SlideTailor: Personalized Presentation Slide Generation for Scientific Papers

1 Introduction

• Prior generation frameworks often overlook individual preferences, although users differ in narrative structure, emphasis, conciseness, and aesthetic choices.The introduction notes that prior work has explored multimodal document-to-slides generation, but typically treats it as direct conversion.Source: §1 · p. 1
• The task conditions paper-to-slides generation on user preferences for narrative flow, emphasis, conciseness, and visual style.Content preferences affect narrative and emphasis; aesthetic preferences govern layout, backgrounds, decorative elements, and style.Source: §1 · p. 1
• A paper-slides sample pair and a .pptx template provide implicit signals of users’ content-structuring and aesthetic preferences.These inputs are convenient to provide, but their embedded preferences are implicit, entangled, and unlabeled.Source: §1 · pp. 1–2
• SlideTailor distills preferences into a profile, reorganizes salient paper content to match them, and structures that content into a coherent slide outline.The outline specifies each slide’s intended message and supporting points.Source: §1 · p. 2
• Chain-of-speech plans narration during outline construction, while template planning and editing produce editable .pptx slides.The paper links anticipated speech with improved slide coherence and clarity, and identifies video presentations as a downstream application.Source: §1 · p. 2
• Experiments report that SlideTailor better aligns with user intent and produces higher-quality slides than existing approaches.The introduction also presents a benchmark of diverse user preferences and interpretable metrics for evaluation.Source: §1 · p. 2

2.1 Document-to-Slides Generation

• Prior document-to-slides work has often treated generation as text summarization, while later methods incorporated visual and layout elements.The cited works include Li et al. 2021, Sun et al. 2021, Costa, Amaro, and Gonçalo Oliveira 2023, Maheshwari et al. 2024, Cachola et al. 2024, Xu and Wan 2021, Fu et al. 2022, and Bandyopadhyay et al. 2024.Source: §2.1 · p. 2
• Most multimodal document-to-slides methods lack diverse user-preference constraints; Persona-Aware-D2S uses four fixed categories, while PPTAgent focuses on layout.The paper positions its formulation as covering both content and layout preferences.Source: §2.1 · p. 2

2.2 Conditional Summarization

• Unlike related preference-conditioned summarization, which includes four predefined preferences in Mondal et al. 2024b, this work targets diverse, fine-grained preferences for slide generation.Other cited work covers personalized recommendations and controllable style or length.Source: §2.2 · p. 2

3 Task Formulation

• Content preferences govern narrative flow, detail, emphasis, and omission, while aesthetic preferences govern layout, theme, colors, typography, and decoration.Source: §3 · p. 2
• Users provide a paper-slides sample pair for content preferences and a .pptx template for aesthetic preferences, rather than detailed textual instructions.The two preference types are relatively orthogonal and can be paired modularly; the task assumes users have a general sense of desired outcomes and self-consistent preferences.Source: §3 · p. 3
• Given a paper, the task is to generate slides that meet general production-quality requirements and follow the preferences embodied in the sample pair and template.The function F jointly performs conditional paper summarization and slide generation.Source: §3 · p. 3

4.1 Key Challenges

• Unlabeled sample pairs and templates provide rich but entangled preference signals, while generation must align content and aesthetics harmoniously.Source: §4.1 · p. 3

4.2 Implicit Preference Distillation

• A dual-branch process infers content preferences from sample paper-to-slides transformations and aesthetic preferences from template roles and parsed geometry, combining them into a symbolic profile.Source: §4.2 · pp. 3–4

4.3 Preference-Guided Slide Planning

• Three agents reorganize paper content to user priorities, create slide-wise outlines with planned speech, and match each slide to a template layout.Source: §4.3 · p. 4

4.4 Slide Realization

• A layout-aware agent maps titles, text, and visuals to template elements, and code applies edits while preserving theme and editability.Source: §4.4 · p. 4

4.5 Downstream: Video Presentations

• Chain-of-speech pairs slides with scripts that could support personalized narration and synchronized presentation videos, including remote teaching or prerecorded talks.A short voice sample could preserve vocal identity through zero-shot text-to-speech.Source: §4.5 · pp. 4–5

• PSP explicitly represents diverse content-structuring and aesthetic preferences, unlike prior benchmarks centered on direct conversion.Source: §5 · p. 5
• The collection includes 50 paper-slides pairs and 10 templates alongside 200 target papers from AI and scientific venues spanning multiple fields.Source: §5 · p. 5
• PSP is reported as the largest benchmark, with up to 100,000 unique input combinations and support for open-ended preference modeling.Source: §5 · p. 5

• The evaluation uses complementary preference-based metrics for user alignment and preference-independent metrics for overall slide quality.Source: §6 · p. 5
• Preference-based measures assess topic coverage, ordering, structural alignment, and aesthetic alignment with the supplied examples and template.Source: Preference-Based Metrics · p. 5
• Table 1 compares existing paper-to-slides generation benchmarks, while Table 2 reports PSP dataset performance for our framework’s two backbone variants versus three state-of-the-art baselines, averaged over 50 target papers.Source: Table 1 · p. 6, Table 2 · p. 6
• Preference-independent measures judge how clearly slides convey the paper and their overall visual appeal.Source: Preference-Independent Metrics · p. 6
• Metrics are normalized to 0–100, with GPT-4.1 used for evaluations and an average reported as Overall.Source: Preference-Independent Metrics · p. 6

7.1 Experimental Setup

• The primary evaluation sampled 50 target papers, each paired with a randomly selected paper–slides example and .pptx template from the benchmark.All systems were constrained to generate 10 slides.Source: §7.1 · p. 6
• The zero-shot MLLM evaluation used GPT-4.1 as both vision and language model, with additional SlideTailor runs using open-source Qwen models.The Qwen models were served through LMDeploy on NVIDIA H200 GPUs.Source: §7.1 · p. 6

7.2 Experimental Results

• 75.8% overall score: SlideTailor with GPT-4.1 ranks highest and outperforms all baselines across preference-based and preference-independent metrics.The method also performs competitively with open-source Qwen2.5 and Qwen2.5VL models, without model-specific tuning.Source: §7.2 · p. 6
• $0.665 per 10-slide deck with GPT, versus $0.016 with Qwen, averaged across five sampled instances.The cost estimates use OpenRouter API pricing as of October 13, 2025.Source: §7.2 · p. 7
• SlideTailor’s qualitative outputs combine text with paper-derived figures, while baselines show issues including missed visuals, style mismatch, and template artifacts.The reported comparison notes that AutoPresent’s generated images may be misleading and PPTAgent can leave unreasonable blank space or placeholder elements.Source: §7.2 · p. 7
• 81.63% win rate: human evaluators preferred SlideTailor over PPTAgent, and human ratings had an average Pearson correlation of 0.64 with MLLM evaluations.Four graduate students evaluated 30 unique cases, each rated by two evaluators.Source: §7.2 · p. 7

7.3 Ablation Studies

• 66.4% → 47.3%: disabling chain-of-speech lowered overall performance, especially general content quality.The authors link narration-aligned slide planning to improved clarity and informativeness.Source: §7.3 · p. 7, Table 3 · p. 7
• Around 10%: removing content-preference distillation degraded coverage, flow, and content structure, as well as other preference-based metrics.The ablation results use a 30-sample subset of the PSP dataset.Source: §7.3 · p. 7, Table 3 · p. 7

8 Conclusion and Limitations

• SlideTailor distills implicit preferences and progressively generates editable slides, using chain-of-speech to connect slide planning with oral narration.The paper also introduces a benchmark that simulates diverse preferences and interpretable metrics for evaluation.Source: §8 · p. 7
• The benchmark covers scientific papers only; the authors identify business reports, educational materials, and advertising as domains for possible extension.They also describe end-to-end multimodal training and more human-aligned evaluation protocols as future directions, noting MLLM perception gaps and self-bias.Source: §8 · p. 7

Appendix

• Supplementary materials cover qualitative results, human evaluation, implementation, benchmark metrics, baseline details, and downstream applications.Source: Appendix · p. 10

A Qualitative Results and Analysis

• SlideTailor’s generated slide flows closely match the sample deck’s content structure and overall style.Figures 6 and 7 show the full inputs and slide-wise flows; additional comparisons appear in Figures 8–10.Source: A Qualitative Results and Analysis · p. 10
• The generated slides combine text and visuals, use template-derived colors and subtle background shapes, and include plots extracted from the target paper.The reported plots match their accompanying text, unlike the generated graphics attributed to AutoPresent.Source: A Qualitative Results and Analysis · p. 10
• SlideTailor follows template backgrounds, palettes, fonts, grids, and motifs, and mirrors the sample’s concise, example-led style.The analysis says it introduces tables and charts at key points and reserves extended discussion for qualitative comparisons and challenges.Source: A Qualitative Results and Analysis · p. 10
• SlideTailor reproduces almost all sections of the sample research-talk flow in the correct order, while PPTAgent preserves only partial macro sections and reorders two sections.The sample flow runs from title and motivation through methods and evaluation to conclusion and future work.Source: A Qualitative Results and Analysis · p. 10

B.1 Detailed Setup

• Four graduate-student annotators evaluated anonymized SlideTailor and PPTAgent outputs across 30 cases, with each case rated by two annotators.The evaluation included preference-based and preference-independent metrics and an overall preference choice; Tables 4 and 5 are captioned as human scoring and overall preference comparisons.Source: §B.1 · p. 10, Table 4 · p. 10, Table 5 · p. 10

B.2 Detailed Scoring Results and Analysis

• In 81.63% of cases, humans preferred SlideTailor; it also consistently outperformed PPTAgent across all human-rated metrics.Source: §B.2 · p. 10
• Pearson r values exceeded 0.6 across metrics between GPT-4.1 ratings and human scores, indicating substantial correlation.Source: §B.2 · p. 10, Table 6 · p. 16
• Human and MLLM scores differed on the same SlideTailor results; human evaluators were generally stricter, particularly on fine-grained visual details.Source: §B.2 · p. 16, Table 8 · p. 16

B.3 Inter-Annotator Agreement

• Average absolute differences between annotators were below 1 point, and 82.5% of case ratings differed by less than 1 point on the 1–5 scale.Source: §B.3 · p. 16, Table 7 · p. 16

C.1 Implicit Preference Distillation

• Content and aesthetic preference distillation use separate instructions, including one for inferring a latent content mapping function.Source: §C.1 · p. 16, Figure 11 · p. 18, Figure 12 · p. 19

C.2 Preference-Guided Slide Planning

• Planning instructions cover paper reorganization, slide-wise outline generation with chain-of-speech, and template-aware layout selection.Source: §C.2 · p. 16, Figure 13 · p. 20, Figure 14 · p. 21

D.1 Preference-based Rubrics

• The preference-based MLLM evaluation uses prompts for assessing content-structure and aesthetic similarity.Source: §D.1 · p. 16, Figure 16 · p. 23, Figure 17 · p. 24

D.2 Preference-independent Rubrics

• The preference-independent MLLM evaluation uses prompts for assessing content and aesthetic quality.Source: §D.2 · p. 16, Figure 18 · p. 25, Figure 19 · p. 26

E.1 Prompt for ChatGPT

• The ChatGPT baseline uses GPT-4o through the web interface, with the target paper, sample pair, and template uploaded alongside a slide-generation prompt.Source: §E.1 · p. 16, Figure 20 · p. 27

E.2 Prompt for AutoPresent

• AutoPresent uses GPT-4.1 and receives extracted raw text from the target paper, sample paper, and slides because it accepts only plain-text input.Source: §E.2 · pp. 16–17, Figure 21 · p. 27

E.3 PPTAgent

• PPTAgent is run using its official open-source implementation.Source: §E.3 · p. 17

F.1 Conditional vs. Unconditional Generation

• With default conditional inputs representing general presentation conventions, SlideTailor can also generate preference-independent slides.Source: §F.1 · p. 17