SlideTailor: Personalized Presentation Slide Generation for Scientific Papers

AI summary

Overview

• Research area: Natural Language Processing, specifically conditional summarization and multimodal document-to-slides generation for scientific papers.
• Technical level: Advanced. The work assumes familiarity with LLM/VLM-based agentic pipelines, benchmark construction, and LLM-as-a-judge evaluation.
• Scope: The paper defines a new task (preference-guided paper-to-slides generation), proposes an agentic framework called SlideTailor, introduces a 100,000-combination benchmark called PSP, and reports automatic and human evaluations against three baselines.

What This Paper Is About

Existing paper-to-slides systems treat generation as a direct one-size-fits-all document conversion, which ignores the fact that presentation design is subjective and that users differ in narrative structure, emphasis, conciseness, and aesthetics. The authors introduce the task of conditioning slide generation on user-specified preferences, supplied not as text instructions but as a paper-slides example pair (for content preferences) and a .pptx template (for aesthetic preferences). The goal is to produce editable, personalized slides that align with the user's implicit preferences, plus a companion speech script.

Key Contributions

1. A new task formulation. Preference-guided paper-to-slides generation, formalized as S = F(D, (D_ref, S_ref), S_tmpl), where the content preferences come from a paper-slides sample pair and the aesthetic preferences from a .pptx template. Preferences are split into two orthogonal dimensions: content preferences (narrative flow, level of detail, emphasis or omission) and aesthetic preferences (layout, background theme, color palette, typography, decorative elements).
2. SlideTailor, a human behavior-inspired agentic framework. A training-free, three-stage pipeline that distills implicit, unlabeled preferences into an explicit symbolic profile, plans slides under that profile, and realizes them by editing the provided template so the output remains editable .pptx.
3. A chain-of-speech mechanism. During slide-wise outline construction, the system simultaneously drafts the oral narration for each slide, aligning slide content with anticipated speech. This is reported to improve coherence and to enable downstream video presentations.
4. The PSP benchmark and interpretable metrics. A dataset with 200 target papers, 50 distinct paper-slides pairs, and 10 academic templates, yielding up to 200 × 50 × 10 = 100,000 unique input combinations, plus four preference-based and two preference-independent metrics.

Main Findings

• No system exceeds 80% overall. On the PSP dataset (Table 2), the best overall score is the authors' GPT-4.1 variant at 75.80, which the paper cites as evidence of the task's difficulty.
• The full framework leads across the board. Ours (GPT-4.1) scores 74.47 Coverage, 66.65 Flow, 72.80 Content Structure, 98.00 Aesthetic (preference-based), 67.64 Content, 75.24 Aesthetic (preference-independent), and 75.80 Overall. It outperforms ChatGPT (Overall 62.86), AutoPresent with GPT-4.1 (48.78), and PPTAgent with GPT-4.1 (67.30).
• The open-source backbone is competitive without tuning. Ours (Qwen2.5 + Qwen2.5VL) reaches 70.19 Coverage, 62.16 Flow, 68.41 Content Structure, 92.80 Aesthetic, 48.84 Content, 72.84 Aesthetic, and 69.21 Overall, ahead of two of the three baselines.
• Content preference distillation matters most for preference alignment. Removing it drops Coverage to 65.80, Flow to 56.83, and Content Structure to 54.67 in the 30-sample ablation (versus 74.82, 68.38, and 66.00 for the full system), a degradation of around 10% on those metrics.
• Chain-of-speech matters most for general content quality. Disabling it drops the preference-independent Content score from 66.40 to 47.33 (described as 66.4% → 47.3%) and Overall from 74.31 to 69.91.
• Humans prefer the proposed system. In a study with four graduate students, 60 independent ratings over 30 unique cases, and each case scored by two evaluators, SlideTailor won 81.63% of comparisons against PPTAgent (40 wins, 9 losses, 11 ties); PPTAgent's win rate was 18.37%. Mean human scores (1-5 scale) were 3.42 for Ours versus 2.85 for PPTAgent.
• Automatic evaluation correlates only moderately with humans. The average Pearson correlation between human and GPT-4.1 (MLLM) ratings is 0.638 (0.683 Content Structure, 0.639 Aesthetic, 0.602 Content, 0.626 Aesthetic). MLLM scores run higher than human scores (average 3.90 versus 3.42 on the 1-5 scale).
• Inter-annotator agreement is reported. Average absolute difference across metrics is 0.84, with 82.50% of ratings falling within a 1-point difference.
• Cost is modest. Over five sampled instances each producing a 10-slide deck, the average cost is $0.665 for the GPT version and $0.016 for the Qwen version, based on OpenRouter API pricing as of October 13, 2025.
• Baseline failure modes are documented. AutoPresent cannot reflect aesthetic preferences from text-only input and produces generated images not faithfully derived from the paper; ChatGPT often omits figures and tables; PPTAgent preserves templates well but misaligns with content structure and sometimes leaves placeholder elements or large blank areas.

Methodology in Plain English

The framework runs in three stages intended to mirror how a person prepares a talk.

Stage 1 — Implicit preference distillation. Users provide a paper-slides example pair and a template. A large language model (for example GPT-4.1) reads the pair and infers how that user abstracts and organizes content, producing a content preference profile covering narrative flow and section-level emphasis, omission, and formatting. In parallel, a vision-language model inspects the template to identify the functional role of each slide-level and element-level component, supplemented by metadata parsed from the raw .pptx file such as bounding box and image positions. The union of both profiles is the conditioning context.

Stage 2 — Preference-guided slide planning. Three LLM-powered agents work in sequence. A paper reorganizer restructures the input paper according to the content preferences, producing a presentation-oriented document rather than a generic summary. A slide outline designer then segments that content into a coherent slide-by-slide outline specifying each slide's intended message and visual cues; at the same time, the chain-of-speech mechanism drafts the spoken narration for each slide. A template selector then matches each planned slide to one slide from the user's template based on the aesthetic schema.

Stage 3 — Slide realization. A layout-aware agent maps planned content (titles, text, visuals) onto concrete template elements such as text boxes and image placeholders, modifying, replacing, or inserting elements as needed. A code agent then generates executable code that applies these edits directly to the .pptx file, preserving the original layout and theme and leaving a fully editable result.

Evaluation. The benchmark supplies both preference-based metrics (Coverage via subtopic intersection-over-union, Flow via Normalized General Levenshtein Distance, Content Structure as a 1-5 LLM judgment, and Aesthetic as a 1-5 VLM judgment of screenshots) and preference-independent metrics (Content and Aesthetic, both 1-5 MLLM judgments). All metrics are normalized to a 0-100 scale, and GPT-4.1 serves as the judge for all evaluations. The main experiment randomly samples 50 papers, pairs each with one randomly selected sample pair and one randomly selected template, and constrains every method including the baselines to 10 slides.

Why This Matters

• Impact on research: It reframes slide generation from a fixed document-conversion problem into a subjective, conditional generation problem, and it supplies the task definition, a benchmark, and metrics so others can build on it. Compared with prior customization work such as Persona-Aware-D2S (limited to four fixed preference categories) and PPTAgent (layout only), the formulation is open-ended and spans both content and aesthetics.
• Real-world applications:
  • Turning a researcher's paper into a conference talk deck that follows the style of their earlier talks and their institution's template.
  • Generating video presentations: the chain-of-speech output (T) can be converted to personalized narration by existing zero-shot text-to-speech systems, optionally with an identity-preserving talking head, supporting remote teaching or pre-recorded conference talks.
  • Producing pre-recorded or asynchronous lecture and seminar material at scale.
  • Producing tailored summaries of technical papers for audiences with different backgrounds, driven by whichever example pair is supplied.
• Industry relevance: The system is training-free, works with either a proprietary model (GPT-4.1) or open-source models (Qwen2.5-72B-Instruct and Qwen2.5-VL-72B-Instruct served through LMDeploy on NVIDIA H200 GPUs), and its reported cost of $0.665 (GPT) or $0.016 (Qwen) per 10-slide deck makes it practical for content-production workflows. Because outputs are standard editable .pptx files, they slot into existing authoring tools.

Future Directions

• Broaden beyond scientific papers. The benchmark covers only scientific papers; the authors suggest extending it to business reports, educational materials, and advertising content.
• Move beyond training-free pipelines. The framework is currently training-free, and the authors flag end-to-end multimodal training for preference-guided slide generation as a promising direction.
• Improve human-aligned evaluation. MLLM-based judgment still deviates from human judgment, and the authors report that MLLMs lack fine-grained human perception and exhibit self-bias; cross-judge setups (for example Qwen judging GPT-based models) appear more consistent with human ratings. Designing better evaluation protocols remains open.
• Scale the downstream video pipeline. The video presentation extension is described as only a step toward realization, with implementation details deferred to the project website rather than fully reported in the paper.

Target Audience

Researchers and practitioners working on document-to-slides generation, conditional and controllable summarization, multimodal content generation, and LLM/VLM agent design. It is also relevant to developers building presentation-authoring or meeting-and-lecture automation tools who need to understand both the modeling approach and the evaluation pitfalls of MLLM-as-a-judge scoring. Readers without a background in LLM agent pipelines will find the framework description accessible, but the benchmark and metric design will be most useful to those already working in the area.

Authors’ abstract

Automatic presentation slide generation can greatly streamline content creation. However, since preferences of each user may vary, existing under-specified formulations often lead to suboptimal results that fail to align with individual user needs. We introduce a novel task that conditions paper-to-slides generation on user-specified preferences. We propose a human behavior-inspired agentic framework, SlideTailor, that progressively generates editable slides in a user-aligned manner. Instead of requiring users to write their preferences in detailed textual form, our system only asks for a paper-slides example pair and a visual template - natural and easy-to-provide artifacts that implicitly encode rich user preferences across content and visual style. Despite the implicit and unlabeled nature of these inputs, our framework effectively distills and generalizes the preferences to guide customized slide generation. We also introduce a novel chain-of-speech mechanism to align slide content with planned oral narration. Such a design significantly enhances the quality of generated slides and enables downstream applications like video presentations. To support this new task, we construct a benchmark dataset that captures diverse user preferences, with carefully designed interpretable metrics for robust evaluation. Extensive experiments demonstrate the effectiveness of our framework.