- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes REDECK, a step-level render-grounded refinement framework for document-to-slide generation. The method introduces atomic editing actions with immediate renderer-derived feedback, combined with a turn-level adaptive critic and a submission-level validation gate. The authors also introduce DECKQUIZ to evaluate content fidelity, spatial correctness, aesthetics, and design quality. Experiments across multiple LLM backbones demonstrate improvements over existing slide-generation agents.
Critical strengths
-
The proposed step-level render-grounded feedback addresses an important limitation of conventional turn-level refinement by detecting layout errors immediately after individual edits.
-
The ablation studies in Table 3 demonstrate the effectiveness of feedback granularity and the complementary roles of step-level observations and turn-level criticism.
-
REDECK consistently outperforms the evaluated baselines across multiple LLM backbones and evaluation dimensions.
-
The shared-start comparisons, external benchmark evaluations, and blinded human study further strengthen the empirical evidence and support the effectiveness of the proposed refinement strategy.
Critical weaknesses
-
The full REDECK framework incurs significantly higher inference costs and latency than existing approaches. Section 4.4 reports approximately 14.5× higher cost and 8.5× longer runtime compared with DeepPresenter.
-
Although REDECK-Lite reduces the overhead, it remains considerably more expensive than the baseline. The lack of fully cost- and latency-matched comparisons makes it difficult to assess the method's practical efficiency relative to its performance gains.
-
DECKQUIZ and the adaptive critic share related evaluation criteria, which may introduce evaluation alignment effects. Although external benchmarks and human evaluations partially address this concern, the possibility of evaluation dependence is not completely eliminated.
-
The primary experiments focus on scientific presentation generation. While the PPTX transfer results are encouraging, broader validation across diverse presentation types and editing environments remains limited.
Questions for authors
-
Can the authors provide a more direct comparison of performance under matched computational costs or latency budgets? This would help clarify whether the substantial additional inference overhead is justified by the observed improvements.
-
How well does REDECK generalize beyond scientific presentation generation? In particular, could the authors clarify its performance relative to existing methods in native PPTX workflows and other presentation scenarios?
2 comments
Native PPTX and broader presentation scenarios. The submitted paper already reports transfer of the unchanged refinement loop to native editable PPTX: replacing DOM geometry with python-pptx geometry and browser rendering with LibreOffice improves T0 from 78.0/58.0/2.80/60.0 to 83.0/82.0/3.30/66.0 on Fid/SCR/Aes/Des, including +24.0 [19.1, 28.9] SCR. This is a refinement-transfer result, not a native end-to-end generation comparison. We add shared-start PPTX controls in which DeepPresenter, screenshot critique, and ReDeck receive the same PPTX T0 and use the same tools and renderer. ReDeck exceeds screenshot critique by +13.5 [9.9, 17.1] SCR and +4.0 [1.6, 6.4] Fid while retaining native editable text frames. PPTAgent is reported separately as an end-to-end reference because it generates its own initial deck and therefore is not part of the shared-start causal comparison. Re-rendering in PowerPoint changes the violation status of 3.5% of slides, mainly because of font metrics, which we retain as a limitation.
We further test three non-scientific scenarios—investor briefings from business reports, lecture slides from textbook chapters, and policy briefings from public reports—with sources clustered as the unit of inference. ReDeck shows SCR gains of +21.4 to +23.3 and Fid gains of +6.9 to +7.4 across all three settings. Aesthetics are unresolved in two scenarios, and DeckDesign is neutral for business reports (+1.8 [−0.9, 4.5]). These experiments support transfer beyond scientific papers in the tested scenarios, but not universal generalization. Keynote, animation, masking, PDF/OCR-only editing, and font-fallback behavior remain untested and will be listed explicitly.
Across all three responses, the revision will foreground evidence already present in the submitted paper, add only verified new results, recommend Lite rather than Full as the practical default, and narrow claims where the evidence remains conditional.
We thank the reviewer for recognizing the value of step-level render-grounded feedback, the Table 3 granularity ablation, and the shared-start, external-benchmark, and blinded-human evidence already included in the paper. We agree that the remaining questions are practical efficiency under matched resources, residual dependence between DeckQuiz and the adaptive critic, and the empirical scope beyond scientific HTML presentations.
Matched cost and latency. Full is substantially more expensive than the baselines: \1.74and3,586sperdeckversus0.12 and 420 s for DeepPresenter. The paper's Lite routing reduces the totals to \0.28and1,450sbutdoesnot,byitself,establishefficiencyatequalresources.Wethereforeaddcomparisonsundermaximumrefinement−budgetcapsof0.10, \0.25,0.60, and \1.65afterasharedT0,aswellashardlatencycutoffsof420sand900s.Baselinesmayscalewithineachbudgetusingadditionalturnsorbest−of−kselection;theirpolicyisselectedonadisjointdevelopmentset,thesamemethod−independentvaliditygateischargedtoeveryarm,andnofailedrunisdropped.At0.10, Lite improves SCR over DeepPresenter by +14.1 [11.2, 17.0] and is preferred by blinded raters on 30/45 tasks (0.67 [0.52, 0.79]). At 420 s, Lite-P improves SCR by +14.9 [11.6, 18.2]. At the largest tested budget, Full has the highest point estimates, but its +1.0 [−0.9, 2.9] SCR increment over Lite is unresolved. At the lowest end-to-end budget of \$0.12, DeepPresenter has higher aesthetics by 0.32 [0.17, 0.47]. These results support Lite as the strongest tested quality–cost operating point, not the stronger claim that ReDeck dominates every metric or every budget.
DeckQuiz and adaptive-critic dependence. We agree that sharing quality criteria can introduce alignment effects even without test-item leakage. We distinguish four layers. First, the data-flow audit finds no route by which ReDeck can access DeckQuiz items, answers, labels, thresholds, or judge outputs, and the probe library was frozen before DeckQuiz was built. Second, the deterministic DOM-based SCR implementation preserves the system ordering and agrees with the VLM evaluator at κ = 0.83. Third, Claude-4.6 and Gemini-3.1 rank ReDeck first across the eight reported module comparisons, with seven intervals excluding zero; Claude Aesthetics is the exception. Fourth, evidence outside the probe taxonomy—including DeckBench, SlidesGen-Bench, AeSlides, blinded human preference, and the proposed human hard-violation audit—preserves the main ordering. These analyses address test-instance reuse, evaluator implementation, and judge choice, but shared constructs remain. We will therefore say that evaluator dependence is reduced rather than eliminated, with DeckDesign identified as the principal remaining limitation.