First systematic study of LLM data-referencing errors when reading tables

arXiv·medium signal

Even when LLMs understand table structure, they make data referencing errors — incorrectly citing or omitting cell values — that corrupt intermediate reasoning steps, not just final answers. This is the first large-scale, systematic evaluation of DREs across models, moving past prior small-scale anecdotes. Directly relevant to builders shipping table-QA, spreadsheet-agent, and analytics features where a silently misread cell poisons the whole chain.