PPTBench Evaluates Coding Agents on Editable Slide Reconstruction

PPTBench introduces a 500-task benchmark for testing if coding agents can reconstruct visual information into structured, editable slide formats. It requires agents to bridge the gap between visual inference and programmatic expression using scientific diagrams as ground truth.