New Benchmark for Audience-Aware AI Slide Generation

A novel benchmark called X+Slides has been developed to address a critical gap in evaluating AI-driven slide generation: the ability to tailor content for different audiences. Existing benchmarks primarily focus on completeness and technical depth, often overlooking the crucial requirement for presentations to resonate with specific groups, such as specialists needing detailed proofs versus decision-makers requiring actionable summaries. X+Slides is built upon a diverse dataset of 113 topics and seven presentation scenarios, utilizing a dynamic evaluation framework with over 8,000 source-grounded probes. It introduces four key metrics: Audience Coverage, which quantifies essential information conveyed; Domain-wise Coverage, indicating types of information covered; Efficiency, measuring utility per attention cost; and Correctness, verifying source support for claims. Experiments with models like DeepPresenter, SlideTailor, and NotebookLM show that while these systems can recover a substantial portion of audience-essential information, they still fall short of complete audience-specific tailoring. For instance, DeepPresenter achieved an Audience Coverage of 0.714, highlighting that visual quality and broad topic coverage alone are insufficient without robust source-grounded and audience-aware evaluation.