Interpretable Table Question Answering via Plans of Atomic Table Transformations

Giang Nguyen, Ivan Brugere, Shubham Sharma, Sanjay Kariyappa, Anh Totti Nguyen, Freddy Lecue

Keywords: Table QA; Interpretability; XAI; LLM-as-a-Judge

TL;DR: We introduce POS, a Table QA method specifically designed for interpretability, decomposing complex queries into atomic natural-language sub-queries, which are then translated into SQL commands to sequentially transform input table into final answer.

Abstract:

Interpretability for Table Question Answering (Table QA) is critical, particularly in high-stakes domains like finance or healthcare.
While recent Large Language Models (LLMs) have improved the accuracy of Table QA models, their explanations for how answers are derived may not be transparent, hindering user ability to trust, explain, and debug predicted answers, especially on complex queries.
We introduce Plan-of-SQLs (POS), a novel method specifically crafted to enhance interpretability by decomposing a query into simpler sub-queries that are sequentially translated into SQL commands to generate the final answer.
Unlike existing approaches,
POS offers full transparency in Table QA by ensuring that every transformation of the table is traceable, allowing users to follow the reasoning process step-by-step.
Via subjective and objective evaluations, we show that POS explanations significantly improve interpretability, enabling both human and LLM judges to predict model responses with 93.00% and 85.25% accuracy, respectively.
POS explanations also consistently rank highest in clarity, coherence, and helpfulness compared to state-of-the-art Table QA methods such as Chain-of-Table and DATER.
Furthermore, POS demonstrates high accuracy on Table QA benchmarks (78.31% on TabFact and 54.80% on WikiTQ with GPT3.5), outperforming methods that rely solely on LLMs or programs for table transformations, while remaining competitive with hybrid approaches that often trade off interpretability for accuracy.

Primary Area: interpretability and explainable AI

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7407