PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Mingxi Cheng, Bei Liu, Kai Qiu, Qi Dai, Dong Chen, Yue Dong, Xiaoqing Zheng, Ji Li, Chong Luo

arXiv:2609.00194v1cs.AI

TL;DR

Document-to-slide generation must jointly handle semantic fidelity and precise spatial layout, but existing refinement commonly delays holistic feedback until after large rewrites. ReDeck uses atomic edits with immediate renderer-derived observations, complemented by turn-level criticism and submission validation, and reports consistent improvements across three model backbones and multiple quality dimensions.

  • Problem

    Document-to-slide generation requires jointly preserving source semantics and arranging dense visual content within a bounded canvas, while common “one version, one feedback” refinement delays critique until after monolithic rewrites.

  • Method

    ReDeck decomposes slide refinement into atomic edits, returns render-grounded observations after each edit, combines them with a turn-level adaptive critic, and applies submission-level hard validation.

  • Results

    Across GPT-5.4, Claude-4.6, and Gemini-3.1, ReDeck consistently outperforms existing slide agents on content fidelity, spatial correctness, aesthetics, and information architecture.

  • Takeaways & Limitations

    DeckQuiz separates content fidelity, spatial correctness, and design quality, while ablations show that step-level render feedback and turn-level criticism provide complementary joint gains.

  • Takeaways & Limitations

    ReDeck requires editable geometry and repeatable rendering; the validation covers HTML/CSS and PPTX but not PDF/OCR-only editing, Keynote, animations, masking, or font fallback.

Abstract

from arXiv · show

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.

1 Introduction

Document-to-slide generation must jointly preserve source meaning and produce readable, valid layouts on a bounded canvas. ReDeck addresses the delayed, monolithic refinement pattern with atomic edits, immediate render observations, multi-granular feedback, and component-level evaluation.

  • Slides require joint optimization of content fidelity, layout validity, and visual readability because they are dense editable artifacts on a bounded canvas.
  • Existing agents commonly follow a “one version, one feedback” loop, rewriting a slide or deck before receiving holistic rendered critique.
  • Monolithic refinement makes it difficult to attribute local failures such as overflow, overlap, clipping, low contrast, and off-canvas placement to individual edits.
  • ReDeck decomposes refinement into atomic edits and returns structured renderer-derived observations after each action, enabling immediate repair of local spatial errors.
  • ReDeck assigns spatial facts to step-level render feedback, semantic and design guidance to a turn-level adaptive critic, and hard-violation validation to submission.
  • DeckQuiz separately measures content fidelity, spatial correctness, and design quality, while evaluations across GPT-5.4, Claude-4.6, and Gemini-3.1 report consistent gains over existing agents.

2 Related Work

Related work spans automated slide generation, LLM feedback and self-correction, and benchmarks for document design and layout. ReDeck targets a remaining gap: slide agents generally lack immediate rendered-state deltas that support repair after each edit.

  • Automated Slide Generation: Automated slide systems increasingly use specialized agents for slide creation, personalization, and narrative structure, but most still evaluate and refine at coarse turn-level granularity.
  • Automated Slide Generation: DeepPresenter inspects rendered output during refinement but does not validate layout integrity after each individual edit, allowing spatial errors to accumulate within a turn.
  • Feedback and Self-Correction in LLM Agents: LLM self-correction research includes verbal feedback, external evidence, execution signals, multi-agent procedures, and real-time observations in web, software, and GUI agents.
  • Feedback and Self-Correction in LLM Agents: Slide agents often lack a rendered-state delta that captures immediate layout regressions, despite related work grounding actions in compiler, test, or screenshot feedback.
  • Benchmarks for Document Design and Layout: Presentation benchmarks increasingly move beyond generic text similarity toward checks for source fidelity, canvas integrity, dimensions, whitespace, and visual balance.

3 Method

ReDeck models slide refinement as atomic edits in a partially observable rendering environment. An inner loop returns immediate technical observations, while an outer critic maintains deck-wide issues and a submission gate enforces the baseline violation constraint.

  • ReDeck reduces feedback delay from O(turn) to O(edit) by combining an inner render-feedback channel with an outer turn-level critic.
  • The environment state contains the editable deck representation, rendered spatial geometry, and compiler or export diagnostics, which are filtered into compact text observations.
  • A persistent issue list supplies static high-level design and content defects throughout a turn, while the agent combines it with each active observation to choose the next action.
  • The compositional action space restricts refinement to atomic structural, content, and visual modifications rather than whole-slide rewrites.
  • Inner Loop: Step-Level Render Feedback: After every atomic edit, ReDeck compiles the deck and returns a filtered observation containing violation changes, failing-element identifiers and bounding boxes, and layout anchors.
  • Inner Loop: Step-Level Render Feedback: Step-level feedback is a soft hint that permits temporary error states, whereas submission invokes a hard check requiring violations to be no greater than the session baseline.
  • Outer Loop: Turn-Level Adaptive Deck Critic: The adaptive critic evaluates source and rendered images, maintains categorized issues, and uses 237 single-purpose probes spanning geometric, source-retrieval, and rubric-based multimodal checks.
  • Outer Loop: Turn-Level Adaptive Deck Critic: The critic exclusively updates issue-list evidence and status, while the agent has read-only access and step-level feedback remains an instantaneous spatial report.

4 Experiments

Experiments evaluate ReDeck with DeckQuiz, ablations, convergence diagnostics, cost controls, and cross-domain tests. ReDeck leads across benchmark modules and domains, while selective frequent feedback improves refinement but incurs substantially higher inference cost.

  • Evaluation setup: DeckQuiz separately measures content fidelity, spatial correctness, aesthetics, and deck information architecture on 100 papers with three seeds.The study compares ReDeck with three published slide agents under matched per-task call caps across three model backbones.
  • Main results: ReDeck achieves the highest score on all four DeckQuiz modules for each agent backbone, with all 12 GPT-5.4 comparisons confirmed after Holm correction.The refinement advantage also persists when starting from either ReDeck or DeepPresenter initial decks.
  • Ablations: Combining action-level render feedback with a turn-level critic yields the best result on all four metrics, whereas each channel alone has narrower benefits.Render feedback substantially improves spatial correctness but reduces design relative to T0; the turn-level critic provides a small design improvement without improving spatial correctness.
  • Ablations: Performance decreases consistently as observation intervals increase from N=1 to N=4 and N=8, while evaluating all probes every turn performs substantially below adaptive scheduling.The exhaustive setting raises additional visual findings whose repairs introduce geometric regressions.
  • Cost and latency: ReDeck’s end-to-end wall-clock is 3586 s and its per-task cost is $1.74, corresponding to 8.5× DeepPresenter time and 14.5× cost.A ReDeck-Lite variant costs $0.28 per task and retains 89.8 SCR.
  • Cross-domain transfer: ReDeck is top-ranked in all five PresentBench domains, with win rates of 0.72 on academia and 0.66–0.68 on the most distant domains.The only rank-order change is on talk, where SlideTailor narrowly exceeds DeepPresenter; aggregate Kendall’s τ is 0.87.
  • External evaluation: Against DeepPresenter, ReDeck improves DECKBench LayoutQ, DeckFaith, and DeckFid by 0.032, 0.023, and 0.022, respectively.It also improves SlidesGen-Bench QuizBank accuracy by 2.6 percentage points and computational aesthetics by 3.2 points.

5 Conclusion

ReDeck refines slides through atomic edits with immediate render observations, complemented by turn-level global criticism and submission-level validation. Across three model backbones, this dual-scale feedback consistently outperforms existing agents, supporting feedback granularity as a broader principle for persistent rendered artifacts.

  • Conclusion: ReDeck decomposes revision into atomic edits and returns renderer-derived observations after each step, reducing the perception–action gap to a single edit.The turn-level critic supplies global design direction, while the submission gate prevents regressions from persisting across turns.
  • Conclusion: Across three model backbones, ReDeck consistently outperforms existing slide agents on content fidelity, spatial correctness, aesthetics, and information architecture.Ablations find that step-level render feedback and the turn-level critic are individually insufficient but jointly super-additive.
  • Broader implication: ReDeck suggests delivering feedback at the granularity where each problem can be most reliably observed and fixed in persistent, inspectable artifacts.The paper identifies web pages, documents, diagrams, and interactive interfaces as possible settings for this principle.

6 Limitations

ReDeck’s validation is bounded by editable-geometry and repeatable-rendering requirements, incomplete format coverage, and non-identical native generator comparisons. The full system also incurs substantial cost and speed overhead, while evaluator dependence and moderate human-study scale remain limitations.

  • ReDeck validates HTML/CSS and PPTX but not PDF/OCR-only editing, Keynote, animations, masking, or font-fallback behavior.
  • The full system costs 14.5× more and runs 8.5× slower than default DeepPresenter.Mean-cost-matched control and ReDeck-Lite clarify the quality–cost trade-off, but routing mitigates rather than removes it.
  • DeckQuiz and the adaptive critic share related quality constructs, so external checks reduce but do not eliminate evaluator dependence.The cited discussion also notes that subjective visual harmony remains difficult to automate.
  • The 45-task human study remains moderate in size, and subjective visual harmony is still difficult to automate.

B Step-Level Feedback Case Study

The case study contrasts refinement without step-level feedback, where edits cascade from three to five hard issues, with ReDeck, which detects regressions after individual edits and recovers to a clean state.

  • Figure 6: Figure 6 compares two trajectories from the same starting slide, with headers showing total hard issues and dashed boxes marking representative active issues.
  • Without step-level feedback: Without per-step rendering, three well-intentioned edits increase hard issues from 3 to 5 through cascading overlap and off-canvas regressions.In Case 02, one critique-only repair turn similarly raises the count from 24 to 31.
  • With step-level feedback: Step-level feedback accepts a bullet-column narrowing after detecting no new hard issue relative to the starting state.
  • With step-level feedback: When widening the chart places its right edge at x = 1340, ReDeck immediately reports the off-canvas regression and rolls back to the last clean state.
  • With step-level feedback: Rollback lets the agent try a different edit shape instead of retrying the same parameter, while the critic and render feedback observe distinct issue types.

C DeckQuiz Human Alignment Study

The DeckQuiz alignment study calibrates automated judges against human annotations across four modules, while related validation studies distinguish item-level calibration, deck-level preference, and SpatialCheck judge reliability.

  • Calibration design: DeckQuiz calibration samples 10 held-out items per module across ContentQuiz, SpatialCheck, Aesthetics, and DeckDesign, totaling 40 items.The modules cover content, hard spatial violations, design axes, and design principles.
  • Calibration design: Table 3 compares inter-rater α among annotators with judge–human α against the specified αjudge–human ≥0.5 demotion threshold.
  • Alignment results: All four modules clear the α = 0.50 threshold on the calibration subset, with Aesthetics closest at αjudge–human = 0.57.
  • External validation: The 45-task forced-choice study provides external deck-level validation against DeepPresenter using 270 judgments from three blinded annotators.
  • SpatialCheck validation: SpatialCheck calibration uses rendered PNGs for cross-format evaluation, while DOM-SCR provides a deterministic HTML consistency check.The VLM judge applies a fixed rubric to four hard-violation families.
  • SpatialCheck validation: DOM-SCR preserves the full VLM-SCR ordering with 92% slide-level agreement, κ = 0.83, MCC = 0.84, and MAE 1.4 points.

F PresentBench Domain-Wise Results

ReDeck ranks first across all five PresentBench domains, with strongest performance in academia and lower but above-chance win rates on more distant domains. The pooled baseline order is DeepPresenter, SlideTailor, then SlideGen, with one talk-domain swap.

  • 0.72 win rate makes ReDeck the top-ranked system in the scientific academia domain.
  • 0.66–0.68 win rates on talk and advertising remain above chance, indicating partial transfer to non-scientific content.
  • ReDeck is top-ranked in all five domains, and aggregate Kendall’s τ = 0.87 exceeds the specified τ ≥0.85 consistency threshold.
  • On talk, SlideTailor narrowly outscores DeepPresenter, 0.51 vs. 0.50, producing the only deviation from the pooled baseline order.

G External Evaluation

External evaluations test ReDeck across published protocols, PPTX rendering, and a pilot-audited probe taxonomy. The results support gains beyond DeckQuiz while defining the benchmark’s construction and scope.

  • External Evaluation: ReDeck gains 0.032 on LQ, 0.023 on DFaith, 0.022 on DFid, 2.6 percentage points on Quiz, and 3.2 points on Aes. relative to DeepPresenter.These external metrics triangulate rather than duplicate DeckQuiz’s four modules.
  • PPTX Transfer: PPTX refinement yields gains of +5.0 on Fid., +24.0 on SCR, +0.50 on Aes., and +6.0 on Des.The result supports transfer to a second editable geometry/rendering stack, but not to PDF/OCR-only editing, Keynote, animations, masking, or font-fallback behavior.
  • Probe Taxonomy: The six probe families derive from external failure dimensions and pass a pilot audit based on residual-other coverage and two-author Cohen’s κ.The audit labels 264 failures from 100 refinement turns, accepting the taxonomy when other ≤5% and κ ≥0.60.
  • DeckQuiz: DeckQuiz evaluates content fidelity, spatial correctness, and design quality separately, enabling attribution of improvements to corresponding refinement components.Its external construction uses answerable multiple-choice questions with same-discipline distractors and per-type accuracy.

N Initial-Deck Attribution and Task-Paired Inference

Task-paired analyses separate initial-deck quality from refinement-loop quality by applying both procedures to both initial-deck sources. ReDeck’s advantage persists across starting drafts and remains after cost scaling.

  • Initial-Deck Attribution: ReDeck improves SCR over the DeepPresenter loop by 19.5 points from ReDeck T0 and 20.2 points from DeepPresenter T0.The draft-by-loop interaction is not clear, with ΔΔ = −0.7 points.
  • Task-Paired Inference: All 12 paired contrasts remain significant after Holm correction, with the smallest contrast being +0.17 [0.02, 0.32] on Aes. versus DeepPresenter.Main-table seed-level standard errors range from 0.5 to 1.2 on 0–100 metrics and 0.03 to 0.07 on Aes.
  • Cost Control: At the 2175-call endpoint, ReDeck exceeds cost-scaled DeepPresenter by +11.6 on Fid., +13.5 on SCR, +0.44 on Aes., and +8.2 on Des.All four task-bootstrap confidence intervals exclude zero; scaling improves DeepPresenter SCR from 72.0 to 78.0 without closing the gap.
  • Regression Honesty: The analysis flags any negative per-layer change because an aggregate gain can conceal a regression in a layer or principle.ReDeck reports such regressions rather than assigning artificial baseline deltas to published systems.
  • Design Attribution: DeckDesign’s aggregate +8.4 gain is accompanied by positive deltas for all five principles, ranging from +5.8 to +11.2.Slide-level gains are +9.5 versus +6.7 for the deck-level layer.

Q Trajectory Event-Flow

ReDeck’s refinement trajectory is dominated by issue resolution, while adaptive scheduling limits persistence and rollback limits regressions. Probing every family each turn instead increases regression and issue drift.

  • Issue Lifecycle: Open → resolved accounts for 68% of issue-turn pairs, while open → persisted accounts for 22%.The persisted share indicates that stable issues can be deprioritized by the adaptive scheduler.
  • Issue Lifecycle: Resolved → regressed occurs in only 8.2% of cases, and per-step observation catches and rolls back 73% of potential regressions.The lifecycle also includes open → dropped when an issue is no longer detected without explicit addressing.
  • Scheduling Effects: The all-probes-every-turn variant raises resolved → regressed to 23.5% by introducing soft findings whose repair edits can break clean layout.The resulting style-to-layout propagation drives open-issue drift rather than failure of a single hard family.
  • Family Diagnostics: The factual family D has the highest reopen rate per issue under all-probes-every-turn despite remaining small in absolute count.Edits to neighboring text can break previously verified claims, making step-level render feedback relevant beyond layout.

R Diagnostic Mini-Ablations

Mini-ablations identify the turn-level critic’s input modality as a key constraint: rendered images preserve spatial and design performance, whereas source-code-only critique does not.

  • Critic Source: Self-Refine and Reflexion differ by less than 1 point on every DeckQuiz module, within observed seed-level variability.This indicates that the NL-critique modality, rather than its source, is the binding constraint.
  • Critic Modality: Removing rendered-image input from the turn-level critic drops SCR by 17.3 and Des. by 6.1 while leaving Fid. approximately flat.The comparison uses the default ReDeck configuration at 88.6/91.5/3.64/71.2.
Loading 2609.00194v1…