🤖 AI

SlideGen: Collaborative Multimodal Agents for Scientific Slide Generation

This paper introduces SlideGen, a collaborative multimodal agent framework that transforms scientific papers into well-structured presentation slides by coordinating narrative planning, multimodal grounding, and layout composition, while also proposing a new geometry-aware density metric to evaluate visual balance and effectiveness.

Original authors: Xin Liang, Zhilin Zhang, Xiang Zhang, Haoran Su, Yiwei Xu, Siqi Sun, Chenyu You

Published 2026-08-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xin Liang, Zhilin Zhang, Xiang Zhang, Haoran Su, Yiwei Xu, Siqi Sun, Chenyu You

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, scientists and researchers spend countless hours translating dense, technical papers into presentation slides. This is a necessary ritual for conferences, classrooms, and seminars, where complex ideas must be distilled into a handful of clear, visually organized pages. The challenge is not merely summarizing text; it is a delicate act of storytelling. A presenter must decide which narrative to tell, which figures to highlight, and how to arrange text and images so they do not crowd one another. If the layout is poor, the audience becomes confused or fatigued, regardless of how brilliant the underlying research might be. For decades, this process has been done almost entirely by hand, a time-consuming task that struggles to keep pace with the growing volume of scientific output.

The core difficulty lies in the fact that creating a good slide deck requires two different kinds of thinking at once. First, one must understand the long, detailed story of the research paper. Second, one must possess a sense of visual design to arrange that story into a format that is easy to read. Most existing computer programs that attempt to automate this task focus only on the text, often producing slides that are cluttered, repetitive, or visually unbalanced. They treat the creation of a presentation as a simple summarization job, missing the crucial step of designing how the information looks on the page.

To solve this, a team of researchers has developed a new system called SlideGen. Instead of relying on a single computer program to do everything at once, they built a collaborative team of specialized digital agents. Think of these agents as a group of experts working together in a studio, where each person has a specific job. One agent acts as the director, reading the entire paper to outline the story and decide what the main points should be. Another agent acts as a curator, carefully selecting the right charts, graphs, and equations to support those points. A third agent functions as a designer, choosing from a library of pre-made layouts to arrange the text and images so they fit together perfectly without overlapping. Finally, a fourth agent writes the speaker notes, providing the script that connects the slides into a smooth, spoken narrative.

The researchers tested this system on 200 recent scientific papers from leading computer science conferences. They found that the collaborative approach produced slides that were significantly better than those generated by other methods. The new system created decks that were not only faithful to the original research but also visually balanced and easy to follow. To measure this, the team developed a new way to score the slides based on how well the content filled the page, avoiding both empty space and visual clutter. Their results showed that the slides generated by this team of agents matched human judgments of quality much more closely than previous attempts.

In a study involving 30 people with experience in academic presentations, the results were striking. When asked to choose a slide deck to use as a draft for their own talks, nearly all participants preferred the ones created by the new system over those made by other automated tools. Many rated the quality of these computer-generated slides as better than what most humans could produce, with some even calling them expert-level. The system succeeded because it did not just copy and paste information; it reasoned about the structure of the presentation and the visual design simultaneously. By breaking the complex task of slide creation into smaller, specialized steps handled by different agents, the researchers showed that computers can learn to handle the artistic and narrative aspects of scientific communication, turning a tedious manual chore into an efficient, automated process.

Technical Summary: SlideGen

Problem Statement

The automatic generation of presentation slides from scientific papers is a uniquely challenging multimodal task that extends beyond simple text summarization. Effective slide decks require not only the extraction of key information but also the joint reasoning over long contexts, the selection and alignment of multimodal evidence (figures, tables, equations), and the deliberate visual design of layouts to ensure clarity and narrative flow.

Existing approaches largely frame this problem as a text generation or content summarization task. While recent methods have improved content selection and some support multimodal inputs, they often treat visual layout as a secondary concern or rely on post-hoc adjustments. Consequently, generated slides frequently suffer from repetitive layouts, poor visual balance between text and figures, and a lack of coherent narrative structure across pages. Furthermore, existing evaluation protocols often rely on text-based metrics (e.g., ROUGE) or VLM-based judges that only partially capture visual organization, sparsity, and fragmentation.

Methodology

The authors propose SlideGen, a collaborative multimodal agent framework designed to coordinate narrative planning, multimodal grounding, and principled layout composition. Unlike monolithic models or rigid formatting rules, SlideGen orchestrates a pipeline of six specialized vision-language agents organized into four sequential stages:

Narrative Planning (Outliner):
The Outliner agent performs global document parsing and generates a high-level narrative structure. It determines the presentation outline, allocating key ideas into a coherent sequence of slides, and generates concise titles and summaries for each segment.
Multimodal Grounding (Mapper & Formulizer):
- Mapper: Assigns visual assets (figures and tables) to specific slides based on semantic alignment with the narrative claims, providing justifications for each assignment.
- Formulizer: Addresses the specific challenge of grounding equations, which lack captions. It crops equations as high-resolution images to preserve typesetting and matches them to slides based on surrounding textual context, avoiding artifacts from LaTeX re-rendering.
Layout Composition (Arranger):
The Arranger agent assembles each slide by selecting a layout template from a predefined library of 19 fixed-slot templates. This library covers diverse structures (e.g., text-only, image-left/right, multi-image grids, formula-centric). The agent selects the optimal template $\ell_i$ for a slide $s_i$ by maximizing compatibility with the content profile (number of bullets, images, equations, and aspect ratios). Crucially, the use of fixed slots prevents element overlap and ensures editability in standard PPTX formats. If a content assignment is infeasible for any template, the system signals the Mapper to redistribute assets.
Global Refinement (Refiner & Speaker):
- Refiner: Performs deck-level adjustments, including merging consecutive text-only slides into two-column layouts, handling text overflow via bounded font scaling and content compression, and extracting a color theme from the source paper's figures to ensure visual cohesion.
- Speaker: Generates presenter notes for each subsection, synthesizing a smooth oral narrative that explains transitions and the rationale behind visual choices, independent of the bullet points.

Evaluation Protocol:
To address the lack of robust visual metrics, the authors introduce Geometry-Aware Density (GAD). This metric quantifies layout quality by combining two components:

Occupancy ( $OM_i$ ): Measures the fraction of the slide canvas covered by content, penalizing deviations from a target density $\tau$ .
Fragmentation Reward ( $FR_i$ ): A concave quadratic function that rewards an optimal number of content elements ( $M^\star$ ), penalizing both sparse layouts and cluttered ones.
The GAD score is calibrated against human judgments using leave-one-deck-out cross-validation.

Key Contributions

SlideGen Framework: A modular, multi-agent system that decomposes slide generation into distinct roles (planning, grounding, layout, refinement) to produce editable PPTX files without requiring reference decks.
Four-Dimensional Evaluation: A standardized protocol evaluating slide generation across (i) visual aesthetics (via GAD), (ii) communication effectiveness (SlideQA), (iii) holistic coherence (VLM-as-Judge), and (iv) textual coherence (Perplexity).
Geometry-Aware Density (GAD): A novel, layout-centric metric that quantifies visual clutter, sparsity, and fragmentation, showing strong correlation ( $\rho = 0.820$ ) with human judgments.
Extensible Layout Library: A fixed-slot template system that prevents element overlap by construction while allowing for diverse slide compositions.

Results

The system was evaluated on a benchmark of 200 peer-reviewed papers from top AI venues (ICLR, ICML, NeurIPS). SlideGen was compared against various baselines, including end-to-end generators (GPT-4o/5 HTML/Image) and multi-agent workflows (PPTAgent, PosterAgent).

Performance: SlideGen consistently outperformed all baselines across all four evaluation dimensions. It achieved the highest SlideQA scores (indicating better content coverage and reasoning), the best VLM-as-Judge scores (design and coherence), and the lowest perplexity (text fluency).
Layout Quality: SlideGen achieved the best GAD scores, demonstrating superior balance between occupied and empty space compared to baselines that often produced either sparse or cluttered layouts.
User Study: In a study with 30 participants (including PhD students and researchers), 93.3% preferred SlideGen's output over competitive baselines. 73.3% rated the quality as "better than most human-made slides," and 6.7% rated it as "top-tier expert quality."
Ablation: Removing the Outliner caused the largest drop in content coverage and language quality, while removing the Arranger caused the most significant degradation in layout density, confirming the necessity of the staged, specialized agent design.

Significance

The paper argues that effective slide generation requires multimodal design reasoning rather than just text summarization. By decomposing the task into collaborative agents, SlideGen bridges the gap between document understanding and scientific communication. The work demonstrates that explicit layout reasoning and multimodal grounding are critical for producing presentations that are both faithful to the source material and effective for human audiences. The proposed GAD metric offers a principled way to evaluate visual organization, addressing a gap in current automated slide generation research.