PaperBanana: Google teaches AI to draw charts that won't embarrass you in front of reviewers

Google and Peking University have introduced PaperBanana, a multi-agent framework that automates the creation of methodology diagrams and statistical charts for scientific publications. MarkTechPost reported the development.

The problem with scientific visualization

A scientific chart has to be accurate, scalable, and compliant with strict publication standards. General-purpose generative models like DALL-E aren't suited to the job: they hallucinate data and don't understand the logic of an experiment. Until now, automation in science has mostly covered text and computation — neural networks have learned to write literature reviews, propose hypotheses, and write code — but visualizing results remained manual work.

How PaperBanana works

Instead of handing the drawing to a single model, PaperBanana splits the task among specialized agents: one plans the diagram's structure, a second writes the code to render it, and a third critiques the result, checking whether the axes are mixed up and whether the fonts are legible. The hierarchy mimics how a small design studio operates.

Context: tools for AI research

Projects like Sakana AI's AI Scientist have already shown that a neural network can take research from idea to draft paper, but those efforts stumbled at the visualization stage. Google, which competes with OpenAI and Anthropic on tools for developers and scientists, is closing that gap with a framework aimed at an academic audience.

Open questions on ethics and quality

If a paper is entirely AI-generated, including its data visualizations, how are reviewers supposed to verify that the data is genuine? PaperBanana builds its charts from real numbers, but the sheer ease of producing them could trigger a flood of low-quality, "nicely packaged" publications.

What does PaperBanana do?

PaperBanana is a multi-agent framework from Google and Peking University that automates the creation of methodology diagrams and statistical charts for scientific publications.

Why aren't general-purpose generative models like DALL-E suitable for

scientific charts?

They hallucinate data and don't understand the logic of an experiment, whereas a scientific chart has to be accurate, scalable, and compliant with strict publication standards.

How are roles divided among the agents in PaperBanana?

One agent plans the diagram's structure, a second writes the code to render it, and a third critiques the result, checking whether the axes are mixed up and whether the fonts are legible.