PaperBanana: How AI Is Reshaping Research Visualisation in Higher Education

What if the most labor-intensive part of creating a research manuscript, such as creating publication-ready figures, could be accomplished in a matter of minutes? Such is the promise of PaperBanana. Created by researchers from Peking University and Google Cloud AI Research, PaperBanana is an AI-based, agentic system that automatically converts scientific text into professional-quality academic figures. Such a breakthrough for students, researchers, and academics in higher education is a paradigm shift in the visualization of scientific knowledge.

In the midst of the rapid evolution of AI-based tools in higher education, PaperBanana is a very specific tool for research visualization. Unlike most AI-based tools, PaperBanana does not employ a single AI agent to accomplish a specific task. Instead, the system employs a cohort of five AI agents to accomplish the tasks of reference gathering, planning, style adjustment, rendering, and self-evaluation of the output. Such figures are of the quality of, if not superior to, those created by human experts.

Key Takeaway: PaperBanana achieved a 72.7% win rate against baseline AI methods in blind human evaluation, with its figures rated superior in conciseness, readability and aesthetics (Zhu et al., 2026).

What Is PaperBanana? The AI Framework Automating Academic Illustrations

PaperBanana is an open-source agent-based system that is developed to facilitate the automation of the production of publication-standard illustrations for academic publications. The system was developed by Dawei Zhu et al. in January 2026 and is available on arXiv under the identifier 2601.23265. The system is a solution to the widely recognized problem of the time-consuming production of methodology illustrations and statistical figures for academic publications.

How Does It Work?

Unlike most image generation systems, the PaperBanana system is based on a multi-agent approach that is supported by the latest vision and language technology. The system operates in two different modes:

Phase 1 – Linear Planning: In this phase, three specific agents are involved.

• Retriever Agent: The Retriever Agent uses a reference dataset to find relevant images that resemble the needed target.
• Planner Agent: The Planner Agent uses the methodology text to create a detailed specification of the needed target.
• Stylist Agent: The Stylist Agent uses academic design principles that are automatically generated from the analysis of hundreds of published images.

Phase 2 – Iterative Refinement: In this phase, two agents are involved.

• Visualiser Agent: The Visualiser Agent uses image generation models to create the image.
• Critic Agent: The Critic Agent uses the original source text to assess errors and give corrective feedback. This process is repeated three times to improve image quality.

Benchmark: PaperBananaBench

The researchers created a benchmark called PaperBananaBench to test this framework. This benchmark consists of 292 test cases that are created from NeurIPS publications from 2025. This is one of the first tests created to assess AI-generated diagrams for academic papers. This test covers four different domains: Agent & Reasoning, Vision & Perception, Generative & Learning, and Science & Applications.

How PaperBanana Is Transforming Higher Education

The implications of the use of artificial intelligence for scientific investigation go beyond the realm of convenience. “PaperBanana” represents a new category of scientific visualization software that could fundamentally change the way knowledge is disseminated throughout the entire academic community.

For Students: Overcoming Barriers to Professional Research Communication

For graduate students, it can take weeks to develop the figures required for theses and publications. Students often do not have the luxury of using professional design software or training in the field. Using “PaperBanana,” a student can enter a description of his/her methodology and create a professionally designed diagram with modern color schemes, layouts, and typography appropriate for the academic community, which could otherwise require extensive manual labor using software such as Adobe Illustrator or PowerPoint.

Did You Know? PaperBanana scored 80.7% in conciseness – significantly outperforming even human experts (50.0%) – meaning it produces cleaner, more focused diagrams with less visual clutter.

For Researchers: Expediting the Publication Process

Active researchers continually face the pressure of publishing, which can cause delays in the authoring process due to issues in the creation of figures. PaperBanana’s prototyping capability allows researchers to create multiple visual representations of the same methodologies, communicate effectively, and conduct various experiments. It also includes the generation of statistical plots, which can be done by producing Python Matplotlib codes to ensure numerical precision in the data-based images created by the researchers.

For Faculty: Advancing Teaching and Supervision

Researchers can utilize PaperBanana as a tool to aid their students in their research and as a means to teach them the principles of efficient scientific communication. By comparing the images created by the students and the images created by the AI, the students can be taught the principles of efficient scientific communication. PaperBanana can also be used to improve the images created by the students, as it can enhance the images created by the students by increasing their aesthetic appeal by up to 56.2% in 292 cases, as demonstrated by the framework’s performance in the test cases.

Applications in Scholarly Practice: Demonstrated Utility of PaperBanana

Enhancements to the Research Workflow

Researchers can utilize PaperBanana as it can be integrated into the research writing process. The researchers can input the text of the methodology section of the research and the corresponding figure’s caption, and the framework can automatically retrieve the necessary reference styles, plan the content of the images, follow the academic design standards, and create the images, all of which can be done without the researchers having to know the principles of design.

Scientific Diagram Generation

The framework can create images of methodologies, which can be done by creating complex diagrams, architectural diagrams, pipeline diagrams, and other diagrams necessary in the publication of AI and computer science-related research. PaperBananaBench demonstrated that the framework can achieve a score of 60.2, which is higher than the human reference score of 50.0. The framework performed the best in the creation of Agent and Reasoning diagrams, which can achieve up to 69.9%, while the lowest performance was seen in the creation of Vision and Perception diagrams, which can achieve up to 52.1%.

PaperBanana performance compared to human experts across five quality dimensions (data from PaperBananaBench; Zhu et al., 2026)

Academic Publication Enhancement

In addition to the creation of new figures, PaperBanana also has the ability to improve existing diagrams. This is because the aesthetic guidelines provided by the Stylist Agent could be used to improve the diagrams created manually. From the results of the empirical tests conducted, it was evident that the enhancement of the diagrams made them look better in 56.2% of cases. This is an indication that the tool could be useful in providing researchers with AI assistance in their work.

The Future of AI in Research and Universities

Emerging Trends in AI Tools for Academic Research

PaperBanana is part of an emerging sector of AI tools for academic research that are transforming universities around the world. From AI writing assistants to literature review tools, the academic research landscape is increasingly being augmented by AI tools. The agent-based approach, which uses several specialized AI agents to work together on tasks, is at the cutting edge of this revolution.

The future of AI in universities is therefore expected to be characterized by these tools becoming an integral part of the research landscape, similar to the current use of statistical software or reference managers. However, their effective integration into academic research must be informed by clear policies.

Ethical Considerations: Responsible AI in Academia

The PaperBanana team recognizes that one of the biggest risks in the use of AI tools in research is “visual hallucination,” in which illustrations created by the tool may include details that are not supported by the source material. The most common errors include “fine-grained connectivity” errors, which include “misaligned arrows” as well as “flow direction” errors (Zhu et al., 2026).

This is an indication that there is a need for researchers to be cautious when using AI tools in their work. The fact that the most common errors are those involving “fine-grained connectivity” is an indication that AI visualization tools must be used in collaboration with human experts.

Important: PaperBanana’s faithfulness score (45.8%) remains below the human reference (50.0%). This means human review for factual accuracy is still essential before any AI-generated figure is published.

Innovation Opportunities

The PaperBanana framework is open-source and available on GitHub. This provides an opportunity for universities to explore, adapt, and extend this technology. Some opportunities that could be taken in the future include:

• Creating editable vector images instead of raster images
• Expanding the domain scope beyond computer science to other natural science, engineering, and social science disciplines
• Creating GUI agents that can be used to run professional computer design software like Adobe Illustrator
• Creating hybrid tools that combine the power of code-driven accuracy in data visualization with AI-driven image creation for conceptual diagrams

Gulf University and AI Innovation: Leading the Digital Revolution in Education

Gulf University understands that artificial intelligence for students and researchers is no longer a promise of the future, but rather an undeniable reality. As the world over tries to figure out the best way to incorporate artificial intelligence into academic scholarship, Gulf University is committed to leading the digital revolution in education.

Innovative research frameworks like PaperBanana are the opportunities that our academic community can benefit from as well as contribute to. Gulf University can assess these new AI technologies, create guidelines on their responsible use, and be at the forefront of the digital revolution in higher education.

We thus strongly encourage our faculty, researchers, and students to explore the opportunities offered by AI-powered research visualization tools, critically, responsibly, and with the utmost commitment to academic integrity.

Join the Conversation: What Do You Think?

The emergence of AI in academic research poses interesting questions, which we think it’s essential to discuss within our academic community:

• Do you spend more hours creating figures than actually researching? How can AI help you be more efficient?
• Should there be a disclosure when AI-powered tools are used to create figures in theses or published works?
• How can we guarantee the accuracy of AI-powered figures?
• How can we guarantee that our students are developing the essential skill of data visualization, despite the emergence of AI-powered automation?

Your input matters. We would love to hear from you. You can share your thoughts in the comments section, or get in touch with the Gulf University Research & Innovation Office to discuss the opportunities offered by AI-powered tools to help you achieve your research objectives.

Take the Next Step: Explore AI-Powered Research Tools

The future of scientific communication is being written today. Whether you’re a student embarking on your very first research project, a faculty member helping shape the next generation of academics, or an established researcher pushing the frontiers of knowledge, AI-powered research tools like PaperBanana can help you improve the quality of your work.

Ready to explore? Visit the PaperBanana project page at https://dwzhu-pku.github.io/PaperBanana/ to learn more about the framework and access the open-source code. Stay tuned for upcoming Gulf University workshops on AI tools for research.

At Gulf University, innovation happens through inquiry. This article advocates for advancing the future of artificial intelligence in higher education through a framework that incorporates responsibility, ethical considerations, and academic standards of excellence that are characteristic of high-quality scholarship.

References and Further Reading

• Zhu, D., Meng, R., Song, Y., Wei, X., Li, S., Pfister, T. and Yoon, J. (2026), “PaperBanana: automating academic illustration for AI scientists”, arXiv preprint arXiv:2601.23265.
• PaperBanana Project Page: https://dwzhu-pku.github.io/PaperBanana/
• PaperBanana.org: https://paperbanana.org/
• Quantum Zeitgeist (2026), “PaperBanana shows automated illustration generation”, https://quantumzeitgeist.com/292-paperbanana-shows-automated-illustration-generation/