PaperBanana: Five-Agent AI System Automates Publication-Ready Diagrams and Plots
Google AI and Peking University unveil PaperBanana, a five-agent agentic framework that turns raw methodology text into publication-ready diagrams and plots, combining a Linear Planning phase (Retriever, Planner, Stylist) with a 3-round Iterative Refinement (Visualizer, Critic); diagrams use image models and plots are produced as executable Python Matplotlib code to ensure data fidelity, and the system outperforms baselines on the NeurIPS 2025 PaperBananaBench in overall score, conciseness, readability, and aesthetics.
