ChartEditor: A Reinforcement Learning Framework for Robust Chart Editing

TL;DR

This work introduces ChartEditVista, a large-scale, automated benchmark for chart editing comprising $7{,}964$ samples across $31$ chart types, with inputs of chart images and natural-language edits and outputs as edited code $C'$. It addresses realism gaps in prior work by removing code-input assumptions and by proposing two fine-grained metrics, the layout and text metrics, to better capture visual and textual edits. Building on ChartEditVista, the authors present ChartEditor, a 3B open-source model trained with a two-stage supervised fine-tuning curriculum and a reinforcement-learning framework called Group Relative Policy Optimization (GRPO) that uses a novel rendering reward to jointly optimize code executability and visual fidelity. Experiments show ChartEditor achieving state-of-the-art performance on ChartEditVista and strong generalization to out-of-domain benchmarks, demonstrating the approach’s practical impact for robust, scalable chart editing with cross-modal alignment.

Abstract

Chart editing reduces manual effort in visualization design. Typical benchmarks limited in data diversity and assume access to complete chart code, which is seldom in real-world scenarios. To address this gap, we present ChartEditVista, a comprehensive benchmark consisting of 7,964 samples spanning 31 chart categories. It encompasses diverse editing instructions and covers nearly all editable chart elements. The inputs in ChartEditVista include only the original chart image and natural language editing instructions, without the original chart codes. ChartEditVista is generated through a fully automated pipeline that produces, edits, and verifies charts, ensuring high-quality chart editing data. Besides, we introduce two novel fine-grained, rule-based evaluation metrics: the layout metric, which evaluates the position, size and color of graphical components; and the text metric, which jointly assesses textual content and font styling. Building on top of ChartEditVista, we present ChartEditor, a model trained using a reinforcement learning framework that incorporates a novel rendering reward to simultaneously enforce code executability and visual fidelity. Through extensive experiments and human evaluations, we demonstrate that ChartEditVista provides a robust evaluation, while ChartEditor consistently outperforms models with similar-scale and larger-scale on chart editing tasks.

Figures (4)

• Figure 1: Comparisons on ChartEditor and other SOTA models, red dots represents ChartEditor and bule represent others.
• Figure 2: Overview of our fully data construction pipeline, it can generate chart editing data and refine them in various aspects. The gray arrows show the data flow inside each stage, the dark blue arrows show the data flow between stages and the red arrows show the data quality control steps.
• Figure 3: Limitations of existing metrics, while MLLM-Based metric suffers from serious hallucination, rule-based metric can not capture the fine-grained visual changes.
• Figure 4: Overview of ChartEditor’s training procedure. ChartEditor is optimized within a GRPO-based reinforcement learning framework, guided by format, execution, and rendering rewards. To further stabilize learning signals, ChartEditor employs a multi-stage training schema that incorporates supervised fine-tuning for cold-start and curriculum RL strategies.