SlideLab is a training‑free multi‑agent framework that generates complete scientific presentation decks directly from research papers. The system first designs a narrative outline, then employs four specialized agents: a content‑planning agent extracts key concepts and structures sections; a visual‑generation agent creates figures or charts from the text; a layout‑refinement agent arranges text and images for readability; and a grounding‑verification agent checks citations, data, and graphics for consistency. All agents iteratively refine a shared slide deck, eliminating error propagation typical of pipeline approaches.
In a blind human preference study, SlideLab was favored over both open‑source and commercial solutions on 77% of papers while consuming roughly one‑quarter the inference tokens of the strongest open‑source baseline. To assess slide quality, the authors introduce ConfArena, an audience‑oriented evaluation framework that simulates a conference room and scores presentations slide by slide. ConfArena’s rankings align closely with human judgments and reliably detect injected faults such as falsified numbers, degraded figures, missing slides, and shuffled order.
Review: SlideLab demonstrates that coordinated multi‑agent collaboration can achieve end‑to‑end paper‑to‑presentation automation, balancing content fidelity and visual design while markedly reducing computational cost, making it a valuable tool for academic communication.