Microsoft is working on a new productivity tool that helps artificial intelligence better understand spreadsheets. Still in the experimental phase, SpreadsheetLLM addresses challenges that are unique to applying AI to spreadsheets, “with their extensive two-dimensional grids, various layouts, and diverse formatting options,” the company explains. Hailed as a significant development in the enterprise space, where spreadsheets are used for everything from data entry to financial modeling and are shared among departments, Microsoft points out that as a research area spreadsheet-optimized AI has generally been overlooked in favor of flashier use-cases.

“Existing language models struggle to understand and reason over spreadsheet contents due to the structured nature of the data and the presence of formulas and references,” Microsoft explains in a research paper. SpreadsheetLLM addresses this challenge by encoding spreadsheet data so LLMs can work with it.

In short, SpreadsheetLLM is a potential breakthrough in that it “paves the way for these models to ‘reason over spreadsheet contents,” SiliconANGLE summarizes.

Beyond that, it can automate repetitive spreadsheet tasks, and offers the additional advantage of allowing users to query spreadsheets using natural language, making their contents accessible to workers who are not familiar with protocols. “This could democratize access to data insights and empower more individuals within an organization to make data-driven decisions,” VentureBeat writes.

“The model uses a novel encoding scheme that preserves the structure and relationships within the spreadsheet while making it accessible to language models,” VentureBeat explains, opening up “exciting possibilities for AI-assisted data analysis and decision-making in the enterprise.”

To accomplish these feats, Microsoft researchers developed an encoding formula called SheetCompressor, which preserves the data relationships and structure while making it LLM-friendly. “SheetCompressor notably compresses the data by up to 96 percent, so LLMs can handle large datasets within their token limits,” SiliconANGLE reports.

There are several other acceleration features, detailed with charts and graphs in the Microsoft research paper and in SiliconANGLE, which notes that experiments by the research team found that SpreadsheetLLM was able to outperform “existing methods by 12.3 percent,” achieving “strong results on spreadsheet question-answering tasks.”

Pitted against well-known LLMs such as OpenAI’s GPT-3.5 and GPT-4 and Meta’s Llama 2, internal testing revealed SpreadsheetLLM “significantly enhanced those models’ ability on spreadsheet understanding tasks.” Using the example of GPT-4, SiliconANGLE says the experimental tool “achieved a table detection score of 78.9 percent,” the rate at which it accurately identified tables.

Though still in the experimental phase, with no announced timeline for release, SiliconANGLE posits “it’s not hard to imagine some kind of ‘Copilot for Excel’ might emerge from this research.”