PaperBanana: Completing the Final Piece of the AI Scientist Puzzle

In the transformative wave where artificial intelligence is reshaping the paradigm of scientific research, we've witnessed the remarkable potential of "AI Scientists" powered by Large Language Models (LLMs). They can read literature, propose hypotheses, write code, and even draft research papers. However, on the road to fully automated research workflows, there has always been an insurmountable "human bottleneck"—the creation of academic illustrations.

For any researcher, creating a figure that accurately conveys methodology while meeting publication-grade aesthetic standards often takes more time than writing the text itself. Today, we introduce PaperBanana (arXiv:2601.23265), the latest research published on arXiv, designed specifically to solve this pain point. It's not just a drawing tool—it's an intelligent agentic framework that marks a crucial step in AI-powered academic automation.

Background: The Visual Gap in Research Workflows

Before PaperBanana, automated research tools primarily focused on text generation and code execution. While existing generative AI (like DALL-E 3 or Midjourney) excels at artistic creation, they often fall short when facing Academic Illustrations. Academic drawing demands extreme rigor:

1. Logical Faithfulness: Must accurately reflect algorithm flows or model architectures without any "hallucinations."
2. Information Conciseness: Remove irrelevant decorations and focus on core information.
3. Layout Standards: Conform to academic journal formatting and aesthetic standards.

Addressing this challenging scenario, a research team from computer science and language computing proposed PaperBanana, aiming to give "AI Scientists" their own "personal illustrator."

Core Architecture: The Agentic Framework

PaperBanana's core innovation lies in its adoption of an Agentic Framework rather than relying solely on an end-to-end model. This design philosophy simulates how a human design team works, breaking down complex drawing tasks to be completed through collaboration between different "Specialized Agents."

The framework is primarily driven by state-of-the-art Vision Language Models (VLMs) and image generation models, executing tasks through precise orchestration.

Deep Dive: Four Core Workflows

According to the paper, PaperBanana's workflow contains four key steps, forming a closed-loop automated pipeline:

1. Retrieve References

Just like human designers need "reference images" before starting work, PaperBanana first retrieves relevant visual references through specialized agents. This step is crucial—it ensures that generated figures aren't created from thin air but are built upon existing visual semantics and academic conventions.

2. Plan Content and Style

After obtaining references, the system enters the planning phase. This step addresses "what to draw" and "how to draw it." Agents need to extract core methodological logic from complex paper text, design figure layouts, and determine visual styles that meet target publication requirements.

3. Render Images

This is the "execution" phase. The system calls image generation models to perform concrete rendering based on the blueprint determined in previous steps.

4. Iteratively Refine via Self-Critique

This is one of PaperBanana's most groundbreaking features. Unlike traditional "one-click generation," PaperBanana introduces a Self-Critique mechanism. This means the system can "review" its generated images, judge whether they accurately express the intended meaning, identify errors or aesthetic issues, and perform multiple rounds of iterative refinement. This feedback loop is key to generating high-quality, publication-ready images.

PaperBananaBench: Establishing a New Academic Illustration Benchmark

To scientifically evaluate PaperBanana's capabilities, the research team constructed a brand-new benchmark—PaperBananaBench.

• Scale and Source: The dataset contains 292 test cases, all carefully selected from papers published at NeurIPS 2025 (Conference on Neural Information Processing Systems). As a top AI conference, NeurIPS paper illustrations represent the highest standards in current academia.
• Focus Area: The test set primarily focuses on Methodology Diagrams—the most complex, core, and difficult-to-automate type of figures in academic papers.
• Diversity: Covers multiple research domains and different illustration styles, ensuring comprehensive evaluation.

The establishment of this benchmark not only validates PaperBanana's capabilities but also provides standardized metrics for future research in this field.

Performance Evaluation: Excellence Across Four Dimensions

In comparative experiments with existing leading baseline models, PaperBanana demonstrated significant advantages. The research team defined four core dimensions for evaluating academic illustration quality:

1. Faithfulness

Generated figures can precisely recreate the algorithm logic and model structures from papers without deviating from original intent. This is the baseline for academic illustration.

2. Conciseness

Figure designs remove redundancy and convey information efficiently. This aligns with academia's "Occam's Razor" principle.

3. Readability

Text, lines, and layouts in figures are clear and easy to read, helping readers understand complex concepts.

4. Aesthetics

Visual effects are professional and beautiful, directly meeting publication requirements (Publication-ready).

Furthermore, the research specifically notes that PaperBanana's capabilities aren't limited to flowcharts or architecture diagrams—it can effectively extend to generating High-quality Statistical Plots. This means it covers virtually all visual material needs in research papers.

Implications and Future: The Future of Automated Research

PaperBanana's release is not just a technical breakthrough—it has far-reaching industry implications:

• Liberating Research Productivity: It frees researchers from tedious drawing work, allowing them to focus more on core intellectual innovation.
• Empowering AI Scientists: For fully automated AI scientist systems, PaperBanana fills the visual output gap, giving them the potential to independently write and format complete papers.
• Standardization and Accessibility: By automatically generating figures that meet standards of top conferences like NeurIPS, it helps raise the average visual quality of papers across academia, making knowledge dissemination more efficient.

Conclusion

PaperBanana successfully tackles the challenging domain of academic illustration through clever combination of VLMs, image generation technology, and agentic workflows. It not only outperforms baseline models on PaperBananaBench but also demonstrates the possibility of achieving high-quality generation through "self-critique."

As the paper states, PaperBanana truly paves the way for automatic generation of publication-ready academic illustrations. For every researcher who has struggled late into the night adjusting PPT or Visio lines, PaperBanana is undoubtedly a ray of light before dawn.

---

Ready to experience PaperBanana?

• Explore our features
• View pricing plans
• Join the waitlist
• Read the FAQ