Abstract :
[en] Spreadsheets play a critical role in modern data work and are one of the most popular tools for data analysis across various domains. Despite their prevalence, spreadsheet tasks often require writing complex formulas, cleaning tabular data, and a thorough understanding of the heterogeneous structure, which are error-prone and time-consuming, even for expert users. The integration of Large Language Models (LLMs) in spreadsheet environments represents a significant paradigm shift, moving data analysis and manipulation from manual formulas to natural language interaction. This survey reviews the emerging landscape of LLM applications for spreadsheet tasks, identifies key methodologies, core capabilities, available benchmarks, and persistent challenges. First, we formulate the spreadsheet intelligence problem as a workflow of independent stages. Next, we categorize existing research works based on the defined stages and further segregate them into a taxonomy of tasks. Moreover, we list downstream tasks and their corresponding benchmarks, forming an end-to-end pipeline. Finally, we discuss open challenges and outline future research directions towards trustworthy LLM systems in spreadsheet environments.
