After texts and images: Microsoft AI now masters Excel tables
AI models that can generate text, images, music and even videos are now well known. However, Microsoft researchers have now presented a system that can also generate complex Excel tables.
Inaccuracies are no longer acceptable
The new SpreadsheetLLM is not yet publicly available, but interested parties can already register by paper familiarize yourself with the model. It is said to be “extremely effective at a wide range of spreadsheet tasks” and has the potential to “revolutionize the management and analysis of spreadsheet data and pave the way for smarter and more efficient user interactions.” Mastering a spreadsheet sounds like a relatively simple matter, but it is a much more difficult task for AI developers than generating various media formats. In images and videos, minor inaccuracies can be easily overlooked.
However, when it comes to using various formulas to get the desired result from data in a table, a high degree of accuracy is essential. A small error can result in completely distorted data, which would render the entire project useless.
step by step
One of the problems with using LLMs in spreadsheets is that they are slowed down by too many tokens (basic units of information that the model processes). To solve this problem, Microsoft developed SheetCompressor, an “innovative coding framework that effectively compresses spreadsheets for LLMs.” LLM approaches understanding a table and its structured data in several stages. First, “structural anchors” are placed that help the AI understand the calculations step by step.
Then, redundant content that tends to cause confusion is removed and a “skeleton version” of the table is created that is easier to analyze. “To improve efficiency, we are moving away from traditional row- and column-wise serialization and using lossless, inverted index translation in JSON format,” Microsoft said. “This method creates a dictionary that indexes non-empty cell text and merges addresses with identical text, optimizing token usage while maintaining data integrity.”
