FGTR: Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning

A research team led by Chaojie Sun has introduced FGTR (Fine-Grained Multi-Table Retrieval via Hierarchical LLM Reasoning), a novel AI method that fundamentally changes how large language models (LLMs) find and extract data from databases. Unlike previous approaches that encode entire tables for simple similarity matching—a process that often includes irrelevant data and struggles with scale—FGTR employs a human-like, two-stage reasoning strategy. It first uses the LLM to identify the relevant database schema elements (like specific columns or tables related to a query) and then retrieves only the precise cell contents needed. This hierarchical process results in a compact, accurate sub-table that directly answers the user's question, drastically cutting through noise.

To validate its performance, the team built two new benchmark datasets based on the established Spider and BIRD benchmarks. Experimental results are striking: FGTR outperforms previous state-of-the-art methods, improving the F2 evaluation metric by 18% on Spider and 21% on BIRD. This significant leap in accuracy demonstrates the method's effectiveness in handling fine-grained, multi-table queries, a task previously underexplored in retrieval research. The paper, currently under review for SIGIR 2026, highlights FGTR's potential to enhance end-to-end performance on downstream tasks that rely on table data, such as complex question answering and data analysis, by making LLM-based retrieval both more precise and computationally efficient.