README.md

metadata

title: Spreadsheet-RL
emoji: 📊
colorFrom: blue
colorTo: green
sdk: static
sdk_version: 1.0.0
app_file: index.html
pinned: false
short_description: RL post-training for spreadsheet agents.
datasets:
  - Spreadsheet-RL/Spreadsheet-RL
tags:
  - spreadsheet
  - excel
  - reinforcement-learning
  - agents

Spreadsheet-RL

Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning

Spreadsheet-RL is an open research effort for training spreadsheet agents that can operate in realistic Microsoft Excel environments. We study how reinforcement learning post-training can improve long-horizon, multi-turn spreadsheet workflows involving formulas, formatting, recalculation, workbook structure, and domain-specific analysis.

This organization hosts the public Spreadsheet-RL releases, including the dataset used for training and evaluation. The companion codebase is available on GitHub.

Latest

• Project page: https://spreadsheet-rl.github.io/
• Paper: https://arxiv.org/abs/2605.22642
• Hugging Face Daily Papers: https://huggingface.co/papers/2605.22642
• Code: https://github.com/Spreadsheet-RL/Spreadsheet-RL
• Dataset: https://huggingface.co/datasets/Spreadsheet-RL/Spreadsheet-RL

What We Release

• Spreadsheet Data Agent: a scalable data-construction pipeline that builds paired initial and oracle final workbooks from realistic spreadsheet tasks.
• Spreadsheet Gym: a multi-turn Excel environment with spreadsheet-native tools, sandboxed code execution, isolated workspaces, and Excel-based reward computation.
• Spreadsheet-RL training stack: an end-to-end GRPO post-training pipeline for spreadsheet agents.
• Domain-Spreadsheet: a domain-specific benchmark covering finance, supply chain, human resources, sales, and real estate workflows.

Highlights

• Qwen3-4B-Thinking-2507 improves from 12.0% to 23.4% Pass@1 on SpreadsheetBench after spreadsheet-native harnessing, comprehensive tool access, and RL post-training.
• On Domain-Spreadsheet, Spreadsheet-RL improves overall Pass@1 from 8.4% to 17.2% across 1,660 evaluation rollouts.
• The released dataset includes ExcelForum training tasks, SpreadsheetBench evaluation tasks, Domain-Spreadsheet tasks, parser-specific parquet files, and workbook archives.

Attribution

Spreadsheet-RL is a research collaboration by authors from the University of Illinois Urbana-Champaign and Meta. Data and code releases are maintained by the corresponding authors at UIUC and are not affiliated with Meta.

Citation

@misc{chi2026spreadsheetrl,
  title         = {Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning},
  author        = {Banghao Chi and Yining Xie and Mingyuan Wu and Jingcheng Yang and Jize Jiang and Zhaoheng Li and Shengyi Qian and Minjia Zhang and Klara Nahrstedt and Rui Hou and Xiangjun Fan and Hanchao Yu},
  year          = {2026},
  eprint        = {2605.22642},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2605.22642},
  url           = {https://arxiv.org/abs/2605.22642}
}