Paper #44: SlideMatcher: Agentic Slide Editing under Item Mismatch

Slide presentations are a central medium for communication in business, education, and research, yet creating professional designs remains time-consuming. Recent advances in AI have enabled automatic slide generation, but most methods rely solely on user instructions and fail to account for reference-based adaptation, where users wish to preserve the visual design of existing slides. In this paper, we study the challenge of Item Mismatch, a common scenario where the number of items in user content differs from that in the reference slide, requiring both content updates and layout adjustments. We introduce ItemChallenge, a benchmark dataset for evaluating slide editing under Item Mismatch, and propose SlideMatcher, an agentic framework that leverages Semantic Groups to analyze and relocate elements effectively. Experiments show that SlideMatcher outperforms existing methods in preserving design fidelity and adapting to new content, enabling users to focus on their content while AI handles the complexity of layout adaptation.