AI agents take over academic paper diagrams

Anyone who has ever rushed a paper submission knows the pain: writing the actual text takes days, but fine-tuning that single methodology diagram to look clean and professional can burn an entire week.

Researchers from Peking University and Google Cloud AI just introduced PaperBanana, an agentic framework designed to automate the painful process of creating publication-ready academic illustrations and plots.

Five agents on a single canvas

Standard text-to-image models usually fail hard at scientific figures, producing messy color blocks and hallucinated gibberish text.

PaperBanana sidesteps this by mimicking a real design agency, distributing the work across five specialized AI agents: Retriever, Planner, Stylist, Visualizer, and Critic.

The Retriever first pulls relevant reference figures from a database. The Planner maps out components and logical flows based on the source text. The Stylist establishes modern color palettes and typography, while the Visualizer renders the image using foundation models like Nano Banana Pro or GPT-Image-1.5.

Finally, the Critic provides feedback in an iterative refinement loop to polish the output.

Beyond generating diagrams from raw text, the system can upgrade existing human-made figures. In direct aesthetic blind tests against original human drawings, PaperBanana won 56% of the time.

For statistical plots, it offers two distinct modes: an image-generation mode that prioritizes visual appeal, and a code-generation mode running Matplotlib to ensure zero numerical hallucinations.

Benchmark gains and tangled arrows

To test how well it performs, the team built PaperBananaBench, curated from 292 methodology diagrams from NeurIPS 2025 publications.

Across this benchmark, PaperBanana beat baseline models by 37% in conciseness and nearly 13% in readability, consistently winning on faithfulness and aesthetics.

Yet, the framework still has a glaring weak spot.

The most frequent failure mode involves edge connections, such as drawing redundant arrows or flipping source and target nodes.

Worse, the Critic agent frequently fails to catch these connectivity bugs, pointing to fundamental perception limitations still lingering in underlying vision-language models.

Why it matters

For academic researchers and technical writers, PaperBanana addresses one of the most tedious manual bottlenecks in the publishing pipeline. Being able to turn rough sketches or raw methodology text into clean visuals could save hours of manual tweaking in vector software.

We think the framework's core architectural choice is smart. Instead of relying on a black-box image model to get everything right at once, it decouples logical planning from aesthetic styling and smartly falls back to Python scripts whenever numerical precision is on the line.

However, users should not blindly trust the final output just yet. In academic diagrams, a misplaced arrow alters the entire scientific meaning. Because the system's own critique agent struggles to spot swapped nodes, human authors will still need to manually verify every single logical connection before submitting. The team plans to release the code and dataset within two weeks, which will be the real test of its everyday utility.