ChartEditBench: Evaluating Grounded Multi-Turn Chart Editing in Multimodal Language Models
#Multimodal Large Language Models #MLLMs #ChartEditBench #multi‑turn chart editing #visual grounding #exploratory data analysis #benchmark #human‑in‑the‑loop visualization
📌 Key Takeaways
- Multimodal LLMs perform well on single‑turn chart generation but their capability in multi‑turn interactive data analysis is underexplored.
- Real‑world data analysis often involves iterative visual refinement, demanding consistent common ground, edit tracking, and adaptation to user preferences.
- ChartEditBench provides a structured, incremental evaluation framework for visually grounded chart editing tasks.
- The benchmark includes datasets, tasks, and metrics that reflect real‑world multi‑turn chart editing scenarios.
- ChartEditBench aims to spur research into more robust MLLMs that can support iterative data exploration and visualization.
📖 Full Retelling
The authors, publishing on arXiv in February 2026, introduce ChartEditBench—a new benchmark designed to evaluate how multimodal large language models (MLLMs) handle iterative, visually grounded chart editing. The benchmark addresses the gap in current research where MLLMs excel at single‑turn chart generation but struggle to support real‑world, multi‑turn exploratory data analysis that requires maintaining common ground, tracking prior edits, and adapting to evolving user preferences.
🏷️ Themes
Multimodal artificial intelligence, Data visualization, Iterative human‑machine interaction, Benchmark development, Exploratory data analysis
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
arXiv:2602.15758v1 Announce Type: cross
Abstract: While Multimodal Large Language Models (MLLMs) perform strongly on single-turn chart generation, their ability to support real-world exploratory data analysis remains underexplored. In practice, users iteratively refine visualizations through multi-turn interactions that require maintaining common ground, tracking prior edits, and adapting to evolving preferences. We introduce ChartEditBench, a benchmark for incremental, visually grounded chart
Read full article at source