ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation
ReDeck enhances slide generation with step-level render feedback, improving spatial accuracy and design quality.
Key Findings
Methodology
ReDeck employs a step-level render-grounded refinement framework, decomposing slide revision into atomic edit actions with renderer-derived observations after each step. It uses multi-granular feedback to balance local repair and global quality, including step-level render feedback, turn-level adaptive critique, and submission-level validation.
Key Results
- ReDeck excels in DeckQuiz benchmark, with content fidelity reaching 88.6%, spatial correctness at 91.5%, and design quality scoring 3.64.
- Compared to SlideGen, SlideTailor, and DeepPresenter, ReDeck shows superior performance across GPT-5.4, Claude-4.6, and Gemini-3.1, especially in spatial correctness and design quality.
- Ablation studies confirm that feedback timing and granularity are critical for reliable slide refinement.
Significance
ReDeck significantly enhances spatial accuracy and design quality in slide generation through its step-level render feedback mechanism. This approach addresses the issue of delayed feedback in traditional methods, offering new insights and tools for the slide generation field with substantial academic and industrial applications.
Technical Contribution
ReDeck introduces a step-level render feedback mechanism, decomposing slide revision into atomic edit actions, providing fine-grained spatial repair and global semantic design guidance. It innovates in feedback granularity and timing, significantly improving slide generation reliability.
Novelty
ReDeck is the first to apply step-level render feedback in slide generation, solving delayed feedback issues in traditional methods and significantly enhancing spatial accuracy and design quality.
Limitations
- ReDeck may face challenges with complex layouts, especially when multiple elements require simultaneous adjustment.
- For certain specific design styles, ReDeck might not provide optimal visual effects.
Future Work
Future work could explore ReDeck's application to other document types, such as web design or report generation. Additionally, further optimization of the feedback mechanism could cater to more complex design needs.
AI Executive Summary
Slide generation is a complex task requiring accurate content selection and precise spatial layout. Traditional methods often use a monolithic feedback mechanism, making local errors hard to fix. ReDeck introduces a novel step-level render feedback framework, decomposing slide revision into atomic edit actions with renderer-derived observations after each step, achieving a 'one edit, one observation' optimization process.
ReDeck balances local repair and global quality through multi-granular feedback. Step-level render feedback addresses spatial errors, turn-level adaptive critique provides semantic and design guidance, and submission-level validation ensures strict layout integrity. Experimental results show ReDeck's superior performance in the DeckQuiz benchmark, significantly outperforming existing slide generation agents.
ReDeck's innovation lies in its fine-grained feedback mechanism and immediate render observations, enhancing spatial accuracy and design quality in slide generation. While there is room for improvement in handling complex layouts, ReDeck's contributions to the slide generation field are undoubtedly significant.
Deep Analysis
Background
Slide generation is a crucial area in document processing, involving complex content selection and spatial layout optimization. Traditional methods often follow a generate-then-render paradigm, leading to delayed feedback issues. Recent advances in large model self-correction have introduced reflection mechanisms, but feedback delay remains a problem.
Core Problem
The core problem in slide generation is optimizing spatial layout and visual readability while maintaining content fidelity. Feedback delay in traditional methods makes local errors difficult to fix, affecting overall quality.
Innovation
ReDeck's core innovation is its step-level render feedback mechanism, decomposing slide revision into atomic edit actions for immediate feedback and correction. This mechanism not only improves spatial accuracy but also provides global semantic and design guidance.
Methodology
- �� Decompose slide revision into atomic edit actions, such as adding images or inserting bullets.
- �� After each action, the rendering environment returns structured observations of the current visual state.
- �� Employ multi-granular feedback: step-level render feedback, turn-level adaptive critique, submission-level validation.
Experiments
The experimental design includes using the DeckQuiz benchmark to evaluate ReDeck's performance in content fidelity, spatial correctness, and design quality. Baselines include SlideGen, SlideTailor, and DeepPresenter, evaluated on the same tasks and models.
Results
ReDeck excels in the DeckQuiz benchmark, with content fidelity reaching 88.6%, spatial correctness at 91.5%, and design quality scoring 3.64. Compared to existing methods, ReDeck shows significant improvements across all metrics.
Applications
ReDeck can be used in academic presentations, business pitches, and other scenarios, helping users quickly generate high-quality slides. Its step-level feedback mechanism is suitable for document generation requiring precise layout and design.
Limitations & Outlook
ReDeck may face challenges with complex layouts, especially when multiple elements require simultaneous adjustment. Additionally, for certain specific design styles, ReDeck might not provide optimal visual effects.
Plain Language Accessible to non-experts
Imagine you're in a kitchen preparing a large meal. Every time you add an ingredient, you immediately taste it to see how it affects the flavor. ReDeck acts like this chef, making small adjustments to a slide and immediately checking the result. This approach helps ReDeck quickly identify and fix issues, rather than waiting until the entire dish is finished to find out something's wrong.
ELI14 Explained like you're 14
Imagine you're playing a building game where every time you place a block, the game tells you if it's in the right spot. ReDeck works like that; when generating slides, it makes small changes and immediately checks the effect. This method lets ReDeck fix problems quickly, just like you can adjust your building in the game.
Glossary
ReDeck (Step-Level Render Feedback)
A framework that decomposes slide revision into atomic edit actions with renderer-derived observations after each step.
Used to optimize spatial accuracy and design quality in slide generation.
DeckQuiz (Benchmark)
A benchmark for evaluating slide generation systems, including content fidelity, spatial correctness, and design quality.
Used to validate ReDeck's performance improvements.
Atomic Edit Actions
Minimal modifications to slides, such as adding images or adjusting colors.
ReDeck uses these actions for step-level optimization.
Multi-Granular Feedback Mechanism
Includes step-level render feedback, turn-level adaptive critique, and submission-level validation.
Balances local repair and global quality.
Renderer-Derived Observations
Structured feedback on the current visual state after each atomic edit action.
Helps ReDeck quickly identify and fix issues.
Open Questions Unanswered questions from this research
- 1 How to further optimize ReDeck's performance on complex layouts, especially with multiple element adjustments.
- 2 How to extend ReDeck's application to more document types for broader use.
Applications
Immediate Applications
Academic Presentation Generation
ReDeck can help researchers quickly generate high-quality academic presentation slides, ensuring accurate content and reasonable layout.
Business Presentation Design
Companies can use ReDeck to quickly design business presentations, improving efficiency and presentation effectiveness.
Long-term Vision
Cross-Domain Document Generation
ReDeck's step-level feedback mechanism can be applied to other document types, such as web design, enhancing overall design quality.
Abstract
Document-to-slide generation is challenging because slides are dense editable artifacts that require both faithful content selection and precise spatial layout. Recent slide agents adopt iterative reflection, but typically follow a monolithic "one version, one feedback" loop: a slide or deck is rewritten, rendered afterward, and critiqued only at the turn boundary. This delayed feedback makes local failures such as overflow, overlap, clipping, and off-canvas placement difficult to attribute and repair. We propose ReDeck, a step-level render-grounded refinement framework that 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 uses 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 hard layout validation. We further introduce DeckQuiz, a benchmark that decouples 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 ablations confirm that feedback timing and granularity are critical for reliable slide refinement.