A new research framework wants AI agents to stop treating spreadsheets like plain text and start reading them the way analysts actually do.
Researchers describe SheetCompass, a graph-guided, memory-driven framework for spreadsheet reasoning, in a paper posted to arXiv on August 17, 2026. The paper argues that large language models struggle with real-world workbooks because most existing tools flatten cross-table references, column dependencies, and spatial layout into a single linear string of text. That process erases the boundaries between tables and the implicit relationships between sheets that a human analyst would spot at a glance. SheetCompass instead builds hierarchical relation graphs that explicitly capture structural links within and across worksheets, paired with a memory component that keeps task-relevant details on hand as an agent works through a workbook.
Spreadsheets remain one of the most common ways organizations store and analyze data, and most LLM-based tools handle them by dumping cell contents into a prompt as raw text. That approach breaks down once a workbook has multiple linked sheets, merged headers, or dependencies that only make sense spatially. Treating structure as a first-class input, rather than something to flatten away before an LLM ever sees it, is a more foundational fix than prompt tricks or bigger context windows.
It is a research proposal, not a shipped product - the real test is whether SheetCompass holds up on messy, inconsistent workbooks in the wild, not the tidy examples that tend to make it into a paper.