Cell-role annotation beats the state of the art on spreadsheet RAG and still hits a ceiling classification cannot pass

arXiv·low signalXBlueskyLinkedInCopy link

The framework splits any spreadsheet into interpretable chunks by annotating cell roles, and the authors report it beats prior work, but the gain comes from enriched context helping answer generation rather than better retrieval accuracy. Their argument is that spreadsheets are two-dimensional unstructured data with continuous relationships and effectively unbounded cell roles, so any finite classifier hits a hard ceiling even at human-level annotation quality. They call for dimensionality-reduction methods that flatten 2D sheets into 1D text instead.