ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation

• Existing slide agents commonly use a monolithic “one version, one feedback” loop, making it difficult to attribute spatial failures to individual edits.The paper situates this approach within generate-then-render systems [1] [2] [3] [4] [5] and LLM self-correction [6] [7] [8] [9] [10].Source: §1 · pp. 1–2
• ReDeck decomposes revision into atomic edits and returns renderer-derived observations after each edit, enabling repair while local layout failures remain attributable.Observations include overflow, overlap, clipping, off-canvas placement, contrast issues, and key element geometry.Source: §1 · pp. 2–3, Figure 1 · p. 2
• ReDeck assigns local spatial repair to step-level render feedback, global semantic and design guidance to a turn-level adaptive critic, and hard-violation checks to a submission gate.The gate requires that the total hard-violation count not increase relative to the session baseline before a turn is committed.Source: §1 · p. 3
• Across GPT-5.4, Claude-4.6, and Gemini-3.1, ReDeck consistently outperforms existing slide-generation agents under a matched per-task LLM-call-count cap.Ablations find that step-level render grounding improves spatial correctness and stabilizes refinement, while the turn-level adaptive critic provides complementary gains in semantic and design quality.Source: §1 · p. 3
• DeckQuiz separately measures content fidelity, spatial correctness, and design quality to attribute improvements to specific refinement components.This avoids relying only on aggregated final-deck scores.Source: §1 · p. 3
• Shared-start controls, external-benchmark evaluation, a 45-task blinded human study, and PPTX transfer test attribution, evaluator dependence, and format scope.Source: §1 · p. 3

Automated Slide Generation

• Most slide-generation systems refine at turn level; DeepPresenter [5] inspects renders but lacks per-edit layout validation, unlike ReDeck’s step-level repair contract.Other systems address personalization and narrative structure, while screenshot-to-web and one-shot layout-generation work do not provide a multi-turn render-grounded repair loop.Source: §2 · p. 3

Feedback and Self-Correction in LLM Agents

• Where agent environments and GUI/vision agents use real-time observations [10] [30] [31] [32] [33], slide agents largely retain holistic or scalar reflection [3] [5] [15]; ReDeck separates step-level facts from turn-level guidance.This decouples objective spatial observations from semantic guidance during long-horizon refinement.Source: §2 · p. 4

Benchmarks for Document Design and Layout

• Presentation benchmarks add verifiable checks for layout and source-aligned content, while DeckQuiz separates hard spatial-violation criteria from subjective design principles to address LLM-judge variance and bias.The cited benchmark landscape also includes cross-format evaluation and instance-specific rubrics.Source: §2 · p. 4

3.1 Problem Modeling and Compositional Action Design

• The refinement state combines editable deck source, rendered geometry, and compiler diagnostics, while atomic structural, content, and attribute edits make layout consequences traceable.A persistent issue list guides actions alongside each step’s compact observation; rollback supports local retry after unsuccessful edits.Source: §3.1 · pp. 4–5

3.2 Inner Loop: Step-Level Render Feedback

• ReDeck compiles after every atomic edit, reporting baseline-relative compiler and DOM violation deltas with element anchors as soft feedback, then rejects submissions unless violations return to baseline.The observations omit screenshots and subjective design scores, and temporary errors during repair are allowed; the submission check uses the same baseline as the step-level reports.Source: §3.2 · pp. 5–6

3.3 Outer Loop: Turn-Level Adaptive Deck Critic

• The adaptive critic maintains structured issues across five categories and schedules three to five probes per turn from a library of 237 checks.Each issue records its category, affected slide, evidence pointer, and status; probes use geometric checks, source retrieval, or multimodal rubrics, with scheduling informed by recent edits, regressions, and resolved issues.Source: §3.3 · p. 6

3.4 The Full ReDeck Framework

• ReDeck’s full loop updates a persistent issue list at each turn, runs atomic edits against that list and step-level observations, and submits through a validation gate.The critic alone can add or localize issues and verify their resolution; the agent can read but not modify issue fields, preventing self-reported success and limiting local optimization from undermining deck-wide coherence.Source: §3.4 · p. 7

4.1 Setup

• DeckQuiz separates content fidelity, spatial correctness, aesthetics, and deck design, while the in-domain split covers 100 stratified papers and comparisons use a matched 150-call cap.Its modules assess facts conveyed, clean rendering, slide appearance, and information organization; a fixed GPT-5.4 backbone judges all systems, with three seeds per model-system-task cell. PresentBench [36] is reserved for cross-domain testing because of known judge cross-model drift.Source: §4.1 · p. 7

4.2 Main results

• ReDeck scores highest on all four DeckQuiz modules for each of three agent backbones, with all 12 GPT-5.4 paired comparisons significant after Holm correction.The comparison uses 100 in-domain tasks with three seeds under a matched per-task LLM-call-count cap; other backbones are descriptive replications.Source: §4.2 · p. 8, Table 1 · p. 8

4.3 Ablations

• Combining action-level render feedback with a turn-level critic yields the best result on all four metrics, while the channels separately favor spatial correction and design quality.Performance falls as the observation interval increases from N=1 to N=4 or N=8; evaluating all probes each turn performs close to T0 and below adaptive scheduling, with additional visual findings linked to geometric regressions. The convergence figures show adaptive scheduling dropping open issues sharply and reaching an approximately 77% per-turn resolution rate, while all-probes reaches about 20%.Source: Figure 3 · p. 8, §4.3 · p. 9, Figure 4 · p. 10

4.4 Cost and latency

• At matched post-T0 inference cost, DeepPresenter scores 77.0/78.0/3.20/63.0, below ReDeck’s 88.6/91.5/3.64/71.2 on Fid./SCR/Aes./Des.ReDeck costs $1.74 per task and takes 3586 seconds end-to-end, respectively 14.5× the cost and 8.5× the wall-clock of DeepPresenter; ReDeck-Lite costs $0.28 per task and retains 89.8 SCR.Source: §4.4 · pp. 9–10

4.5 Robustness and validity

• On PresentBench, ReDeck ranks first in all five domains, and cross-domain ranking consistency is τ = 0.87, above the 0.85 threshold.Against DeepPresenter, ReDeck improves DECKBench LayoutQ, DeckFaith, and DeckFid by 0.032, 0.023, and 0.022; on SlidesGen-Bench it improves QuizBank accuracy by 2.6 percentage points and computational aesthetics by 3.2 points. ReDeck wins human pairwise judgments over DeepPresenter on overall quality in 33/45 cases and source faithfulness in 31/45; PPTX transfer improves the four DeckQuiz modules from 78.0/58.0/2.80/60.0 to 83.0/82.0/3.30/66.0, and internal-validity audits do not reverse the headline ranking.Source: §4.5 · pp. 10–11, Figure 5 · p. 10

5 Conclusion

• ReDeck decomposes slide revision into atomic edits and returns renderer-derived observations after each step, enabling local layout failures to be detected and repaired as they arise.A turn-level adaptive deck critic provides global design direction, while a submission-level validation gate prevents regressions from persisting across turns.Source: §5 · p. 11
• Across three model backbones, ReDeck consistently outperforms existing slide agents on content fidelity, spatial correctness, aesthetics, and information architecture.Shared-start controls isolate the refinement loop’s effect from initial-deck quality.Source: §5 · p. 11
• Ablations show that step-level render feedback and the turn-level critic are individually insufficient but jointly super-additive.Shared-start controls isolate the refinement loop from initial-deck quality.Source: §5 · p. 11
• The conclusion proposes matching feedback granularity to where a problem can be reliably observed and fixed, and suggests this principle may extend beyond slides to web pages, documents, diagrams, and interactive interfaces.The proposed scope is iterative generation of persistent, inspectable artifacts where symbolic edits have nonlinear rendered consequences.Source: §5 · p. 11

6 Limitations

• ReDeck is validated for editable HTML/CSS and PPTX, not PDF/OCR-only editing, Keynote, animations, masking, or font-fallback behavior.Native end-to-end comparisons also combine each system’s initial generator and refinement loop; a shared-start experiment isolates loop quality across two initial-deck sources without making the native generators identical.Source: §6 · p. 11
• The full system costs 14.5× more and runs 8.5× slower than default DeepPresenter; routing mitigates, but does not remove, the quality–cost trade-off.A mean-cost-matched control and ReDeck-Lite clarify that trade-off.Source: §6 · p. 11
• External benchmarks, non-author SCR labels, a deterministic DOM cross-check, and blinded human preference reduce evaluator coupling, but do not eliminate evaluator dependence.The 45-task human study remains moderate in size, and subjective visual harmony is difficult to automate.Source: §6 · pp. 11–12

A Initial-Draft Pipeline and Rendering Environment

• The appendix documents the bootstrap that creates session-start deck d0 and the rendering environment that produces rendered geometry rt and diagnostics Ct, outside the refinement loop.ReDeck ablations share d0, renderer, and extractor; the loop’s action space, feedback, critic, and submission gate are specified separately.Source: A Initial-Draft Pipeline and Rendering… · p. 16
• Bootstrap parsing converts PDF or LaTeX papers into structured sections and separately extracts tables, figures, and equations as referenceable entities with captions and source locations.A single LLM call then assigns slide roles and selects source spans in a JSON deck blueprint.Source: A Initial-Draft Pipeline and Rendering… · p. 17
• Bootstrap selects from 15 base layouts by slide role, content density, and figure count, then generates each slide’s code independently from its blueprint entry, template, and source spans.The resulting HTML/CSS deck d0 is shared by ReDeck variants; “no refinement” or “T0” means d0 rendered without the loop, while published baselines use native pipelines under a matched LLM-call cap.Source: A Initial-Draft Pipeline and Rendering… · p. 17
• HTML slides render to PNG in headless Playwright/Chromium at 1280 × 720; a DOM extractor computes bounding boxes, overlap, overflow, and off-canvas indicators alongside compiler diagnostics Ct.Step-Level Render Feedback uses the extractor after atomic actions, while the Adaptive Deck Critic uses it with rendered PNGs and deterministic-geometry probes.Source: A Initial-Draft Pipeline and Rendering… · p. 17
• For final SCR reporting, all systems are rendered to PNG and judged by SpatialCheck VLM; the DOM extractor provides ReDeck’s internal feedback, not the cross-system SCR score.Source: A Initial-Draft Pipeline and Rendering… · p. 17

B Step-Level Feedback Case Study

• Without per-step observations, the baseline batches edits, then introduces a chart-callout overlap, an off-canvas chart, and a metric strip that collides with bullet text.The critic reports issues only at turn boundaries, allowing the regression chain to accumulate before the next verdict.Source: B Step-Level Feedback Case Study · pp. 17–18
• Per-edit rendering catches ReDeck’s off-canvas chart regression immediately, enabling rollback to the latest clean state and a different edit shape that resolves the original spatial issues.After narrowing the bullet column, a chart-widening attempt extends to x = 1340; ReDeck rolls it back, then moves the takeaway into a full-width bottom band and darkens the title.Source: B Step-Level Feedback Case Study · p. 18
• The turn-level deck critic and step-level renderer detect different signals: the critic flags a narrative problem, and the next render check confirms its repair did not regress geometry.The critic finds that the chart alone is hard to parse in three seconds; ReDeck adds a three-card metric strip below it.Source: B Step-Level Feedback Case Study · p. 18
• Without step-level feedback, hard issues rise from 3 to 5 (and Case 02 issues rise from 24 to 31), while ReDeck’s final state retains zero hard geometric issues.Figure 6 summarizes the contrasting trajectories; the baseline’s new edits trigger cascading overflow and density regressions before turn-boundary feedback.Source: B Step-Level Feedback Case Study · p. 18, Figure 6 · p. 19

C DeckQuiz Human Alignment Study

• Three experienced annotators independently labeled 40 held-out items (10 per module), with judge–human Krippendorff’s α ≥0.5 set as the reliability threshold.The 120 judgments covered ContentQuiz correctness, SpatialCheck violation counts, Aesthetics axis scores, and DeckDesign Likert scores; modules below threshold would be demoted from Tab. 1 to diagnostic use.Source: C DeckQuiz Human Alignment Study · p. 20, Table 3 · p. 20
• αjudge–human = 0.57 for Aesthetics, the lowest module alignment; all four modules cleared the α = 0.50 demotion threshold.Figure 7 pairs human inter-rater agreement with judge–human alignment for each module; Aesthetics was the first candidate for demotion if future re-calibration fell below 0.50.Source: C DeckQuiz Human Alignment Study · p. 20, Figure 7 · p. 21

D Deck-Level Human Pairwise Study

• The blinded deck-level study compared ReDeck with DeepPresenter on 45 GPT-5.4 tasks using majority-voted human choices for presentation quality and faithfulness.Three annotators judged each pair; ReDeck’s preference rate was aggregated by task-level majority vote with Wilson 95% confidence intervals, and no task had a majority “no preference” outcome.Source: D Deck-Level Human Pairwise Study · p. 21, Table 4 · p. 21

E SpatialCheck VLM-Judge Validation

• SpatialCheck’s SCR judge scores rendered slides uniformly across PPTX, PDF, and HTML, classifying four hard-layout violation families before computing each deck’s clean-slide fraction.A GPT-5.4 vision model evaluated one PNG slide at a time for text clipping, element overlap, out-of-bounds elements, and image clipping; two non-author raters adjudicated reference labels for 150 sampled slides.Source: E SpatialCheck VLM-Judge Validation · p. 22, Table 5 · p. 22
• 0.89 primary clean/dirty balanced accuracy; DOM-SCR preserved the full VLM-SCR ordering with 92% slide-level agreement (κ = 0.83 [0.79, 0.86], MCC = 0.84 [0.80, 0.87]).Across 100 tasks × 3 seeds × 6 variants, DOM-SCR had a deck-level bias of −0.8 points [−1.3, −0.3] and MAE 1.4 points [1.0, 1.8]; this was a second implementation of the same violation taxonomy, not an independent task definition.Source: E SpatialCheck VLM-Judge Validation · p. 22

F PresentBench Domain-Wise Results

• ReDeck’s win rate is 0.72 on scientific academia and 0.66–0.68 on distant talk and advertising domains.Scientific academia overlaps with the in-domain training distribution; the lower distant-domain win rates indicate partial transfer of the feedback loop’s benefits to non-scientific content.Source: F PresentBench Domain-Wise Results · p. 22
• The pooled baseline order is DeepPresenter > SlideTailor > SlideGen, with SlideTailor narrowly ahead of DeepPresenter on talk (0.51 vs. 0.50).This sole per-domain second-versus-third swap contributes to the aggregate Kendall’s τ = 0.87; the passage attributes the talk result to SlideTailor’s template-driven priors transferring better to less content-dense talk slides.Source: F PresentBench Domain-Wise Results · p. 22

G External Evaluation

• Against DeepPresenter, ReDeck gains 0.032 [0.018, 0.046] on LQ, 0.023 [0.010, 0.036] on DFaith, 0.022 [0.008, 0.036] on DFid, 2.6 percentage points [0.8, 4.4] on Quiz, and 3.2 points [1.5, 4.9] on Aes.Quiz measures answerability from extracted slide content rather than complete source fidelity, so these results triangulate rather than duplicate DeckQuiz’s four modules.Source: G External Evaluation · p. 23, Table 6 · p. 23

H Transfer to PPTX

• On GPT-5.4 PPTX refinement, ReDeck gains +5.0 on Fid., +24.0 on SCR, +0.50 on Aes., and +6.0 on Des., supporting transfer to a second editable geometry/rendering stack.The SCR gain has a 95% CI of [19.1, 28.9]; the result does not establish support for PDF/OCR-only editing, Keynote, animations, masking, or font-fallback behavior.Source: H Transfer to PPTX · p. 23, Table 7 · p. 23

I Judge Configuration

• All DeckQuiz VLM judges use a fixed GPT-5.4 backbone, while agent calls retain each system’s released defaults; deterministic judge decoding, structured outputs, and one retry standardize scoring.When GPT-5.4 generates a deck, judge and generator share a model family, a residual self-preference limitation that the authors validate against human annotations; parse failures after retry are under 0.3% of main-run calls.Source: I Judge Configuration · pp. 23–24

J Pilot Audit, Taxonomy Mapping, and Probe Coverage

• A 100-turn pilot yielded 264 hand-labelled failures; the six-family taxonomy met its acceptance criteria of at most 5% residual “other” failures and Cohen’s κ ≥0.6.Two authors independently coded failures using a written codebook, then adjudicated disagreements; the figure reports that both criteria were satisfied.Source: J Pilot Audit, Taxonomy Mapping, and… · p. 24, Figure 8 · p. 25

K DeckQuiz Construction and Protocol

• DeckQuiz builds paper-grounded questions for 100 arXiv papers without official slides, using four reviewer-style types—Contribution, Method, Experiment, and Limitation—and 2,000 total questions.Questions are generated from each paper alone, paired with answer-bearing source spans of at most 200 tokens, and screened so correct answers cannot be guessed from world knowledge; the bank has 500 questions per type.Source: K DeckQuiz Construction and Protocol · pp. 25–26
• DeckQuiz uses four-choice, same-discipline distractors and rendered-deck judging to test whether slides communicate paper-specific facts, with per-type accuracy separating omissions from rubric-phrasing mismatches.Each paper has five questions per type, and the judge reads rendered deck PNGs; the comparison with existing benchmarks emphasizes DeckQuiz’s per-loop-component attribution alongside spatial, fidelity, and design signals.Source: K DeckQuiz Construction and Protocol · p. 26

L DeckDesign Principles and Scoring

• DeckDesign scores five principles across deck-level narrative progression and consistency and slide-level segmentation, signal and emphasis, and structural correspondence.Each principle uses a 0–3 scale; the two deck-level principles and three slide-level principles are weighted by their counts (2/5 and 3/5) to form the Des. score.Source: L DeckDesign Principles and Scoring · pp. 26–27

M Probe Library Composition and Per-Task Call Budget

• The probe library contains 237 atomic checks in 40 groups across six families, and each task has a matched 150-call cap combining agent, turn-level, and step-level calls.At each turn, scheduler πP selects 3–5 checks from rendered PNGs, open issues, modified slides, and the full catalog; it rechecks each open issue on a modified slide and carries unselected issues forward.Source: M Probe Library Composition and… · p. 28

N Initial-Deck Attribution and Task-Paired Inference

• Formal GPT-5.4 inference averages three seeds per task and uses task-bootstrap CIs, paired permutation tests, and Holm correction; Bgeom's 37 checks gate hard, while Bvisual's 85 style checks gate soft.Gemini-3.1 and Claude-4.6 are descriptive replications.Source: N Initial-Deck Attribution and… · p. 28
• ReDeck improves SCR over the DeepPresenter loop by 19.5 points from ReDeck T0 and 20.2 points from DeepPresenter T0, with no clear draft-by-loop interaction (∆∆= −0.7 points).Seven of eight metric contrasts have 95% CIs excluding zero; the +0.03 Aes. contrast from DeepPresenter T0 is inconclusive.Source: N Initial-Deck Attribution and… · p. 28, Table 11 · p. 28
• All 12 ReDeck-versus-baseline contrasts have CIs excluding zero and remain significant after Holm correction; the smallest is Aes. versus DeepPresenter (+0.17 [0.02, 0.32], adjusted p = 0.024).The 12 comparisons use task-bootstrap 95% CIs after averaging seeds within task.Source: Table 11 · p. 28, N Initial-Deck Attribution and… · p. 29

O Absolute Inference Cost

• At the 2175-call endpoint, ReDeck exceeds cost-scaled DeepPresenter by +11.6/+13.5/+0.44/+8.2 on Fid./SCR/Aes./Des., with all four task-bootstrap CIs excluding zero.The tested scaling policy raises DeepPresenter SCR from 72.0 to 78.0 but does not close the gap; the cap matches mean post-T0 cost, not exact per-task spend.Source: O Absolute Inference Cost · p. 29, Table 13 · p. 29
• ReDeck-Lite reduces mean cost from $1.74 to $0.28 and latency from 3586 s to 1450 s, while changing Fid./SCR/Aes./Des. by −0.8/−1.7/−0.03/−0.8.It retains high-level planning and final submission on GPT-5.4 while routing recurrent edits and both verification channels to GPT-5.4-nano; routing mitigates but does not remove the full system's cost trade-off.Source: O Absolute Inference Cost · p. 29, Table 12 · p. 29

P Per-Layer Regression Honesty

• DeckDesign reports T0-to-Tfinal changes across deck-level P1/P5 and slide-level P2/P3/P4; any negative layer delta is flagged, and a negative principle within a positive layer is reported as a sub-row.Published baselines receive no artificial zero-delta rows because they lack ReDeck's explicit repair trajectory.Source: P Per-Layer Regression Honesty · pp. 29–30
• All five per-principle DeckDesign deltas are positive (+5.8 to +11.2), with slide-level gains (+9.5) exceeding deck-level gains (+6.7), so aggregate +8.4 conceals no principle regression.Step-level render feedback resolves slide-local layout violations affecting P2 and P3, while content choices constrain P4 and the deck-level pair.Source: P Per-Layer Regression Honesty · p. 30, Table 14 · p. 30

Q Trajectory Event-Flow

• Issue events comprise open →resolved when addressed, resolved →regressed when a later edit revives an issue, open →persisted when its family is not probed, and open →dropped when it is no longer detected.Dropped issues are not explicitly addressed.Source: Q Trajectory Event-Flow · p. 30
• Under ReDeck, 68% of issue-turn pairs are open →resolved, 22% are open →persisted, and 8.2% are resolved →regressed; per-step observation rolls back 73% of potential regressions.Persisted issues indicate that the adaptive scheduler deprioritised stable issues.Source: Q Trajectory Event-Flow · p. 30
• Resolved →regressed rises to 23.5% under All-probes-every-turn, as repeated style findings prompt edits that break previously clean layout.Style-motivated edits propagating into layout regressions, rather than any single hard family failing, drive open-issue drift.Source: Q Trajectory Event-Flow · pp. 30–31
• Per-family traces show Bgeom collapses fastest under ReDeck, while Bvisual grows under All-probes-every-turn and its rewrite edits lift Bgeom alongside it.Under ReDeck, soft families A, C, and E decay steadily once issues enter the persistent list and structural choices stabilise.Source: Q Trajectory Event-Flow · p. 30
• Family D remains small in absolute count but has the highest per-issue reopen rate under All-probes-every-turn because edits to neighbouring text break verified claims.Step-level render feedback therefore matters beyond layout, particularly for factual correctness.Source: Q Trajectory Event-Flow · p. 31

R Diagnostic Mini-Ablations

• 17.3-point SCR and 6.1-point Des drops, with Fid. approximately flat, follow replacing rendered-image input with source code in the full ReDeck loop.The step-level rendering channel remains intact; drops are relative to default ReDeck in Tab. 2.Source: R Diagnostic Mini-Ablations · p. 31, Table 15 · p. 31
• Self-Refine [8] and Reflexion [9] differ by less than one point on every module, within seed-level variability, making critique modality—not judge source—the binding constraint.These implementations disaggregate the per-turn NL-critique row of Tab. 2.Source: R Diagnostic Mini-Ablations · p. 31