Focus on Evidence: Relational-Structure Enhances LLM Effectiveness in TableQA

Table Question Answering (TableQA) requires reasoning over natural language questions and structured tables, and remains challenging due to noisy evidence and complex multi-step reasoning. Recent Large Language Model (LLM)-based approaches typically adopt decomposition–reasoning–validation pipelines that combine Chain-of-Thought (CoT) decomposition, Direct Prompting (DP), Python Agent (PyAgent) execution, and self-validation. However, these methods still largely rely on unconstrained LLM reasoning during decomposition, which may introduce unsupported sub-questions and irrelevant evidence expansion. We propose

idence- ware (EVA), a framework that explicitly grounds decomposition on question-relevant table evidence before reasoning begins. EVA first extracts evidence-aware relational structures that associate question semantics with supporting table attributes, thereby constraining sub-question generation within a grounded evidence space. EVA further introduces relation-guided validation and integrates both DP and PyAgent reasoning to improve robustness across different reasoning patterns. Experiments on datasets across multiple LLMs demonstrate that EVA consistently improves reasoning reliability and achieves strong performance across diverse TableQA settings.