PaperlayerView exact revision on arXiv

Source-linked AI summary

SlideLab: Audience-Centered Scientific Slide Generation and Evaluation

Vidushee Vats, Karun Sharma, Yuxia Wang

arXiv:2609.30294v2cs.CLcs.AI

TL;DR

Scientific presentations must do more than summarize papers: they need coherent sequencing and clear explanations for audiences encountering the work as it unfolds. SlideLab plans and iteratively refines research presentations, while ConfArena simulates audience interaction to evaluate them. In a blind human preference study, SlideLab was preferred on 77% of evaluated papers.

  • Problem

    Scientific slide systems still struggle to jointly ground content in the paper, organize a coherent narrative, and present visual elements, while common evaluations overlook how audience understanding develops during a talk.

    Source: §1 · p. 1

  • Method

    SlideLab plans a coherent narrative before progressively constructing and refining slides, while ConfArena simulates an audience following the presentation slide by slide and asking questions.

    Source: §1 · p. 1, §3.1 · p. 3, §4 · p. 4

  • Results

    SlideLab was preferred for 77% of evaluated papers in a blind human preference study against three representative open- and closed-source systems.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The work presents SlideLab and ConfArena as contributions toward research on scientific presentation generation and its evaluation.

    Source: §6 · p. 8

    Experiments are limited to conference-style presentations generated from research papers; applicability to educational, business, or technical talks remains for future work.

    Source: Limitations · pp. 8–9

Abstract

from arXiv · show

Scientific presentations are more than summaries of research papers. They need to present the work in a coherent sequence, explain the main ideas clearly, and help the audience follow the presentation. We present SlideLab, a training-free multi-agent framework for generating scientific presentations from research papers. SlideLab first plans the presentation narrative, then builds and iteratively refines a shared slide deck using agents for content planning, visual generation, layout refinement, and grounding verification. In a blind human preference study, SlideLab was preferred over both open-source and commercial systems on 77% of papers while using roughly 4 times fewer inference tokens than the strongest open-source baseline. We also introduce ConfArena, an audience-oriented evaluation framework that simulates a conference room and assesses presentations slide by slide. ConfArena matches human system rankings and detects injected presentation problems, including falsified numbers, degraded figures, dropped slides, and shuffled slide order.

1 Introduction

Scientific slide generation must organize and visually explain research for an audience encountering it for the first time, while evaluation must account for how understanding develops during a talk. SlideLab generates presentations through narrative planning and iterative refinement, and ConfArena evaluates them through simulated audience interaction.

1 Introduction

  • Scientific presentations must jointly select paper content and present it visually, while current systems still struggle with grounding, narrative organization, and visual elements.Audiences encounter the work for the first time as the talk unfolds, so decks must support understanding in sequence rather than merely summarize the paper.Source: §1 · p. 1
  • SlideLab plans a coherent presentation narrative before specialized agents progressively construct and refine the deck, emphasizing source faithfulness, organization, and visual aesthetics.The paper also describes the framework as training-free.Source: §1 · pp. 1–2
  • ConfArena simulates an audience following slides in sequence and asking questions when communication breaks down, assessing presentations beyond static deck-level evaluation.The paper reports strong agreement between the environment and human judgments.Source: §1 · p. 2
  • SlideLab was preferred for 77% of evaluated papers in a blind human preference study against three representative open- and closed-source systems.The framework is training-free and progressively constructs and refines slide decks.Source: §1 · p. 2

2 Related Work

Prior slide-generation research spans extractive summarization, editing existing decks, and end-to-end synthesis, while evaluation benchmarks mostly assess presentations as static artifacts. The paper positions ConfArena as addressing the overlooked sequential audience experience.

Early Extractive Methods

  • Early extractive methods selected and compressed salient document content but primarily operated on text and predated modern LLM-based systems.They established foundations for automatic slide generation.Source: §2 · p. 2

Editing-Based Generation

  • Editing-based systems adapt reference slides, templates, or structured deck objects, producing consistent designs but constraining layout flexibility toward template filling.PPTAgent edits retrieved slides; SlideGen uses agents for planning, layout selection, and refinement; Talk-to-Your-Slides edits structured objects.Source: §2 · p. 2

End-to-End Scientific Slide Generation

  • End-to-end systems synthesize presentations without reference decks, but approaches differ in scope: some omit presentation-level coherence while others target aesthetics alone.AutoSlides and DeepPresenter use multistage agentic pipelines for scientific presentations.Source: §2 · p. 2

Slide Deck Evaluation

  • Existing benchmarks use LLM judges, reference decks, or specialized failure-mode tests, yet assess presentations as static artifacts rather than sequential audience experiences.The paper positions ConfArena alongside, not as a replacement for, existing benchmarks.Source: §2 · pp. 2–3

3 SLIDELAB

SlideLab addresses narrative, visual, flow, layout, and grounding problems through a staged generation and refinement process. Its agents plan the presentation, add needed visuals, restructure slides, and inspect rendered layouts and factual support.

3 SLIDELAB

  • SlideLab reduces cost by 3.6× and token consumption by 4× compared to DeepPresenter.The framework responds to annotation feedback about poor narrative flow, grounding, layout aesthetics, and high cost.Source: §3 · p. 3

3.1 Structured Narrative Planning

  • A Planner reads paper sections, figures, and tables to propose presentation blueprints; a second agent critiques candidates and selects the most coherent flow.Each slide blueprint specifies its title, takeaway, content source, visuals, and whether to generate a new visual; removing the Planner causes the largest performance drop in the reported ablation.Source: §3.1 · p. 3

3.2 Concept Figure Design

  • SlideLab reuses source figures when sufficient and generates a diagram only when a concept needs one unavailable in the paper.The Visual Generator uses relevant paper sections and slide context to create a style-matched figure, then checks its fit and alignment on the slide.Source: §3.2 · pp. 3–4

3.3 Editorial Flow Restructuring

  • A post-generation Compositor can reorder or remove slides, merge nearly empty slides, reposition figures, and add visual slots to text-heavy slides.It does not redesign or create slides, and checks edits against the full paper.Source: §3.3 · p. 4

3.4 Layout Debugging and Grounding

  • A multimodal LayoutDebugger inspects rendered slides and can return corrected HTML for up to three rounds; a separate long-context LLM checks grounding against the paper.The checks target visual layout defects and unsupported claims, incorrect numbers, or other factual errors.Source: §3.4 · p. 4

4 ConfArena: Evaluation Environment

ConfArena simulates a conference presentation with an examiner and distinct attendee personas who follow the slides and participate in Q&A. It evaluates presentation events and aggregates measures of grounding, figures, design, narrative, and coverage.

4.1 Simulation Overview

  • ConfArena simulates a talk with an examiner and three attendee personas, who see only the slides and build understanding as the presentation progresses.The presenter answers questions using the full paper, while the examiner reads it as an evaluation reference.Source: §4.1 · p. 4, §4 · p. 4

4.2 Attendees

  • The expert reviewer, learner, and cross-field attendee focus respectively on scientific rigor, understandability, and accessibility without specialized background knowledge.Each persona uses a focus, question trigger, and short role description to represent a distinct audience perspective.Source: §4.2 · p. 4

4.3 Presentation

  • Before the talk, an examiner prepares a paper-based checklist; during each slide, the examiner checks grounding while attendees update understanding and decide whether to ask questions.Questions are posed to the presenter, and the examiner identifies whether they arose from omitted, unclear, unreadable, or out-of-scope material.Source: §4.3 · p. 5

4.5 Final Assessment

  • ConfArena aggregates observations into grounding and figure errors, design and figure-use scores, narrative errors, and key-point coverage.Metrics are rank-normalized within each paper and averaged with equal weight for an environment score from 0 to 1.Source: §4.5 · p. 6

5 Experiments

Experiments used ArcBench papers and a held-out set of machine-learning papers, comparing systems through automated benchmarks and blind human preference; ConfArena was also checked against human rankings and targeted presentation failures.

5.2 Human Preference Evaluation

  • Annotators criticized Manus for text-heavy slides, DeepPresenter for layout and factuality issues, and Kimi Slides for overcrowding despite generally well-structured decks.Kimi Slides received the fewest factual-error comments, while DeepPresenter received the most.Source: §5.2 · p. 6
  • 80% of annotations (96/120) favored SlideLab; it was selected for 23 of 30 papers, with Kimi Slides preferred for the other seven.Neither DeepPresenter nor Manus was selected.Source: §5.2 · p. 6, Table 2 · p. 7

5.3 Evaluation with ConfArena

  • SlideLab achieved the highest overall ConfArena score and led on every evaluation axis except figure errors, while maintaining high coverage without sacrificing factual correctness, visual quality, or structure.Kimi Slides had the fewest figure errors but lower source-paper coverage; DeepPresenter had more grounding errors and weaker narrative flow.Source: §5.3 · p. 7
  • Removing the Planner caused the largest ablation drop; removing the LayoutDebugger also significantly changed results, while removing the Compositor or using paper figures only caused smaller drops.The Planner removal left the Generator without a structured blueprint; the LayoutDebugger removal mainly affected visual design and illegible-figure penalties.Source: Table 3 · p. 7, §5.3 · p. 7

5.4 Performance by Existing Metrics

  • SlideLab scored highest on all PPTEval dimensions and SlidesGen-Bench, while ranking second on PresentBench behind Kimi Slides.The rankings were consistent with ConfArena’s, suggesting the improvements were not specific to the proposed evaluation environment.Source: Table 4 · p. 7, §5.4 · p. 7

Does ConfArena Reflect Human Judgment?

  • ConfArena alone detected all four planted failures on their intended axes; PPTEval lacks a missing-content metric, while PresentBench and SlidesGen-Bench failed to detect all four.PresentBench and SlidesGen-Bench have relevant metrics but did not detect every targeted failure.Source: Table 6 · p. 8, Does ConfArena Reflect Human Judgment? · p. 8
  • Humans ranked SlideLab first and Kimi Slides second, as did ConfArena, PPTEval, and SlidesGen-Bench; PresentBench instead ranked Kimi Slides first.Humans tied Manus and DeepPresenter.Source: Does ConfArena Reflect Human Judgment? · p. 8
  • Targeted perturbations mainly shifted their intended metrics: shuffled order raised narrative defects, incorrect values grounding errors, smaller figures readability penalties, and removing a method slide reduced coverage.The remaining metrics changed only slightly.Source: Table 7 · p. 8, Does ConfArena Reflect Human Judgment? · p. 8

6 Conclusion

SlideLab generates scientific presentations through narrative planning, progressive slide refinement, and multimodal grounding, while ConfArena evaluates them through audience-oriented slide-by-slide interaction and Q&A. Experiments, including blind human evaluation, show more effective presentations with substantially fewer inference tokens than existing systems.

  • SlideLab is a training-free framework that generates scientific presentations through structured narrative planning, progressive slide refinement, and multimodal grounding.ConfArena assesses presentations from the audience’s perspective through slide-by-slide interactions and Q&A.Source: §6 · p. 8

Limitations

The experiments cover conference-style presentations generated from research papers, while broader presentation domains remain a possible direction for future work.

Limitations

  • The experiments are limited to conference-style presentations generated from research papers.SLIDELAB is described as largely domain-agnostic, but this broader scope is not established by the reported experiments.Source: Limitations · p. 8
  • Narrative planning, grounding, and iterative refinement are design principles that may generalize beyond scientific presentations.Their applicability to educational, business, and technical presentations remains for future work.Source: Limitations · p. 9
  • Other presentation domains can differ in audience, communication objectives, and presentation style.These differences are cited as reasons to investigate broader applicability.Source: Limitations · p. 9

Ethical Considerations

SLIDELAB is intended to assist scientific communication, not replace human authorship or judgment, and generated slides require review before public use.

Ethical Considerations

A.1 SLIDELAB Configuration

The configuration lists the models and maximum rounds for SLIDELAB stages, whose tool-use loops produce a final output when their round limit is reached.

A.1 SLIDELAB Configuration

  • Table 8 lists the model and maximum-round configuration for each SLIDELAB stage.Source: §A.1 · p. 10
  • The Planner, Slide Generator, and Compositor each run as tool-use loops with specified round limits.Source: §A.1 · p. 10
  • When an agent reaches its round limit, it must produce its final output from what it has gathered so far.Source: §A.1 · p. 10

A.2 Model Configuration and Ablation Results

The supplementary experiments compare model configurations, system and component performance, evaluation variants, computational cost, human-rating procedures, and figure-related results.

A.2 Model Configuration and Ablation Results

  • Removing the Planner lowers the environment score from 0.74 to 0.53 under GPT-5.4 and from 0.78 to 0.62 under Gemini, the largest drop among component removals.Ablations used MiMo-V2.5-Pro throughout the pipeline, including planning, layout debugging, and grounding verification.Source: §A.2 · p. 10, Table 9 · p. 11

A.3 ConfArena Configuration

  • SlideLab achieves the highest ConfArena environment score and leads on five of six metrics, while Kimi Slides has the fewest grounding errors.The comparison uses Gemini-3.8-Flash for evaluation; SlideLab uses MiMo-V2.5-Pro for text agents, while PPTAgent and DeepPresenter use GPT-5.5.Source: §A.3 · p. 10, Table 10 · p. 12

A.4 Evaluation Configs

  • The configuration table lists planner candidate-plan generation with GPT-5.5 (60 calls), candidate critic and selector with GLM-5.2 (1), and plan-grounding verification with GPT-5.5 (1).Source: §A.4 · p. 11

A.5 Sequential Evaluation and Audience Ablations

  • Static-examiner and ConfArena scores correlate less for coverage and narrative errors; narrative errors are 0.80 versus 0.90 for Kimi Slides and SlideLab under static scoring, but 2.40 versus 3.94 under ConfArena.Removing Q&A or using only an examiner reduces reported narrative problems more than replacing attendee roles.Source: §A.5 · pp. 11–12, Table 12 · p. 12, Table 13 · p. 12

A.6 Computational Cost

  • SlideLab costs approximately $0.58 per paper, runs in under twelve minutes, and uses about 700K tokens per deck, roughly 4× fewer than DeepPresenter’s ∼3000K.The paper attributes the token difference mainly to reading the paper once for planning and generating slides against a fixed blueprint rather than repeatedly reflecting over the whole deck.Source: §A.6 · p. 12, Table 15 · p. 13

A.7 Human Evaluation Interface

  • Annotators choose the deck they would present, then rate it on six dimensions using a five-point scale.Annotators were known to the authors, participated voluntarily, and knew their annotations would be shown for analysis.Source: §A.7 · pp. 12–13, Table 16 · p. 13

A.8 Detailed Perturbation Scores

  • Kimi Slides has fewer figure errors than SlideLab in one comparison (0.17 vs. 0.26), while SlideLab scores higher on figure use (3.80 vs. 3.08).The appendix also describes Table 17’s baseline-versus-perturbed benchmark scores across 15 papers for assessing detection of injected failure modes.Source: §A.8 · p. 13, Table 17 · p. 14
Loading 2609.30294v2…