Microsoft researchers have developed SpreadsheetLLM, a method designed to help language models work with spreadsheets that are too large or complex for conventional AI analysis. The core idea is simple: instead of forcing a model to process every cell in a bulky file, SpreadsheetLLM converts the spreadsheet into a compact representation that preserves the information the system needs.
That matters because spreadsheets are a common format for scientific and financial work, but their size and structure can push AI systems beyond practical limits. According to the team, SpreadsheetLLM can reduce the amount of data by up to 96 percent without dropping important information, opening the door to analysis that was not possible before.
Why spreadsheets are difficult for language models
Language models are built to process sequences of text. Spreadsheets, by contrast, are structured grids with rows, columns, cell addresses, repeated values, empty areas, number formats and table boundaries. A spreadsheet can contain several different layouts in the same file, which makes it harder for an AI system to understand where one meaningful table begins and another ends.
The source problem is not only size. A large spreadsheet can contain a mix of dense and sparse areas, repeated labels, numbers with related formats and empty cells that add little value for analysis. If all of that is serialized in a traditional row-by-column form, the model may spend too much of its available capacity on structure rather than meaning.
SpreadsheetLLM addresses that by making the spreadsheet smaller before the language model analyzes it. The method does not simply delete data at random. It uses three techniques to identify the shape, contents and formats of the spreadsheet so the model receives a condensed version that still carries the key signals.
The three techniques behind SpreadsheetLLM
The first technique is Structural Anchors. It looks for heterogeneous rows and columns that may mark table boundaries. It also removes distant homogeneous rows and columns, then builds a condensed skeleton version of the spreadsheet. This gives the model a clearer view of the spreadsheet layout without requiring every cell to be included.
The second technique is Inverted-Index Translation. Instead of using a conventional row-by-column serialization, it creates a JSON-format inverted-index translation. This approach builds a dictionary of non-empty cell text and merges cell addresses that contain identical text. The purpose is to reduce token use while maintaining data integrity.
The third technique is Data Format Aggregation. It extracts number format strings and data types from adjacent numerical cells, then groups cells with similar formats or types. That helps preserve meaningful numerical structure while avoiding a full cell-by-cell representation.
Together, these techniques let SpreadsheetLLM capture the information that defines a spreadsheet: where tables are likely located, which text values matter, which cells repeat, and how related numerical data is formatted. The result is a version of the spreadsheet that is much smaller, but still useful for language-model analysis.
Accuracy gains show up most clearly in large files
The researchers tested SpreadsheetLLM with several AI models, including OpenAI's GPT-4 and open-source models such as Llama 2. In table recognition tasks, the system reached 79 percent accuracy. That was 13 percentage points higher than the previous best score reported in the source article.
The largest spreadsheets showed the biggest benefit. For those files, accuracy improved by 75 percentage points compared with conventional techniques. The reason given is practical: the compact format kept the data within language-model token limits, so the model could analyze files that would otherwise exceed those limits.
This is an important distinction. SpreadsheetLLM is not described as a new language model. It is a way to prepare spreadsheet data so existing models can work with it more effectively. By changing the representation, the method changes what the model can realistically process.
For scientific and financial spreadsheets, that could be especially useful. Those files often depend on structure as much as content. A number by itself is rarely enough; its meaning depends on the row, column, nearby labels, formatting pattern and table region around it. SpreadsheetLLM is designed to preserve those signals in a more compact form.
Chain of Spreadsheet adds a query workflow
The researchers also developed a technique called Chain of Spreadsheet, or CoS, for answering complex questions about spreadsheets. Rather than asking the model to answer directly from the entire file representation, CoS splits the task into two steps.
- First, the system identifies the relevant table area.
- Then, it generates the answer from that selected area.
That two-step process gives the model a narrower target. Instead of searching across the whole spreadsheet at once, it focuses on the part most likely to contain the answer. Using this method, the system achieved 74 percent accuracy on spreadsheet question-answering tasks.
This workflow reflects the broader logic of SpreadsheetLLM. The method is not just about making spreadsheets smaller. It is about guiding the model toward the right structure before asking it to reason over the data.
What still needs work
The researchers also note limitations. SpreadsheetLLM does not currently consider formatting details such as background colors, even though those visual cues could contain useful information. In many spreadsheets, formatting can help separate categories, flag important cells or make table regions easier to interpret.
The team also sees room to improve the semantic condensation of text cells. That matters because text in spreadsheets can carry labels, descriptions and categories that shape the meaning of nearby values. Compressing that content while keeping its meaning is a harder problem than reducing repeated numbers or empty cells.
Even with those limits, SpreadsheetLLM points to a practical direction for AI spreadsheet analysis. Instead of treating a spreadsheet as a plain block of serialized cells, the method treats it as a structured object that needs translation. For large scientific and financial files, that translation may be the difference between a model that cannot fit the data and a model that can produce a usable answer.