ReDeck Improves AI Document-to-Slide Generation with Step-Level Feedback

Generating presentation slides from documents using AI has been challenging due to the need for both accurate content selection and precise visual layout. Existing AI slide agents often use an iterative "one version, one feedback" loop, where an entire slide or deck is rewritten, rendered, and then critiqued. This delayed feedback makes it difficult to pinpoint and fix specific layout issues like text overflow, overlapping elements, or off-canvas placement. Researchers have introduced ReDeck, a novel framework designed to address these limitations. ReDeck breaks down the slide revision process into atomic edit actions and provides immediate, render-derived observations after each step. This transforms refinement into a "one edit, one observation" cycle, allowing for much more granular and timely error correction. To ensure both local accuracy and overall quality, ReDeck employs multi-granular feedback. It uses step-level render feedback for spatial errors, a turn-level adaptive critic for semantic and design guidance, and a submission-level gate for final layout validation. Evaluated against leading models like GPT-5.4, Claude-4.6, and Gemini-3.1, ReDeck consistently outperforms other slide-generation agents, demonstrating the critical importance of feedback timing and granularity for reliable slide refinement.