ConvDeck Enables Conversational, Stage-Specific Slide Generation

Creating effective presentations from academic papers is an inherently iterative process, involving multiple cycles of generation, critique, and revision. While some multi-agent AI systems for slide generation incorporate internal critique loops, and conversational approaches allow for post-generation refinement, these methods often limit user involvement to the very end or keep the internal revision process opaque. This restricts a user's ability to influence crucial aspects like narrative flow, content emphasis, and overall structure early on. Researchers have developed ConvDeck, a new multi-agent pipeline designed to address this gap. ConvDeck integrates user interaction throughout the slide generation process by introducing stage-specific feedback loops. This allows users to iteratively refine both the presentation outline and the final slide deck at the precise stages where these decisions are made. The core of ConvDeck's refinement mechanism involves agents that can "think, speak, and act." This enables them to either directly apply user edits or engage in a conversation to clarify feedback and discuss revision options. Evaluations show that this approach of distributing conversational feedback improves user satisfaction with their goals while successfully preserving the narrative coherence, content quality, and visual presentation of the generated slides.