Researchers at Microsoft have developed a new LLM designed to better leverage AI in spreadsheets like Excel and Google Sheets.

Researchers from Microsoft have published a research paper under the original name SpreadsheetLLM, proposing an experimental LLM specifically designed for spreadsheets. Existing LLMs are unable to understand the structured data, also due to the token limit. The experimental SpreadsheetLLM uses a new encoding mechanism that puts the contents of a spreadsheet into a new format that is more readable by the LLM. The model is still in the experimental phase, but the first results are positive.

LLM-friendly style

Spreadsheets are widely used in the business world: from simple formulas to complex financial models. However, existing LLMs are unable to understand the nature of these data and complex formulas. To address this challenge, Microsoft researchers have developed a special LLM for spreadsheets.

To ensure that LLMs can better understand the structured nature of data in spreadsheets, the researchers developed a new encryption mechanism called “SheetCompressor.” This method encodes data in an LLM-friendly manner while preserving the structure and relationships of the data.

In addition to SheetCompressor, the researchers highlighted another feature called Structural Anchor Extraction, which identifies the key rows and columns that define table structures. In addition, the inverted index translation method is applied, which efficiently encodes the cell contents and addresses to minimize redundancy.

potential

The experiments showed that SpreadsheetLLM achieved positive results in the spreadsheet recognition test, outperforming existing methods by 12.3 percent. SpreadsheetLLM was applied to several well-known LLMs such as GPT-3.5, GPT-4, and Llama 2. The researchers found that the ability of these models to understand spreadsheets was improved, with GPT-4 achieving a table recognition score of 78.9 percent.

The researchers strongly believe in the potential of this model, although there are currently limitations with more complex spreadsheet formats. SpreadsheetLLM could perform automated tasks or make spreadsheets more accessible to people who have difficulty using them.

SpreadsheetLLM is currently a research project. Microsoft has not yet announced any plans to bring this model to market.