The paper introduces PaperBanana, an advanced multi-agent framework designed to automate the creation of professional academic methodology diagrams and statistical plots. The system utilizes a specialized workflow involving retriever, planner, visualizer, and critic agents that collaborate to transform technical text into high-quality illustrations through iterative refinement. To rigorously measure performance, the researchers developed PaperBananaBench, a benchmark derived from modern AI publications that evaluates images based on faithfulness, conciseness, readability, and aesthetics. The study demonstrates that this agentic approach significantly outperforms standard generative models, producing results that align closely with human-drawn figures. Additionally, the documentation provides comprehensive style guides and evaluation protocols aimed at standardizing the visual language of scientific discovery. The framework ultimately seeks to democratize design resources for scientists while maintaining technical accuracy through human-in-the-loop oversight.

References:

• Zhu D, Meng R, Song Y, et al. PaperBanana: Automating Academic Illustration for AI Scientists[J]. arXiv preprint arXiv:2601.23265, 2026.