← News

Agents & MCPSources: arXiv · 20 Nov 2026

SheetAlign trains small open-weight models to automate spreadsheet edits

A research team reports a framework that lets locally deployed open-weight LLMs handle spreadsheet tasks by separating reasoning into distinct stages and rewarding edit progress.

In plain words

The authors present SheetAlign, a system that helps AI models edit spreadsheets automatically. It breaks each task into stages — where to act, what to change, and how to implement it — and trains the model using rewards for edit progress. On a public benchmark, it raised a small model's accuracy substantially.

What was announced

The framework, described in an arXiv paper by the SheetAlign authors, targets locally deployed open-weight LLMs so that spreadsheet analysis can be automated while sensitive data stays on-site and deployment costs stay controlled. Complex spreadsheet tasks demand both task reasoning and workbook understanding, and predefined operation interfaces add tool-selection and invocation requirements, the paper notes.

SheetAlign organizes policy learning into isolated stages covering where to operate, what to change, and how to implement transformations. Task-guided inspection and API grounding tie code generation to workbook observations and available operations, while stage-isolated learning optimizes program generation and review with edit-progress rewards. At inference, execution consistency picks programs based on agreement among resulting workbook states. The authors report that their grounded task interface lifted Qwen3-4B's average accuracy on SpreadsheetBench from 13.5% to 29.2% without training, and that with learning plus execution consistency SheetAlign reached 32.8% on SpreadsheetBench and 30.1% on Domain-Spreadsheet, outperforming comparable baselines. Code is to be released publicly upon publication.

What it does not show

  • The evidence is the paper's abstract only; full method details, experimental setup and peer review are not available in the supplied excerpt.
  • Reported accuracy figures are the authors' own benchmark results and have not been independently confirmed.
  • The code is not yet available; the paper says it will be released upon publication.

For security and IT teams

Scope
The framework is aimed at locally deployed open-weight LLMs for spreadsheet manipulation, with the stated goals of protecting sensitive data and controlling deployment costs.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.