PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.00226v1cs.AI

TL;DR

Paper-to-slide generation must accommodate presenter-specific goals, but existing iterative systems either keep critique internal or restrict user feedback to post-hoc deck editing. ConvDeck distributes conversational refinement across outline and slide stages using agents that can reason, speak, and act, and its experiments report higher user-goal satisfaction while remaining competitive on slide-quality metrics. The prototype still relies heavily on simulated users and has overhead, visual-refinement fragility, and risks of factual errors or omissions.

  • Problem

    Existing systems limit user participation by keeping refinement internal or introducing feedback only after the complete deck, restricting control over structural presentation decisions.

  • Method

    ConvDeck is a multi-agent pipeline with stage-specific conversational loops over the outline and rendered deck, using agents that can reason, speak, clarify, and act.

  • Results

    ConvDeck achieves higher user-goal satisfaction than post-hoc editors while remaining competitive on slide-quality metrics.

  • Takeaways & Limitations

    Stage-aligned feedback lets users address high-level structure before rendering and slide-level presentation details after rendering.

  • Takeaways & Limitations

    Experiments rely primarily on an LLM-based user simulator, while larger human studies remain future work; the system also adds overhead and can produce visual or factual errors.

Abstract

from arXiv · show

Automatic academic paper-to-slide generation is inherently iterative, because creating an effective presentation requires repeated cycles of generation, critique, and revision. Recent multi-agent systems partially acknowledge this through internal critique-and-revise loops, while conversational approaches allow users to refine generated slide decks through dialog. However, these refinement processes either remain largely closed to the user or introduce feedback only after a complete deck has been produced, limiting the user's ability to participate in the iterative refinement of narrative flow, content allocation, and presentation emphasis. To address this gap, we introduce ConvDeck, a multi-agent pipeline for conversational paper-to-slide generation that distributes interaction across the pipeline through stage-specific loops, allowing users to iteratively refine both the presentation outline and the final slide deck at the stages where each kind of decision is made. These loops are driven by a refinement mechanism in which agents can think, speak, and act, enabling them to either directly apply edits or respond conversationally to clarify user feedback and discuss revision options. Our evaluation shows that stage-specific conversational feedback improves user-goal satisfaction while preserving narrative coherence, content quality, and visual presentation.

1 Introduction

Paper-to-slide generation requires decisions about content, narrative, emphasis, and detail that depend on presenter goals. ConvDeck addresses limitations of single-pass and post-hoc approaches by placing conversational refinement at outline and slide stages.

  • Paper-to-slide generation must adapt content selection, narrative organization, emphasis, and detail to the presenter’s goals, audience, and time budget.The same paper can support multiple valid decks, including methodology-focused, results-oriented, and instructional presentations.
  • Recent systems use multi-stage pipelines and internal critique-and-revise loops, but these processes largely keep iterative refinement within the system.Specialized components address aspects such as visual design and narrative flow.
  • Conversational systems commonly collect user feedback only after the complete deck is generated, forcing one post-hoc stage to absorb all requested changes.This gives users direct feedback control but places structural and surface-level revisions together at the end.
  • Post-hoc slide revision is inefficient for structural problems because outline structure, content allocation, and narrative order were determined earlier in the pipeline.Surface-level language control cannot reliably address decisions encoded in earlier intermediate representations.
  • ConvDeck introduces stage-specific conversational loops for revising high-level outline decisions before rendering and slide-level details after rendering.Outline refinement covers narrative flow, section emphasis, ordering, and coverage; slide refinement covers content, visual organization, figures, layout, and other details.
  • Its agents can reason, speak, and act, enabling direct edits or clarification and discussion when feedback is ambiguous or underspecified.The evaluation measures user-goal satisfaction alongside narrative coherence, content quality, and visual presentation.

2 Related Work

Related work spans multi-agent slide-generation pipelines and conversational systems that refine completed decks. ConvDeck differs by distributing interaction across outline and slide-generation stages and by evaluating presenter-specific goal satisfaction.

  • Multi-agent paper-to-slide generation: Recent automatic slide-generation systems decompose planning, content selection, construction, and visual refinement into specialized agents.Some systems are reference- or template-driven, while others emphasize narrative or discourse structure.
  • Multi-agent paper-to-slide generation: Multi-agent approaches incorporate generation, critique, and revision cycles that reflect iterative document authorship.These systems use internal iteration to polish generated decks.
  • Conversational paper-to-slide generation: Conversational slide-generation systems let users provide natural-language feedback, but typically introduce conversation only after a complete deck has been generated.Post-hoc edits entangle surface properties such as font size with structural properties such as topic ordering.
  • Conversational paper-to-slide generation: ConvDeck distributes interaction across outline generation and slide generation, aligning each refinement with the pipeline component best suited to perform it.This contrasts with systems that restrict users to editing final slides.
  • Evaluation: Prior evaluations emphasize slide quality, reference-deck similarity, or learning outcomes, whereas ConvDeck adds explicit presenter-specific user-goal satisfaction evaluation.The evaluation uses diverse randomly sampled requirements for generated decks.

3 Conversational Paper-to-Slide Generation via ConvDeck

ConvDeck transforms a paper PDF into a deck through preprocessing, outline planning, conversational refinement, slide generation, and rendered-slide refinement. Its agents update intermediate artifacts through dialogue, clarification, tools, and localized edits.

  • Pipeline overview: ConvDeck accepts a paper PDF plus optional target-audience and presentation-duration inputs and processes them through five sequential stages.The stages are preprocessing, outline generation, conversational outline refinement, slide generation, and conversational slide refinement.
  • Preprocessing: Preprocessing extracts structured paper text and an asset dictionary containing figures and tables with captions.These artifacts ground subsequent generation, and Docling performs the parsing and asset extraction.
  • Outline Generation: Outline Generation builds a discourse-aware initial outline using an RST-based discourse tree, a global commitment, and a planner–reviser–critic cycle.The commitment incorporates target audience and presentation duration into the deck’s high-level intent.
  • Conversational Outline Refinement: Conversational Outline Refinement lets users revise high-level structure and narrative from a preview of slide titles and discussion ideas before slides are rendered.An Outline Refiner applies operations including adding, editing, splitting, and merging outline elements.
  • Refinement mechanism: The refinement agent uses Think+Speak for clarification and Think+Act for executing requested modifications, with retrieval available for requested information beyond the paper.Edits are implemented as local updates rather than regenerating the full outline.
  • Slide Generation: Slide Generation converts the refined outline into a draft deck using a constructor for content, figures, tables, layouts, and structured slide specifications, followed by aesthetic refinement.The constructor produces structured JSON for the slide specification.
  • Conversational Slide Refinement: Conversational Slide Refinement operates on rendered slide images, enabling localized correction of overflow, undersized figures, crowded layouts, and other visual issues.The stage observes actual visual output rather than only the slide specification.
  • Editable representation: ConvDeck translates slide specifications into JavaScript compiled into PPTX, providing an editable structured state for refinement.PPTXGenJS is used to compile the JavaScript representation into a PPTX file.

4 Experimentation

ConvDeck evaluates stage-specific conversational refinement through unmet-goal studies, per-stage improvements, ablations, natural-interaction studies, and overall-quality comparisons. Results show substantial gains in presenter-goal satisfaction while preserving quality, with performance depending on capable generation backbones and adding modest computational cost.

  • Evaluation Setup: The evaluation uses five categories of presenter-specific goals, sampled through an unmet-goal protocol and assessed with LLM judges and human users.The protocol compares baseline satisfaction, refinement outcomes, per-stage gains, and standard slide quality.
  • Baseline Comparison: 79% of initially unmet goals are satisfied overall, with ConvDeck leading on Content Inclusion/Exclusion at 90% and Narrative Structure at 82%.ConvDeck’s structural advantage is attributed to outline-level changes propagating across the deck before rendering.
  • Human Evaluation: 98% Success Rate shows that ConvDeck renders reliably, while human studies report 96.0% satisfaction for initially unmet goals and 93.0% for participant-defined goals.The natural-interaction studies include 35 sessions for initially unmet goals and 20 sessions for participant-defined goals.
  • Per-Stage Improvement: Outline refinement raises outline-relevant satisfaction from 37.7% to 92.0% (+54.3 points), while slide refinement raises slide-relevant satisfaction from 68.4% to 86.4% (+18.0).The largest outline gain is Narrative Structure (+71), while slide gains concentrate on Figure/Table Usage (+36) and Style & Wording (+29).
  • Overall Quality and Cost: ConvDeck is preferred over every baseline on stronger backbones, reaches 87% against author-prepared slides under the GPT judge on GPT-5, and costs $0.82 per paper.On GPT-5, ConvDeck wins 91% over AutoSlides; margins compress on Qwen3-VL, indicating that refinement amplifies rather than replaces backbone capability.

5 Conclusion

ConvDeck distributes conversation across two stage-specific refinement loops: one over the outline before rendering and another over the rendered deck. Experiments show higher user-goal satisfaction than post-hoc editors while remaining competitive on slide quality and affordable.

  • ConvDeck distributes conversational refinement across the outline before rendering and the rendered slide deck.The two loops align feedback with the pipeline stage best suited to apply each revision.
  • Stage-aligned refinement achieves higher user-goal satisfaction than post-hoc editors can structurally reach.
  • ConvDeck remains competitive on slide-quality metrics while adding conversational refinement.
  • The pipeline remains affordable despite distributing conversation across multiple refinement stages.

6 Limitations

ConvDeck's evaluation relies primarily on an LLM-based user simulator, while larger human studies remain future work. The system also incurs conversational overhead, can fail during visual edits, and is weaker with smaller open-source backbones; generated decks require expert verification.

  • Experiments primarily use an LLM-based user simulator, with a complementary human study rather than a larger human evaluation.Larger human studies are left for future work to test whether simulated feedback matches broader presenter behavior.
  • Multi-turn conversational refinement adds computational overhead relative to single-pass generation.
  • VLM-based visual refinement occasionally emits inaccurate figure coordinates, causing layout failures.
  • Generation quality drops on the Qwen3-VL-32B backbone because smaller models have less capacity for structured editing operations.Results on this backbone are interpreted as a lower bound, with scaling to larger open-source backbones left for future work.
  • ConvDeck is a research prototype rather than a replacement for expert review, because generated decks may contain factual errors, omissions, or overstated claims.The authors recommend manual verification before presentation or public release.

B Implementation Details

The implementation uses shared models, structured paper and asset representations, speak–act conversational loops, and editable slide specifications. Baselines share models, feedback limits, and rendered information, while ConvDeck retains retrieval and intermediate-representation editing as its contributions.

  • Generation backbones: ConvDeck uses GPT-5, Gemini 3 Pro, and Qwen3-VL-32B-Instruct as interchangeable backbones across its agents.A single selected backbone drives ConvDeck agents from Stages 2 through 5.
  • Preprocessing: Preprocessing converts the PDF into markdown and an asset dictionary containing figures, tables, captions, dimensions, aspect ratios, and citation mappings.
  • Conversational refinement configuration: Both conversational stages use a speak–act loop in which a user simulator supplies feedback or “Ready,” and refiners apply localized edits.The simulator runs for up to five rounds with early stopping, while refiners have per-round action budgets.
  • Rendering pipeline: The rendering pipeline uses 16:9 slides, 14 reusable layout templates, and PPTXGenJS to create structured, editable slide programs.The Slide Refiner operates on this editable state during slide refinement.
  • Evaluation fairness: Conversational baselines use the same models, feedback-round limits, and rendered information as ConvDeck.Retrieval and intermediate-representation-based editing remain unavailable to baselines because they are ConvDeck contributions.

C User Study Details

The user studies evaluate goal satisfaction through sampled presenter requirements, human judgment, and natural interaction, while the implementation details describe the five-stage generation pipeline. The system builds discourse-aware outlines, commits to high-level intent, constructs asset-grounded slides, and applies aesthetic refinement.

  • User Study: A human study samples 30 papers and asks participants to judge whether final revised slides satisfy unmet goals, with paper order randomized.Participants receive links to the original papers, and results are reported by category alongside judge-rating correlations.
  • Natural Interaction: Natural-interaction studies let participants interact directly with ConvDeck using a GPT-5 backbone.
  • Pipeline: The five-stage pipeline preprocesses the paper, generates a discourse-aware outline, refines it conversationally, renders slides, and conversationally refines the deck.
  • Narrative-Driven Outline Generation: The Discourse Parser builds hierarchical trees whose rhetorical relations guide content grouping and narrative ordering.Paragraphs become leaves, while nucleus-satellite and multinuclear relations organize higher-level spans.
  • Narrative-Driven Outline Generation: The Commitment Builder creates a five-field Global Commitment conditioned on target audience and presentation duration.Its fields include a snapshot, core content, talk contract, narrative spine, and section plan.
  • Narrative Refinement: The Slide Planner drafts outlines from rhetorical groupings, while the critic, judge, and reviser perform a critique cycle before later outline refinement.
  • Slide Generation: The Slide Deck Constructor selects assets, assigns one of 14 layouts, generates text, and records citations in a structured slide specification.The Aesthetic Refiner then adds visuals, balances density, colors text, and applies bold formatting.

G Discussion on Retrieval Tools

Retrieval is evaluated separately from ConvDeck’s stage-specific conversational mechanism. Retrieval provides a modest quality gain and has a measurable accuracy limitation, while the conversational mechanism remains the primary source of gains.

  • ConvDeck without retrieval still outperforms every baseline, indicating that gains arise primarily from the conversational mechanism.
  • Retrieval is always called when requested, and the final deck always includes the requested baseline slide.
  • 20% of retrieval cases return the wrong paper, causing the resulting slide to be based on that paper.Retrieval accuracy is the only imperfect step reported in the retrieval-goal test.

H ConvHTML Baseline Implementation

ConvHTML adapts an HTML slide-generation baseline with a refinement loop that mirrors ConvDeck’s conversational slide refinement. Its HTML representation enables direct fine-grained visual edits.

  • ConvHTML augments an HTML slide-generation baseline with a user-simulator refinement loop mirroring Conversational Slide Refinement.
  • HTML output enables fine-grained visual edits directly, avoiding coordinated XML updates required by the PPTXGenJS representation.
  • A VLM selects five currently unsatisfied evaluation goals after rendered slides are inspected.

I ArcBench Benchmark

ArcBench is a curated benchmark of oral paper–slide pairs with author-prepared references and dense visual content. Its predefined goals support evaluation of content, narrative structure, deck composition, figure and table usage, and style.

  • Benchmark construction: ArcBench contains 100 oral paper–slide pairs with author-prepared reference decks from top-tier CV/ML venues.ConvDeck uses the benchmark unchanged.
  • Benchmark construction: Every ArcBench paper has at least three figures and at least three tables, supporting assessment of content coverage, figure fidelity, and narrative quality.
  • Benchmark construction: ArcBench imposes oral-talk and content-density requirements unlike prior datasets targeting general scientific papers or adjacent slide-generation settings.
  • Goal design: The 50 predefined user goals are organized into five categories covering content inclusion/exclusion, narrative structure, deck composition, figure and table usage, and style and wording.
  • Goal design: Outline-relevant goals are addressed during Stage 3, while figure, table, and style goals are addressed during Stage 5.
  • Conversational evaluation: The simulator reviews the outline or rendered deck against quality criteria for at most five rounds or until it emits “Ready”.Stage 3 evaluates narrative flow, content coverage, and audience-duration fit; Stage 5 additionally evaluates result interpretation and visual communication.

L Overall-Quality Evaluation Protocol

Overall quality is evaluated with complementary VLM-judge rubric, pairwise preference, and content-coverage quiz protocols. Judge agreement is stronger for text-oriented dimensions than visual layout, and judge leniency materially affects comparisons with human references.

  • Evaluation protocol: The evaluation combines VLM-as-Judge rubric scores, pairwise A/B preference, and a Q/A quiz probing content coverage.
  • Evaluation protocol: Two independent VLM judges score each presentation over three iterations, with rubric means and pairwise majority votes reported.
  • Evaluation protocol: Each deck receives four 0–10 dimension scores based on ten binary checklist criteria per dimension.The dimensions are Text Quality, Narrative Flow, Visual Layout, and Visual–Text Alignment.
  • Inter-judge agreement: Text Quality, Narrative Flow, and Visual–Text Alignment show strong inter-judge correlations of r=0.94, r=0.85, and r=0.92, respectively.
  • Inter-judge agreement: Visual Layout agreement is lower at r=0.68 overall, dropping to r=0.18 for AutoSlides and r=0.28 for SlideGen.Overall-quality and narrative-flow pairwise verdicts show moderate Cohen’s κ agreement of 0.42 and 0.55.
  • Judge leniency: The GPT-5 judge reports a 55.3% win rate against author-prepared slides for GPT-5-generated decks, versus 16.5% for Gemini.Absolute-score differences between judges are +5 to +7 points, showing that automated comparisons depend heavily on judge choice.

M.4 Same-Family Self-Preference

The evaluation examines whether judges favor decks generated by their own model family, separating global leniency from same-family preference. Both judges show modest absolute preference, but GPT-5 exhibits a large asymmetric effect in pairwise comparison against author-prepared decks, motivating multi-judge reporting.

  • Cross-generator agreement: Figures 14 and 15 correlate per-paper scores between GPT-5, Gemini 3 Pro, and Qwen3-VL-32B generation backbones under GPT-5 and Gemini evaluators.Each figure contains three generator-pair panels.
  • Self-preference index: SPI measures same-family preference as the difference between Qwen and Gemini leniency gaps, with positive values indicating over-crediting of the judge’s own family.The neutral Qwen anchor estimates global leniency alone.
  • Absolute scoring: Both judges show a small but consistent same-family preference in absolute scoring.Table 16 reports the same-family self-preference index against the neutral Qwen anchor.
  • Pairwise human comparison: +17.0 points: the GPT-5 judge over-credits GPT-5-generated decks in the pairwise-vs-author-prepared comparison.The Gemini judge shows essentially no corresponding self-preference, at −0.3.
  • Evaluation implications: Judge choice matters most in pairwise comparison with author-prepared decks, where a single judge could drive an automated-decks-beat-human-slides claim.The authors therefore report both judges and pair their verdicts rather than averaging them.
  • Evaluation implications: Low inter-judge agreement and low cross-generator transfer for visual-layout scores make those dimensions a noisier complement to pairwise overall-quality comparisons.The evaluation leans on the more consistent overall-quality comparison for main conclusions.

N Per-Stage Token Breakdown

ConvDeck’s four LLM-driven stages distribute token use unevenly: input context dominates, slide-side stages process more visual material, and outline stages generate most output. The pipeline also documents conversational examples and several observed failure modes.

  • Token usage: 227.3K input and 54.8K output tokens per paper: the average full-run usage across ConvDeck’s four LLM-driven stages.The four stages are Outline Generation, Conversational Outline Refinement, Slide Generation, and Conversational Slide Refinement.
  • Token distribution: Input tokens dominate output by roughly four to one because agents repeatedly condition on the paper, outline or slide specification, previews, and accumulated feedback.Slide-side stages are somewhat heavier because they process figure and table assets and rendered images.
  • Token distribution: Output tokens concentrate in the outline stages, while slide stages emit fewer tokens because refiners apply localized edits instead of regenerating the deck.This reflects the use of editing functions during slide refinement.
  • Conversational overhead: The two conversational stages account for a substantial fraction of the budget, representing the cost of interactivity relative to single-pass generation.The end-to-end cost-quality trade-off is reported separately.
  • Refinement examples: Stage 3 examples show outline refinement adding a retrieved related-work slide or splitting an overloaded slide, pairing feedback with reasoning and an editing-function call.The examples illustrate how structural changes occur before slide rendering.
  • Failure cases: Figure 16 illustrates rendering defects including overflow, wrong figure placement, and unwanted figure removal.These are representative ConvDeck failure cases.
  • Failure cases: The refiner can over-edit by altering or dropping content that the feedback did not target, including an unmentioned figure during a split or merge.This occurs despite preservation rules.
  • Failure cases: Edits and layouts degrade on the weaker Qwen3-VL-32B backbone relative to GPT-5 and Gemini 3 Pro.This is identified as backbone-dependent quality.

Q Prompts

The prompt appendix specifies how ConvDeck’s conversational refiners and user simulator transform feedback into controlled outline and slide edits. It distinguishes content operations, visual overrides, global patches, and staged user-goal evaluation.

  • Q Prompts: The appendix identifies ConvDeck-specific prompts for conversational refiners, the user simulator, paper summarizer, evaluation judges, and ConvHTML, while Stage 2 and Stage 4 prompts are reused from ArcDeck.Runtime placeholders such as {{presentation}} are substituted with corresponding inputs.
  • Q.1 Conversational Refinement Prompts: The Outline Refiner receives a JSON slide list with title, content, and discussion_idea fields and rewrites the entire list while preserving narrative coherence.Its prompt frames the task as incorporating free-form user feedback into the current outline.
  • Q.1 Conversational Refinement Prompts: Allowed outline operations include reordering, targeted removal, adding requested background, splitting or merging slides, and editing titles or content.The prompt requires preserving slides and ordering unless the user specifically requests a change.
  • Q.1 Conversational Refinement Prompts: When background or related work is requested, the refiner uses arxiv_search with a concise topical query and omits extraction_query for ordinary related-work summaries.Targeted extraction is reserved for details such as metrics, baselines, ablations, or limitations.
  • Q.1 Conversational Refinement Prompts: The Slide Refiner edits both content and layout through slide-plan edits, per-slide visual overrides, and surgical global JavaScript patches.Visual controls include font sizes, text-box rectangles, image rectangles, and figure hiding.
  • Q.1 Conversational Refinement Prompts: Prompt rules preserve slide identity by subsection title, prohibit invented filenames, retain T14_2Text structure, preserve unmentioned content, and maintain LaTeX formatting.Per-slide overrides are preferred over global patches for localized changes.
  • Q.1 Conversational Refinement Prompts: The prompt maps common visual issues to scoped actions, such as template changes for one-image slides, overrides for isolated problems, and global patches for repeated template-wide problems.Resized rectangles must remain in bounds and preserve clearance from other elements.
  • Q.2 User Simulator Prompts: The user simulator gradually guides the assistant using presentation settings and goals, evaluating coverage, flow, audience fit, tone, detail, sectioning, slide count, and relevance.It can request additions, removals, reordering, combination, splitting, clarification, shortening, expansion, or reframing.
Loading 2609.00226v1…