SheetMind: New LLM Framework Achieves 55% Success Rate on Spreadsheet Automation Tasks

Researchers have developed SheetMind, a multi-agent system that automates spreadsheet operations through conversational instructions. The framework uses specialized large language model agents working in tandem to understand user requests, generate commands, and verify accuracy—addressing a common productivity bottleneck.

How it works#

The system operates as a hierarchical pipeline with three distinct roles:

• Manager Agent breaks down complex user instructions into smaller, manageable subtasks
• Action Agent translates these subtasks into structured machine commands using a formal grammar specification (Backus-Naur Form)
• Reflection Agent checks whether the generated actions actually match what the user intended

This three-tier approach allows SheetMind to handle nuanced requests that might confuse simpler automation tools.

Performance and testing#

When evaluated on the SheetCopilot Benchmark—a suite of 221 real-world spreadsheet tasks—SheetMind achieved flawless execution success while reaching 54.8% functional correctness (meaning the correct result was produced). This outperformed the previous leading system, SheetCopilot, which managed 44.3% accuracy. Critically, SheetMind maintained perfect reliability, meaning every command executed without errors, even when the result wasn't entirely accurate.

Ablation studies on a separate test set confirmed that all three agents working together consistently beat any stripped-down two-agent variant, validating the hierarchical design.

Real-world deployment#

The team built SheetMind on GPT-3.5-Turbo and integrated it directly into Google Sheets as a native Workspace extension, making it accessible to ordinary users without requiring technical setup.