GDP.xlsx
Test whether agents can reason through professional spreadsheets.
Overview
GDP.xlsx from Surge AI tests whether an AI agent can read a professional Excel workbook the way a domain expert would: following data across tabs, noticing formatting that carries meaning, and applying rules written elsewhere in the file. Surge's position is that real spreadsheets behave more like small software systems than tables, so answering a simple-sounding question can require reconstructing a workbook's logic.
What a task looks like
Each task pairs a short, colleague-style workplace request with a spreadsheet or CSV file, and the agent explores it with tools inside a Linux container. In Surge's published example, a product-certification tracker makes two products look ready for production, but a rule on another tab allows a pending certification only for one manufacturer, so only one product qualifies.
How it’s scored
An LLM judge (Gemini 3.8 Flash in the reference harness) checks the agent's final response against task-specific must-have rubric criteria, each scored pass or fail, and the trial reward is the mean of those verdicts. The reference protocol runs 5 attempts per task, averages them per task, and reports the unweighted mean across tasks.
- Released
- Sep 30, 2026
- Size
- 70 tasks (58 Excel files and 12 CSV files, 393 worksheets in total); 3 sample tasks are public.
- Access
- Partly private: only 3 sample tasks are published on Hugging Face (Apache-2.0), and the evaluation harness is on GitHub (Apache-2.0); Surge asks readers to contact it for access to the full benchmark. Tasks were written by domain experts and passed three rounds of review.
Summarized by Stubbi from the public sources listed below. Stubbi has not run this benchmark.
Published leaderboard
Explore scores and evaluation settings on the publisher’s leaderboard.
Open leaderboardCoverage
70 workbook tasks across 12 domains, including finance, real estate, healthcare, and engineering.
Methodology
Agents inspect worksheets, formulas, formatting, comments, and cross-sheet dependencies. Expert-written tasks are graded against required rubric criteria after three rounds of review.