ECHO: Explainable Co-editing with Human-in-the-loop Operations for Presentation Slide Refinement

Abstract

Authoring and refining presentation slides is time-consuming in academic and professional settings. Although generative AI lowers the barrier to creating initial drafts, its black-box, one-way workflow often limits fine-grained control. A formative study with 10 frequent presentation authors identified trial-and-error anxiety, invisible intermediate decisions, cross-slide inconsistency, ambiguous spatial references, and fear of irreversible edits. We present ECHO, an interactive slide-refinement system that combines natural-language instructions with direct visual selection. ECHO translates multimodal intent into schema-constrained operation plans through a Plan-Confirm-Execute loop, maintains global style and user habit memory, routes spatially ambiguous requests to a vision-language model, and supports byte-exact rollback. We further introduce CoEdit-Eval, a multi-level framework for evaluating intent mapping, spatial grounding, execution safety, and rendered visual quality. Across four foundation models, ECHO raises Target Hit@1 from 0% for text-only baselines to 55–85%. A within-subjects study with 14 participants shows a 25.8% reduction in completion time and a 20.8% reduction in NASA-TLX workload. These results demonstrate how explicit execution boundaries can make AI-assisted document refinement more controllable, transparent, and reversible.