DV-World benchmarks data visualization agents in realistic settings. It includes 260 tasks across DV-Sheet, DV-Evolution, and DV-Interact, with hybrid evaluation of both values and visual quality.
DV-World evaluates data visualization agents in more realistic workflows.
Many existing benchmarks focus on simplified chart generation. DV-World tests agents in software-grounded settings that require editing spreadsheets, adapting reference visuals, and clarifying ambiguous requests.
DV-World covers three core settings: spreadsheet execution, visual adaptation, and interactive clarification.
Agents work directly in spreadsheets to create, repair, and organize visualizations.
Agents adapt reference visuals into correct outputs for new data and new frameworks.
Agents must ask the right questions before acting when user intent is unclear.
Each tab shows the main paper-reported score for the selected task.
Browse error cases by task instead of scanning one long mixed gallery.
Readability, structure, and data accuracy problems inside spreadsheet-native chart creation.
Labels and layout fight each other until the visual becomes difficult to read.
A creation case where both chart structure and encoded values drift away from the intended result.
Integrity failures, destructive regressions, and diagnostic blindness in targeted chart repair.
The intended fix leaks side effects into parts of the chart that should remain stable.
The repair spreads damage instead of containing it.
Cases where the layout exists, but the business framing, completeness, or design logic still fails.
The dashboard exists, but does not answer the business question well.
Visual composition and analytical logic are no longer aligned.
Wrong-vs-right library transfer examples that show whether visual semantics survive migration.
The side-by-side pair makes semantic drift visible immediately.
Great for showing how layout and semantics can break even in expressive libraries.
Ambiguity, avoidance, and inquiry failures under user-simulator pressure.
The agent proceeds instead of clarifying unresolved ambiguity.
The system fails to ask the questions it should have asked before acting.
Replace the anonymous metadata once the public version of the paper is finalized.
@article{dvworld2026,
title = {DV-World: Benchmarking Data Visualization Agents in Real-World Scenarios},
author = {Anonymous Authors},
journal = {arXiv preprint arXiv},
year = {2026},
url = {./8190_DV_World_Benchmarking_Dat.pdf}
}