AI Papers Reader

Personalized digests of latest AI research

View on GitHub

No More Broken Formulas: How AI Is Learning to Master Microsoft Excel

Spreadsheets are the unsung heroes of modern business, holding together everything from household budgets to multi-billion-dollar Wall Street financial models. Yet, for all their power, they are famously fragile. A single misplaced formula or shifted column can break an entire workbook. This fragility has long stymied artificial intelligence; while modern large language models (LLMs) can write poetry or draft essays, they consistently fumble when tasked with executing complex, multi-step operations inside Microsoft Excel.

To bridge this gap, researchers from the University of Illinois Urbana-Champaign and Meta have developed Spreadsheet-RL, a groundbreaking reinforcement learning framework designed to train AI agents to interact with spreadsheets like human power-users.

To understand why spreadsheets stump conventional AI, consider a simple task: deleting every column in a sheet where the header contains the word “description.” A typical AI agent attempting this by writing ad-hoc Python code might iterate through the columns from left to right. However, as soon as the agent deletes Column 1, Column 2 shifts left to become the new Column 1. The naive AI, continuing its loop to the next index, ends up skipping columns or deleting the wrong data entirely.

Spreadsheet-RL solves this by abandoning raw, unpredictable code-generation in favor of a specialized, “spreadsheet-native” environment. First, researchers developed Spreadsheet Gym, a sandbox environment powered by actual Microsoft Excel. This allows the AI to interact with advanced Excel features, such as complex formulas and dynamic arrays, under realistic execution rules.

Crucially, they equipped the AI with a custom “Tool Harness.” Instead of writing raw code from scratch, the AI uses structured, guardrailed tools. For example, it uses delete_columns (which automatically handles shifting indices) and fill_formula (which correctly translates relative cell references, like adjusting A1 to A2 as a formula is dragged down a column). The agent is trained to follow an intuitive, human-like workflow: inspect the relevant ranges, modify the workbook using specialized tools, and verify the output using Excel’s calculation engine before finalized.

To train the agent, the researchers turned to on-policy reinforcement learning (specifically, the GRPO algorithm). Rather than relying on expensive human-annotated data, they built an automated pipeline to scrape real-world problems and solutions from popular online message boards like ExcelForum, translating them into thousands of training scenarios. If the AI agent successfully edits a spreadsheet to match the target “oracle” result, it receives a positive reward; if it breaks a formula or fails the task, it learns from the error and adjusts its policy.

The results are striking. When applied to the open-source model Qwen3-4B-Thinking, Spreadsheet-RL nearly doubled its success rate on SpreadsheetBench, a standard industry benchmark, pushing its accuracy from 12.0% to 23.4%.

Furthermore, to test how well the AI generalizes to professional scenarios, the team curated a new “Domain-Spreadsheet” benchmark containing 1,660 complex tasks in finance, human resources, and supply chain management—such as calculating corporate Value-at-Risk or modeling debt-service coverage ratios. Under this rigorous evaluation, the Spreadsheet-RL trained agent improved its pass rate from 8.4% to 17.2%, outperforming several much larger baselines.

By open-sourcing the training pipeline, datasets, and environment, the researchers have laid the groundwork for a new generation of reliable, autonomous digital assistants—promising to save humans from the dread of the #REF! error once and for all.