Large Language Models (LLMs) achieve strong results on table question answering, yet their accuracy drops as table size grows. This degradation is not uniform across question types: localization‑sensitive queries suffer heavily from irrelevant table content, while queries that require broader evidence can still benefit from reasoning over the entire table.
Motivated by this observation, we propose a question‑adaptive framework that dynamically switches between localized sub‑table reasoning and full‑table reasoning. The framework first constructs question‑specific sub‑tables via operation‑aware table decomposition, then uses the predicted question type to select the appropriate reasoning mode.
To evaluate evidence selection, we introduce silver reference sub‑tables and build SLQA, a benchmark based on real‑world long tables. Experiments on WikiTQ and SLQA show that localization greatly improves lookup and local reasoning questions, and that adaptive selection between localized and full‑table reasoning yields the best overall performance.
These results highlight that long‑table QA requires not only deciding how to localize, but also when to localize.
Review