PPTBench: 500 scientific diagrams turned into editable PPTX, Astra tops out at 77.34

Queen Isabell @Queen_1o1Originally posted 2026.09.29

Open original on X

Turning a diagram into clean, editable code sounds simple until you try it. PPTBench tests this with 500 scientific diagrams reconstructed as editable PPTX files. Even the top model, GPT-6 Astra High, reaches only 77.34/100. A tiny detail like a flipped arrow can change the meaning—faithful design-to-code reconstruction still has a long way to go.

This post quotes another post: x.com/einsiaai/status/2104941770433814889