LIGHT-TABLE-CHAIN: Simplifying and Enhancing Chain-Based Table Reasoning

The objective of table reasoning tasks is to accurately extract answers from complex tabular data based on specific queries. While chain-based table reasoning methods have proven effective for stepwise inference, traditional approaches often encounter inefficiencies and redundancies due to their re-liance on lengthy and intricate chains of operations that refine the original data. To address these challenges, we propose LIGHT-TABLE-CHAIN, a novel approach that streamlines the reasoning process and enhances overall efficiency. Our method begins by transforming the user query into a target-oriented final table, which serves as a guiding framework for subsequent reasoning operations. We introduce a guided pruning strategy that leverages this final table's structure to iteratively refine and optimize the original data, thereby minimizing unnecessary operations. This method enables large language models (LLMs) to accurately select pruning operations, constructing an efficient pathway while ensuring the accuracy and reliability of the reasoning process. Extensive experiments on benchmark datasets, including TabFact and WikiTQ, demonstrate that LIGHT-TABLE-CHAIN achieves state-of-the-art performance, significantly surpassing existing methodologies in table reasoning.