Multi Table QA:Evaluating Modern LLM Strategies on Table0 Understanding

Abstract

Tables constitute the majority of structured data in enterprise environments. While single-table question answering has received significant attention, research on multi-table reasoning remains limited. Compared to single-table QA, multi-table QA requires schema alignment, relational inference, and scalable context man- agement. We introduce tool-augmented reasoning as a paradigm for multi-table QA, and systematically study two complementary strategies: (1) free-form tool interaction, where models iteratively call exploration and computation tools, and(2) structured agent workflows, which stage tool-use into exploration, preparation, and analysis phases. Using the TQA-Bench dataset, we evaluate GPT-4o-mini and GPT-5-mini across both small (8k) and large (128k) database scales under varying tool context constraints. We show that tool augmentation substantially improves robustness and accuracy over direct prompting, with gains up to +28 percentage points. Structured workflows yield further benefits for weaker models on large databases, but regress for stronger models, revealing a scale–capacity trade-off in the value of structure. These results establish tool-augmented reasoning as a powerful paradigm for multi-table QA. Structure aids weaker models under scale but constrains stronger models, underscoring that tool-use strategies must be adapted jointly to model ability and database complexity.