Got it! Here’s the story in English, around 500 words:

The Last Night Before Defense

Lin Xiao had been awake for forty-eight hours straight when the clock ticked past 2 a.m.

Her dissertation sat open on the screen — five years of research, three failed experiments, one unexpected breakthrough, and four chapters that her advisor had called “rigorous and compelling.” But Chapter Four had a problem. A red-ink problem. Her advisor’s note from last week still burned in her memory: “The methodology figure looks amateur. Not publication-ready.”

Defense was in seven hours.

She opened Adobe Illustrator and stared at a blank canvas. She wasn’t bad at design — she just couldn’t make five interconnected agents look like anything other than a kindergarten flowchart. Every time she got the arrows right, the labels became unreadable. Every time the labels were clear, the layout collapsed. An hour passed. Then another.

Her roommate, Priya, shuffled out of the bedroom in socks and a worn MIT hoodie. She took one look at Lin Xiao’s screen and said nothing for a moment.

“Try Paper Banana,” she finally said, filling a glass of water.

“What?”

“Paper banana dot app. I used it for my NeurIPS submission last month. Just paste your methods section in and let it run.” She shrugged. “Saved me a full day.”

Lin Xiao was skeptical. She had tried AI image generators before — they hallucinated connections between components that didn’t exist, produced charts with made-up numbers, and generated diagrams that looked impressive until a reviewer actually read them. Academic figures weren’t decoration. They were arguments.

But she had six hours left. Skepticism was a luxury she couldn’t afford.

She pasted her entire methodology section — six hundred words describing her multi-agent retrieval-augmented framework — into the text box and hit generate.

The system went to work. She watched the status bar cycle through its pipeline: Retrieving references… Planning layout… Applying style guidelines… Rendering… Critic review…

Three minutes and forty seconds later, an image appeared on her screen.

Lin Xiao leaned forward.

The diagram was clean. The five agents were arranged in a logical left-to-right flow, each labeled precisely with the terminology from her own paper. The color palette was muted and professional — the kind she’d seen in ICML proceedings. The arrows indicated directionality without cluttering the space. A feedback loop between the Critic and Visualizer agents curved elegantly back through the pipeline, exactly as her methodology described.

She zoomed in on the labels. Every word was correct. No hallucinated connections. No invented data.

She sat back and exhaled for what felt like the first time in days.

Priya leaned over her shoulder, squinting at the screen. “See? Told you.”

Lin Xiao exported the figure at full resolution, dropped it into her slide deck, and finally closed Illustrator — the empty canvas still untouched.

At 9:07 the next morning, the committee chair paused on that slide and said, “This is one of the clearest architecture diagrams I’ve seen in a defense this year.”

Lin Xiao smiled and said nothing about the three minutes and forty seconds it had taken to make it.