Generating slides from documents is difficult because slides must preserve content fidelity while arranging elements precisely. Existing slide agents follow a monolithic “one version, one feedback” loop: a slide or deck is rewritten, rendered afterwards, and critiqued only at the end of a turn. This delayed feedback makes it hard to locate and fix local issues such as overflow, overlap, clipping, or off‑canvas placement.
ReDeck decomposes slide revision into atomic edit actions and returns renderer‑derived observations after each step, turning refinement into “one edit, one observation.” To balance local repair with global quality, ReDeck employs multi‑granular feedback: step‑level render feedback for spatial errors, a turn‑level adaptive critic for semantic and design guidance, and a submission‑level gate for strict layout validation.
We also introduce DeckQuiz, a benchmark that separates content fidelity, spatial correctness, and design quality. Across GPT‑5.4, Claude‑4.6, and Gemini‑3.1, ReDeck consistently outperforms existing slide‑generation agents, and ablation studies confirm that the timing and granularity of feedback are crucial for reliable slide refinement.
Review