ConvDeck: Conversational Paper-to-Slide Generation via Stage-Specific User Feedback
Tarik Can Ozden, Sachidanand VS, Furkan Horoz, Ozgur Kara, Dilek Hakkani-Tür, Junho Kim, James Matthew Rehg
cs.AI
Quick verdict
ConvDeck's stage-specific feedback loops lift user-goal satisfaction (outline 37.7→92%) and preserve deck quality, but gains depend on frontier LLMs.
Summary
ConvDeck is a five-stage multi-agent pipeline that turns a paper PDF into a presentation deck with user interaction embedded at two pipeline points rather than only after the deck is finished. After Docling preprocessing and ArcDeck-based outline generation, a Conversational Outline Refinement stage lets users reshape narrative structure before any slide is rendered; after slide generation, a Conversational Slide Refinement stage edits the rendered deck. Both refiners use a think-speak-act loop — clarifying questions or direct edits via local edit functions — plus arXiv/web retrieval, with PPTXGenJS code as editable state. Using an LLM user simulator on ArcBench, outline refinement lifts outline-stage goal satisfaction from 37.7% to 92.0% and slide refinement raises slide-stage goals from 68.4% to 86.4%; removing both conversational stages drops average satisfaction from 82% to 18% (Table 5, 20 papers). A small human study (35 and 20 sessions) reports 96.0%/93.0% satisfaction. On GPT-5/Gemini 3 Pro backbones ConvDeck is preferred over all baselines and reaches 87% preference versus author-prepared decks under the GPT-5 judge, at $0.82/paper and 940 s. Limitations: most evaluation relies on a simulated user; the GPT-5 judge shows +17-point same-family preference in the vs-author comparison; gains largely vanish on Qwen3-VL-32B; visual edits sometimes emit wrong coordinates.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
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
- Distributing conversational feedback to outline and slide stages raises user-goal satisfaction far above what post-hoc editing reaches (outline goals 37.7%→92.0%; slide goals 68.4%→86.4%; Sec 4.2).
- Each conversational stage improves mainly the goal categories it is positioned to act on, and removing both collapses average satisfaction from 82% to 18% (Table 5).
- Conversational refinement preserves standard slide quality: ConvDeck is preferred over all baselines on GPT-5/Gemini 3 Pro and is the automated method closest to author-prepared decks (87% under the GPT-5 judge; Sec 4.3).
- Stage-specific conversation adds modest overhead: $0.82/paper, 281K tokens, 940 s, roughly 10% over the non-conversational ArcDeck backbone (Fig 5).
- Adding interaction at the global-commitment stage yields no benefit (Table 6), bounding where in the pipeline interaction is useful.
Strengths
Stage-attribution evidence directly supports the core routing claim
Table 5 (20 papers) shows each conversational stage improves mainly the categories it is positioned to act on: removing Stage 5 collapses Figure/Table (75→10) and Style (70→0); removing Stage 3 erodes Narrative (90→70) and Style (70→40); removing both drops average satisfaction 82→18. Table 6 adds an honest null result: interaction at the global-commitment stage yields zero satisfaction gain for +4.8% tokens and +4.5% runtime.
Exemplarily candid judge-bias analysis
Appendix M quantifies inter-judge agreement (Visual Layout r=0.09 between GPT-5 and Qwen generators), a leniency gap (identical GPT-5 decks win 55.3% vs author-prepared under the GPT-5 judge but 16.5% under Gemini, Table 15), and same-family self-preference (SPI +17.0 for the GPT-5 judge, Table 16), then changes evaluation practice accordingly (M.5: both judges reported throughout).
Round-wise dynamics give actionable deployment guidance
Full cost/latency transparency and published prompts
Weaknesses
The headline quality number rests on the self-preferring judge
Evaluation is dominated by an unvalidated LLM user simulator
Gains concentrate on frontier proprietary backbones
No visible uncertainty quantification and one unexplained anomaly
The slide-stage editing mechanism is fragile
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ConvDeck (arXiv 2609.00226; UIUC) addresses conversational paper-to-slide generation. The central claim is that user interaction should be *distributed* across a generation pipeline — one conversational loop over the outline before any slide is rendered (Stage 3), another over the rendered deck (Stage 5) — rather than confined to post-hoc editing of a finished deck, where structural decisions (narrative order, content allocation) cannot be revised cheaply. Three contributions are stated: the two-loop conversational pipeline built on the authors' ArcDeck backbone, a think–speak–act refinement mechanism, and a user-goal satisfaction evaluation. Supporting numbers: outline-stage goals rise from 37.7% to 92.0% and slide-stage goals from 68.4% to 86.4% after conversation (Sec 4.2); ConvDeck is preferred over all baselines on GPT-5/Gemini 3 Pro backbones and reaches 87% preference against author-prepared decks under the GPT-5 judge (Sec 4.3), at $0.82/paper and 940 s — roughly 10% overhead over non-conversational ArcDeck (Sec 4.5).
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 ConvDeck and how is it different from other AI slide generators?
ConvDeck is a multi-agent system that converts a research paper PDF into a presentation deck. Unlike systems that let you chat only after the full deck is generated, it embeds two conversational refinement loops inside the pipeline: one on the slide outline before any slide is rendered (narrative, ordering, content coverage) and one on the rendered slides (figures, layout, wording). Its agents can ask clarifying questions or directly apply edits, and can pull extra material from arXiv/web search.
How much does ConvDeck improve user-goal satisfaction?
In the paper's simulated-user experiments on ArcBench, outline-stage goals that were initially unmet rose from 37.7% to 92.0% satisfied after outline conversation, and slide-stage goals from 68.4% to 86.4% after slide conversation. Removing both conversational stages drops average satisfaction from 82% to 18%. A small human study (35 and 20 sessions) reported 96.0% and 93.0% satisfaction.
How much does ConvDeck cost and how long does it take per paper?
The paper reports about $0.82 and 281K tokens (227K input / 54K output) per paper with 940 seconds end-to-end runtime — roughly 10% overhead over its non-conversational backbone ArcDeck ($0.54, 451 s), and cheaper than the PPTAgent baseline ($1.39).
Does ConvDeck work with open-source models?
Largely no, per the paper's own results: on the open-source Qwen3-VL-32B backbone, pairwise win rates compress toward 50%, meaning conversational refinement amplifies capable frontier backbones (GPT-5, Gemini 3 Pro) rather than substituting for them. Visual refinement is also fragile — the refiner sometimes emits wrong figure coordinates.
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.
- Sampling gaps in the provided excerpt cover Tables 1-3, parts of Sec 4.1-4.2 (evaluation-protocol details), App C (human-study details), and App G (retrieval-ablation results, cut mid-sentence); judgments relying on those sections are lower confidence.
- No statistical-test reporting is visible in the provided text; it may exist in the sections hidden by the sampling gaps.
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.