Reasoning by Commented Code for Table Question Answering

2026-01-31

AI Summary

This work addresses the challenge that large language models often disrupt the two-dimensional structure of tables during linearization, leading to inaccurate numerical reasoning and limited interpretability. To mitigate this, the authors propose a multi-step code generation framework that, for the first time, incorporates explicit natural language annotations during program synthesis. By decomposing table-based question answering into annotated, executable Python programs, the approach enhances the model’s understanding of tabular structure, improves numerical accuracy, and increases reasoning transparency. Built upon Qwen2.5-Coder-7B-Instruct and integrating multi-line annotated program synthesis with a lightweight answer selection mechanism, the method achieves 70.9% accuracy on WikiTableQuestions. When further combined with an end-to-end model, performance improves to 84.3%, substantially outperforming the Repanda baseline (67.6%).

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Search and Retrieval-Augmented AI: Web question answering
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Table Question Answering (TableQA) poses a significant challenge for large language models (LLMs) because conventional linearization of tables often disrupts the two-dimensional relationships intrinsic to structured data. Existing methods, which depend on end-to-end answer generation or single-line program queries, typically exhibit limited numerical accuracy and reduced interpretability. This work introduces a commented, step-by-step code-generation framework that incorporates explicit reasoning into the Python program-generation process. The approach decomposes TableQA reasoning into multi-line executable programs with concise natural language comments, thereby promoting clearer reasoning and increasing the likelihood of generating correct code. On the WikiTableQuestions benchmark, the proposed method achieves 70.9\% accuracy using Qwen2.5-Coder-7B-Instruct, surpassing the Repanda baseline (67.6\%). Integrating the proposed framework with a robust end-to-end TableQA model via a lightweight answer-selection mechanism yields further improvements. This combined approach achieves up to 84.3\% accuracy on the WikiTableQuestions benchmark.

Problem

• Table Question Answering
• Large Language Models
• Structured Data Reasoning
• Numerical Accuracy
• Interpretability

Innovation

• commented code generation
• table question answering
• step-by-step reasoning
• executable program synthesis
• interpretable LLM