# SheetAgent: Towards A Generalist Agent for Spreadsheet Reasoning and Manipulation via Large Language Models

**Authors:** Yibin Chen, Yifu Yuan, Zeyu Zhang, Yan Zheng, Jinyi Liu, Fei Ni, Jianye Hao  
**arXiv:** [2403.03636](https://arxiv.org/abs/2403.03636)  
**Date:** 2024-03-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.03636v2/

## Key points

- Recent advancements in spreadsheet manipulation involve the integration of Large Language Models (LLMs) for automating tasks and enhancing efficiency.
- LLMs have shown promise in simpler operations but are limited in more complex scenarios requiring intricate reasoning challenges.
- The introduction of the $\textbf{SheetRM}$ benchmark addresses this gap by encompassing long-horizon tasks across multiple categories that demand manipulation based on reasoning-dependent factors from real-life complexities.
- The innovative $\textbf{SheetAgent}$ is proposed as an autonomous agent comprising three collaborative modules - Planner, Informer, and Retriever - leveraging LLM capabilities for advanced reasoning and precise spreadsheet manipulation without human intervention.
- SheetAgent showcases notable improvements ranging from 20% to 30% in pass rates across various benchmarks compared to baseline models through iterative task reasoning and reflection mechanisms.
- This enhanced precision underscores SheetAgent's superior table reasoning abilities, contributing towards developing a generalist agent tailored for spreadsheet reasoning and manipulation using LLMs.
- The research paper detailing these findings has been accepted by the Large Language Models and Cognition conference at ICML 2024. Interested individuals can explore further insights and visualizations of SheetAgent's capabilities at https://sheetagent.github.io.

## Summary

In the realm of spreadsheet manipulation, recent advancements have seen the integration of Large Language Models (LLMs) for automating tasks and enhancing efficiency. While LLMs have shown promise in simpler operations, their application has been limited to more complex scenarios involving intricate reasoning challenges. To address this gap and cater to real-world demands, a new benchmark known as $\textbf{SheetRM}$ has been introduced. This benchmark encompasses long-horizon tasks across multiple categories that require manipulation based on reasoning-dependent factors stemming from real-life complexities. To tackle these challenges, the innovative $\textbf{SheetAgent}$ has been proposed as an autonomous agent leveraging the capabilities of LLMs. Comprising three collaborative modules - namely the $\textit{Planner}$, $\textit{Informer}$, and $\textit{Retriever}$ - SheetAgent excels in advanced reasoning and precise spreadsheet manipulation without requiring human intervention. Through iterative task reasoning and reflection mechanisms, SheetAgent showcases its prowess by delivering notable improvements ranging from 20% to 30% in pass rates across various benchmarks when compared to baseline models. This enhanced precision in spreadsheet manipulation underscores SheetAgent's superior table reasoning abilities. The work conducted by Yibin Chen, Yifu Yuan, Zeyu Zhang, Yan Zheng, Jinyi Liu, Fei Ni, and Jianye Hao culminates in a significant contribution towards developing a generalist agent tailored for spreadsheet reasoning and manipulation using Large Language Models. The research paper detailing these findings has been accepted by the Large Language Models and Cognition conference at ICML 2024. For further insights and visualizations pertaining to SheetAgent's capabilities, interested individuals can explore additional information available at https://sheetagent.github.io.