ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation
Muzhao Tian, Zezi Zeng, Yifan Yang, Xin Gao, Yan Li, Zisu Huang, Xiaohua Wang, Changze Lv, et al.
cs.AI
Quick verdict
ReDeck's per-edit render feedback beats turn-level slide refinement on DeckQuiz; rigorous ablations, but 14.5x cost and self-built metrics temper impact.
Summary
ReDeck reframes document-to-slide refinement from monolithic 'one version, one feedback' loops to 'one edit, one observation': an LLM agent makes atomic edits (structural, content, or attribute changes, plus rollback and submit) and after each edit receives a renderer-derived observation — a violation delta against the session-start baseline plus a geometry anchor of element bounding boxes. A turn-level adaptive critic samples 3–5 probes from a 237-check library to maintain a persistent issue list, and a submission gate rejects any turn whose hard-violation count exceeds the baseline. The authors also introduce DeckQuiz, a four-module benchmark (ContentQuiz fidelity, VLM-judged SpatialCheck, Aesthetics, DeckDesign) over 100 arXiv papers. With GPT-5.4, ReDeck reaches 88.6 fidelity / 91.5 spatial clean-rate / 3.64 aesthetics / 71.2 design, beating SlideGen, SlideTailor, and DeepPresenter across the GPT-5.4, Gemini-3.1, and Claude-4.6 backbones under a 150-call cap. Gains persist in shared-start controls, PresentBench cross-domain transfer, external benchmarks, PPTX transfer, and a 45-task blinded human study (73% preference). Ablations show both feedback channels are needed. Main limitations: 8.5x latency and 14.5x cost versus DeepPresenter, reliance on author-designed evaluators judged by GPT-5.4, formal significance testing only for the GPT-5.4 backbone, and a moderate-size human study.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Delivering renderer-derived observations after each atomic edit ('one edit, one observation') repairs local layout failures at the moment they occur, substantially improving spatial correctness over turn-boundary feedback.
- Step-level render feedback and turn-level adaptive critique are individually insufficient but jointly necessary; quality degrades monotonically as the observation interval grows (N=1 > 4 > 8) and exhaustive probe scheduling destabilizes refinement.
- ReDeck outperforms SlideGen, SlideTailor, and DeepPresenter on DeckQuiz across GPT-5.4, Gemini-3.1, and Claude-4.6 under a matched 150-call cap, with the loop advantage persisting under shared-start, cost-matched, external-benchmark, and blinded human evaluation.
- The refinement principle transfers across rendering stacks (HTML/CSS to PPTX via python-pptx + LibreOffice) and across the five PresentBench domains.
- DeckQuiz decouples content fidelity, spatial correctness, and design quality, enabling component-level attribution that aggregated final-deck scores cannot provide.
Strengths
Ablations isolate feedback timing and granularity cleanly
Tab. 2 separates paradigm (NL critique 68.2 SCR / screenshot 78.4 / ReDeck 91.5), components (render-only 80.2 SCR but Des. 59.4 below T0; critic-only 60.8 SCR), and frequency (SCR 91.5 → 87.4 → 82.1 for N=1/4/8). The scheduling ablation is particularly informative: all-probes-every-turn collapses to ~T0 (62.8 SCR) with App. Q showing resolved→regressed rising from 8.2% to 23.5%.
Attribution controls beyond the headline table
Shared-start 2×2 (App. N, Tab. 11a) shows the loop adds +19.5/+20.2 SCR from either T0 with negligible draft-by-loop interaction (ΔΔ=−0.7); the cost-matched control (Tab. 13) scales DeepPresenter to 2175 calls/$1.74 and it still trails by −11.6 Fid./−13.5 SCR; Fig. 5 adds cross-domain PresentBench transfer with Kendall's τ=0.87.
Evaluator validation is unusually thorough
Honest cost and regression reporting
Weaknesses
Circularity between method and evaluation
Formal statistics cover only one of three backbones
Severe cost/latency overhead limits practical adoption
Main-table protocol matches calls, not tokens or cost
Submission gate is baseline-relative, so dirty starting decks can commit with violations
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ReDeck (Microsoft + SJTU/XJTU/Fudan; arXiv 2609.00194v2) targets document-to-slide generation. The central diagnosis (§1, Fig. 1) is that existing slide agents refine monolithically — 'one version, one feedback' — so local spatial failures (overflow, overlap, clipping, off-canvas) cannot be attributed to individual edits. The claims: (1) step-level render-grounded refinement (atomic edit → renderer observation) repairs local failures as they occur; (2) multi-granular feedback (step-level render facts, turn-level adaptive critic, submission-level hard gate) is jointly necessary; (3) DeckQuiz decouples fidelity, spatial correctness, and design for component-level attribution; (4) ReDeck consistently beats published slide agents across three backbones. This is a method paper with a bundled benchmark and a general 'feedback at the right granularity' principle (§5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ReDeck?
ReDeck is a framework (from Microsoft and university collaborators) for turning documents into presentation slides. Instead of rewriting whole slides and getting one critique per round, an LLM agent makes small atomic edits and receives a renderer-generated report after every single edit — including whether the edit introduced overflow, overlap, clipping, or off-canvas problems — plus a per-round design critic and a final layout gate that blocks submission if hard violations exceed the starting baseline.
How much better is ReDeck than previous slide-generation agents?
On the paper's new DeckQuiz benchmark with GPT-5.4, ReDeck scores 88.6/100 content fidelity and 91.5% spatial clean-rate, versus DeepPresenter's 76.8 and 66.8 respectively; it also leads on aesthetics (3.64 vs 3.47) and deck design (71.2 vs 55.1). Wins hold across three LLM backbones, on external benchmarks, in cross-domain transfer, and in a blinded human study where annotators preferred ReDeck in 73% of 45 paired comparisons.
What does ReDeck cost to run?
On GPT-5.4 it costs about $1.74 and takes roughly 60 minutes (3,586 seconds) per deck — 14.5× the cost and 8.5× the latency of the DeepPresenter baseline. A cheaper 'ReDeck-Lite' variant that routes routine edits and checks to a smaller model costs $0.28 per deck with only about a 1–2 point quality drop on most metrics.
Is the code or benchmark available?
The paper points to github.com/microsoft/ReDeck and states that prompt templates, the 237-probe definitions with thresholds, and the 2,000-question DeckQuiz bank are released with the codebase; evaluation code is MIT-licensed. Reproducing the full result tables requires paid access to frontier models (GPT-5.4, Gemini-3.1, Claude-4.6).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text inside figures (Figs. 1–8) is interleaved and garbled by PDF extraction; figure content is judged from captions, tables, and surrounding prose rather than the figures themselves.
- The provided metadata lists 8 authors, but the full paper byline lists 17 (Microsoft, SJTU, XJTU, Fudan); this review follows the full byline. The provided text otherwise runs complete through Appendix R with no evidence of end truncation.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.