X+Slides

X+Slides benchmark evaluates audience-conditioned slide generation from documents

Researchers introduce X+Slides, a benchmark for evaluating LLM-based slide deck generation that incorporates target audience as a first-class evaluation dimension. Built on 113 topics and seven presentation scenes, it uses 8,133 source-grounded probes and four metrics covering audience coverage, domain coverage, efficiency, and correctness. Experiments on DeepPresenter, SlideTailor, and NotebookLM reveal that current systems recover substantial but incomplete audience-essential information, with NotebookLM achieving the highest audience coverage of 0.853 at the tested threshold.