream LLMs' tabularreasoning. Extensive experiments show that GTR exhibits superior cross-tablequestion-answering performance while maintaining high deployment efficiency,demonstrating its real-world practical applicability.