Automated Data Visualization: PlotGen Uses Multi-Agent LLMs for Precise Scientific Charts

The visualization of scientific data is essential for transforming raw data into understandable visual representations. It enables the identification of patterns, the creation of forecasts, and the presentation of data-driven insights. However, inexperienced users in particular often face challenges, as the selection of suitable tools and the mastery of visualization techniques can be complex.

Large Language Models (LLMs) have recently demonstrated their potential in supporting code generation, but they often struggle with accuracy and require iterative debugging.

A new approach to addressing these challenges is PlotGen, an innovative multi-agent framework that aims to automate the creation of precise scientific visualizations. PlotGen orchestrates multiple LLM-based agents that collaborate to solve complex visualization tasks.

The Architecture of PlotGen: An Interplay of Specialized Agents

At the heart of PlotGen is the division of labor between specialized agents. A so-called "Query Planning Agent" breaks down complex user requests into individual, executable steps. A "Code Generation Agent" then translates pseudocode into executable Python code that creates the visualization. The real innovation of PlotGen, however, lies in the integration of three feedback agents, which are responsible for the iterative refinement of the generated diagrams.

These feedback agents – a "Numeric Feedback Agent," a "Lexical Feedback Agent," and a "Visual Feedback Agent" – utilize multimodal LLMs to improve the data accuracy, text labels, and visual correctness of the plots through self-reflection. The Numeric Feedback Agent checks the numerical consistency of the data in the diagram. The Lexical Feedback Agent optimizes the text components, such as axis labels and titles, to ensure clarity. The Visual Feedback Agent evaluates the overall appearance of the diagram and suggests adjustments to optimize the visual representation. Through this iterative feedback system, PlotGen achieves high precision and reduces the need for manual debugging.

Convincing Results Compared to Established Methods

Extensive experiments with the MatPlotLib dataset show that PlotGen outperforms strong baseline models. The results demonstrate an improvement of 4-6 percent compared to established methods. This leads to increased user confidence in LLM-generated visualizations and increased productivity, especially for inexperienced users, as the time spent debugging errors in the plots is reduced.

Future Prospects and Potential of PlotGen

PlotGen represents a promising step towards automated data visualization. The combination of specialized agents and multimodal feedback enables the creation of precise and meaningful diagrams. Future research could focus on expanding the functionality of PlotGen, for example, to enable interactive visualizations or support for other programming languages. The potential of PlotGen lies in simplifying complex visualization tasks and empowering users to analyze and present data more effectively.

Bibliography: - https://www.arxiv.org/abs/2502.00988 - https://www.arxiv.org/pdf/2502.00988 - http://paperreading.club/page?id=281463 - https://github.com/tmgthb/Autonomous-Agents - https://aclanthology.org/2024.findings-acl.701.pdf - https://www.researchgate.net/publication/375063078_Generative_Agents_Interactive_Simulacra_of_Human_Behavior - https://paperswithcode.com/paper/matplotagent-method-and-evaluation-for-llm - https://www.researchgate.net/scientific-contributions/Varun-Manjunatha-2107363883 - http://128.84.21.203/list/cs/pastweek?show=1926 - https://github.com/bowen-upenn/Agent_Rationality