Semantically Similar, Yet Not Answerable: Diagnosing the Semantic-Answerability Gap in Table RAG
The study identified a 'Semantic-Answerability Gap' in Table RAG, where dense retrievers find semantically similar tables but fail to select the uniquely answerable one. Using TCR-Bench, researchers found that semantic accumulation and weak row-column binding cause this failure. Introducing Answerability-Aware Reranking (AAR) significantly improved performance by explicitly judging query-table interaction during reranking.