source: arxiv artificial intelligence: deepslide: from artifacts to presentation delivery
level: research
most ai slide tools focus only on making slides look good. deepslide instead supports the whole presentation process. it helps with requirement gathering, time-budgeted narrative planning, evidence-grounded slide and script generation, attention augmentation, and rehearsal. the system uses a controllable logical-chain planner with per-node time budgets. a lightweight content-tree retriever grounds the content in sources. markov-style sequential rendering with style inheritance creates slides. sandboxed execution with minimal repair ensures slides render correctly.
the system is a human-in-the-loop multi-agent setup. it breaks down presentation preparation into steps that a person can guide and adjust. the planner builds a logical flow with time limits for each part. the retriever pulls in relevant evidence to back up claims. the renderer produces slides that follow a consistent style. if something fails to render, the system tries a minimal fix. this approach aims to improve not just the slide deck but also the pacing and narrative of the talk.
researchers tested deepslide across 20 domains. they used a dual-scoreboard benchmark that separately measures static artifact quality and dynamic delivery excellence. this benchmark helps show if a system makes good slides and also supports effective delivery. the results indicate that deepslide can produce presentations that are both visually sound and well-structured for speaking. the work highlights the gap between generating slides and preparing a full presentation, and offers a way to bridge it.
why it matters: it shifts ai presentation tools from slide creation to full delivery support, helping speakers prepare better talks with evidence and pacing.
source: arxiv artificial intelligence: deepslide: from artifacts to presentation delivery