A Compact Formula-Aware Spreadsheet Abstraction for Large Language Models

| Source: arXiv AI

Tags: spreadsheets, R1C1 formulas, token efficiency, LLM context compression, static analysis

A formula-aware abstraction compresses large spreadsheet ranges into positions, types and R1C1 formula patterns before LLMs see the data, cutting token use on a 50-query evaluation.

Details

LLMs struggle with spreadsheets because their grid structure does not map well to sequential text. This paper proposes an abstraction that represents large homogeneous ranges by their positions, types, R1C1 formula patterns and selected boundary values. An algorithm detects these ranges and serializes them, and the abstraction is given to the LLM before detailed content. On 50 spreadsheet-query cases, the abstraction alone answered some queries, and when more data was needed it reduced how much had to be sent, lowering token consumption. The single-author paper reports better reliability and accuracy, though the evaluation is small.