LightVA: Lightweight Visual Analytics with LLM Agent-Based Task Planning and Execution

Yuheng Zhao

Yuheng Zhao1

Junjie Wang

Junjie Wang1

Linbin Xiang

Linbin Xiang1

Xiaowen Zhang

Xiaowen Zhang1

Zifei Guo

Zifei Guo1

Cagatay Turkay

Cagatay Turkay2

Yu Zhang

Yu Zhang3

Siming Chen

Siming Chen1

1 Fudan University · 2 University of Warwick · 3 University of Oxford

IEEE Transactions on Visualization and Computer Graphics (2024)

Paper Appendix

Abstract

LightVA introduces a lightweight visual analytics framework that supports LLM agent-based task planning and execution for human–AI collaborative data analysis. It enables users to explore datasets through an adaptive and recursive workflow involving a planner, executor, and controller. LightVA reduces the complexity of developing and using VA systems by dynamically generating tasks, visualizations, and insights. The system demonstrates efficiency and interpretability through a usage scenario and expert evaluation, advancing human–AI collaboration in visual analytics.

Framework

The LightVA framework connects human goals, analytical tasks, and visual insights through an LLM agent-based recursive workflow. It includes three roles: a planner for task recommendation and decomposition, an executor for task execution and visualization generation, and a controller that manages their coordination. This design reduces human effort and supports adaptive task-driven visual exploration.

LightVA Framework Diagram

System

The LightVA system integrates an agent-based task planning pipeline with an interactive interface that includes four main views: Chat View, Visualization View, Task Flow View, and Data View. Users interact with the LLM agent via natural language, exploring recommended or user-defined tasks while the system automatically generates visualizations and insights. Linked-view exploration enables dynamic coordination among multiple charts for richer data understanding.

LightVA System Interface

Demo Video

BibTeX

@article{Zhao2025LightVA,
  title={LightVA: Lightweight Visual Analytics with LLM Agent-Based Task Planning and Execution},
  author={Zhao, Yuheng and Wang, Junjie and Xiang, Linbin and Zhang, Xiaowen and Guo, Zifei and Turkay, Cagatay and Zhang, Yu and Chen, Siming},
  journal={IEEE Transactions on Visualization and Computer Graphics (2024)},
  year={2025},
  url={https://zyh1222.github.io/LightVA/}
}