Narrative-Driven Paper-to-Slide Generation via ArcDeck
New framework uses discourse trees and agent coordination to preserve a paper's logical narrative flow.
A research team from Georgia Tech and KAIST has unveiled ArcDeck, a novel AI framework that tackles the complex task of converting dense academic papers into coherent presentation slides. Published on arXiv, the system reframes the problem as one of narrative reconstruction rather than simple summarization. Its key innovation is an initial parsing stage that builds a discourse tree—a structural map of the paper's logical flow—and establishes a 'global commitment document' to lock in the author's core intent. This ensures the high-level argument isn't lost in translation.
ArcDeck then employs a multi-agent refinement process, where specialized AI agents take on distinct roles to iteratively critique and revise a presentation outline. This collaborative, role-specific coordination is designed to enhance narrative flow and logical coherence far beyond what direct text-to-slide models can achieve. To rigorously evaluate their approach, the team also introduced ArcBench, a newly curated benchmark of academic paper-slide pairs, providing a standardized testbed for future research in this domain.
- ArcDeck models a paper's logical flow using discourse trees and a 'global commitment document' before generating slides.
- It uses a multi-agent system where specialized AI agents iteratively critique and refine the presentation outline for better coherence.
- The team released ArcBench, a new benchmark dataset, and results show their method significantly improves narrative flow over existing approaches.
Why It Matters
Automates a tedious but critical academic and professional task, potentially saving researchers hours while improving presentation quality.