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

With the rapid advancement of large language models (LLMs), there has been an increasing focus on LLM-based table retrieval. However, existing studies primarily concentrate on single-table queries, implementing them through similarity matching after encoding the entire table. These methods often result in low accuracy due to their coarse-grained encoding, which incorporates much query-irrelevant data, and are inefficient when dealing with large tables, failing to fully utilize the reasoning capabilities of LLMs. Moreover, multi-table queries remain under-explored in retrieval tasks. To address this, the paper proposes a hierarchical multi-table query method based on LLMs: Fine-Grained Multi-Table Retrieval (FGTR), a new retrieval paradigm that employs a human-like reasoning strategy. Through hierarchical reasoning, FGTR first identifies relevant schema elements and then retrieves the corresponding cell contents, ultimately constructing a concise and accurate sub-table that aligns with the given query. To comprehensively evaluate the performance of FGTR, two new benchmark datasets based on Spider and BIRD were constructed. Experimental results show that FGTR outperforms previous state-of-the-art methods, improving the F_2 metric by 18% on Spider and 21% on BIRD, demonstrating its effectiveness in enhancing fine-grained retrieval and its potential to improve end-to-end performance on table-based downstream tasks.

Table reasoning tasks involve performing reasoning and deriving answers over structured tables, widely used in healthcare, finance, and corporate financial reporting. Distinct from textual data, tabular data exhibits a two-dimensional structure with complex logical relationships among rows, columns, and cell contents, posing significant NLP challenges. Common tasks include table question answering, table fact verification, table-to-text generation, and table semantic parsing. As retrieved tables serve as critical inputs for reasoning, effective table information retrieval becomes pivotal. Reducing redundant table content directly enhances model accuracy, computational efficiency, and complex reasoning capability.

In natural language processing (NLP), LLMs have made significant strides, opening new frontiers for table reasoning tasks. However, their application to structured data still faces persistent challenges due to inherent limitations. First, LLMs’ inherent context window constraints pose critical limitations when processing large-scale tables containing thousands of rows or columns despite their proficiency in conventional text tasks. Common mitigation strategies for large tables include table segmentation or setting a token budget for content extraction, which risks information loss and subsequent degradation of reasoning accuracy. Second, LLMs demonstrate constrained capabilities in comprehending tabular structures and logical relationships, attributed to:

Hence, it is necessary to overcome these issues through fine-grained multi-table retrieval, as table-based reasoning constitutes a complex task requiring human-like cognitive processes and fine-grained comprehension. However, existing studies still exhibit two main limitations. First, current approaches primarily encode the entire table or encode rows and columns to retrieve question-relevant contents and construct sub-tables. Such methods depend heavily on holistic table representations, resulting in limited understanding of table structures and low encoding efficiency. In contrast, as illustrated in Figure, humans typically first attend to the table’s schema structure—including column names, attributes, and sample entries—to rapidly grasp the informational framework and overall theme. On this basis, they subsequently focus on the columns required by the question, locate specific cell contents, and obtain the relevant rows needed to solve the problem. Second, existing fine-grained table retrieval methods tend to oversimplify real-world scenarios by restricting retrieval to a single table, neglecting the fact that in practical cases, the required information is often distributed across multiple tables and must be integrated through inter-table joins and relationships.