Reference Methodology

Researchers found that showing the model any good diagram teaches it structure better than fine-tuning or detailed prompting.

The PaperBanana team experimented with fine-tuning versus simply retrieving and showing reference examples. Result: showing the model good examples taught structure more effectively than anything else. The model learns layout, spacing, typography patterns, and visual hierarchy from examples in a way that instruction prompts can't replicate. This is why the retrieval stage exists — it gives the model a visual vocabulary to work from.

Use rich, data-driven visual publications like Visual Capitalist as reference material — they communicate complex information purely through visuals.

Visual Capitalist is an entire website where news is communicated purely in visuals — no essays, no paragraphs, just information design. "Ranked: The Jobs Most Exposed to GenAI" becomes a beautiful infographic. This style — data-dense, visually clear, professionally polished — is exactly what the agent team pipeline excels at reverse engineering and recreating for new topics.

The goal isn't to copy — it's to extract the visual DNA from a reference image and apply it to a brand new scenario.

Take an existing image — say, a Visual Capitalist infographic about global coffee consumption with beautifully designed cups themed like national flags. Feed it to your research agent. The agent extracts: color palette, composition structure, typography hierarchy, visual metaphors, data presentation style. Then apply that extracted style to a completely new topic — AI investment by country, for example. Same visual language, different data.