Automating Research Diagrams With an Agentic AI Framework

Google has developed PaperBanana, an AI framework that can generate publication-ready research figures from natural language descriptions, addressing the frustration researchers face in creating visualizations.

Details

PaperBanana is Google's solution to the common frustration researchers face when trying to create publication-ready figures from their data. The framework uses an agentic architecture with four agents that work together to generate figures that meet strict publication standards, rather than just producing rough drafts that require extensive manual cleanup. The Planner agent interprets the natural language description, selects the appropriate chart type, and identifies necessary data transformations. The Code Generator then translates this specification into executable code, which the Executor runs in a sandbox to capture the output and any errors. The Critic agent evaluates the generated figure against publication standards and provides structured feedback with severity ratings and suggested fixes. This critic-generator loop continues until the figure quality threshold is met, typically in 3-5 iterations. The agentic architecture and structured feedback approach are key innovations that distinguish PaperBanana from previous 'natural language to chart' tools, which often failed to bridge the gap between a plausible first attempt and a publication-ready figure.