VisEditBench: Can Vision-Language Models Edit Visualization Code from Multimodal Feedback?

2026-08-10

AI Summary

Existing benchmarks lack evaluation of visualization code editing capabilities driven by multimodal feedback, such as chart images paired with textual instructions. To address this gap, this work proposes VisEditBench—the first multimodal benchmark tailored for this task—comprising 1,395 human-annotated editing samples that reflect real-world scenarios involving feedback-guided repair and reference-guided redrawing. Additionally, we introduce VisEditAgent, a framework that integrates vision-language models, code execution validation, rendering comparison, and iterative refinement to enable render-driven closed-loop editing. Experimental results reveal that current state-of-the-art vision-language models struggle with visually aligned tasks like style transfer, with the best-performing model, Claude-4.6-Sonnet, achieving a pass rate of 74.46%. Notably, VisEditAgent significantly improves GPT-4o’s pass rate from 55.75% to 67.99%.

Abstract

Vision-language models (VLMs) have shown strong capabilities in generating visualization code from textual or visual specifications. However, real-world visualization authoring is inherently iterative: users frequently revise existing visualizations to repair flawed charts or adapt them to desired styles. Existing benchmarks primarily evaluate generation from scratch, leaving visualization code editing from multimodal feedback largely unexplored. We introduce VisEditBench, a benchmark of 1,395 human-annotated visualization code-editing tasks grounded in realistic visualization workflows and failure cases. VisEditBench covers two practical settings: feedback-guided repair, where models revise visualization code using buggy or marked charts together with textual feedback, and reference-guided restyling, where models modify code to match a target chart image. Evaluating 20 state-of-the-art VLMs reveals that visualization code editing remains challenging: Claude-4.6-Sonnet achieves the best overall pass rate of 74.46%, while most open-source models remain below 50%. Performance is particularly weak on visually grounded style adaptation, where Claude-4.6-Sonnet achieves only 55.71%. To establish a strong baseline, we further propose VisEditAgent, a render-grounded editing framework that iteratively generates, executes, validates, and refines candidate edits. Built on GPT-4o, VisEditAgent improves overall pass rate from 55.75% to 67.99%, demonstrating the importance of render-grounded feedback for faithful visualization editing. We will release VisEditBench at https://github.com/vis-nlp/VisEditBench.

Problem

• visualization code editing
• vision-language models
• multimodal feedback
• iterative visualization authoring
• style adaptation

Innovation

• visualization code editing
• vision-language models
• multimodal feedback
• render-grounded editing
• VisEditBench