Microsoft unveils software that lets LLMs work with spreadsheets

A team of programmers and AI specialists at Microsoft developed an AI tool called SpreadsheetLLM that applies large-scale language model capabilities to spreadsheets. In their study, published on the arXiv preprint server, the group developed SheetCompressor, an encoding framework that efficiently compresses spreadsheets for use by large language models (LLMs).

LLMs like ChatGPT are well known, but as more people use them, their skill gaps become more apparent. One of them is the ability of these models to understand spreadsheets. Because of their unique organization and capabilities, spreadsheets remain a mystery to LLMs, which prevents their use as a tool for businesses.

In this new study, the Microsoft team created a tool that reorganizes a spreadsheet into a format that master’s students can use as a data source. And, as the team points out, it’s based on a concept called SheetCompressor, a programming tool that allows AI to manage and analyze data to gain insights from spreadsheets.

To implement SheetCompressor, the researchers divided it into three main functions: compression, translation, and aggregation of data formats. The first was implemented by adding what the team describes as anchors to a spreadsheet to help an LLM understand what the spreadsheet does.

Once in place, the rows and columns are replaced with a skeletonized table. Translation modules are then used to remove empty cells or repeating values. Applying a lossless reverse index translation to JSON allows for aggregation of data formats.

The team also added other modules to handle unique situations, such as adjacent cells with similar number formats. The result is a tool that allows master’s students to use spreadsheets as a data source in a variety of ways.

The research team suggests that SpreadsheetLLM paves the way for using LLM technology to revolutionize the way spreadsheets are used; from automating data entry to analyzing it, to presenting complex information in a way that is accessible to people from diverse backgrounds. And this, they further suggest, will make spreadsheets and the data they contain much more accessible and useful.

© 2024 Science X Network