SlideTailor: Personalized Presentation Slide Generation for Scientific Papers

Wenzheng Zeng*, Mingyu Ouyang*, Langyuan Cui*, Hwee Tou Ng†,

National University of Singapore

AAAI 2026

📄 Paper   |   🤗 Daily paper   |   🤗 Dataset   |     🖼️ Poster   |     ▶️ Video   |   ▶️ Slides   |   ▶️ X (Twitter)  

This repository contains the official implementation of the AAAI 2026 paper "SlideTailor: Personalized Presentation Slide Generation for Scientific Papers".

🔆 Overview

We argue that presentation design is inherently subjective. Users have different preferences in terms of narrative structure, emphasis, conciseness, aesthetic choices, etc.

So in this work, we ask: Can we better model such diverse user preferences for personalized paper-to-slides generation?

We make the following contributions:

• Task: We introduce and properly define a new task that conditions paper-to-slide generation on user-specified preferences.
• System: We propose a human behavior-inspired agentic framework, SlideTailor, that progressively generates editable slides in a user-aligned manner.
• Evaluation: We construct a benchmark dataset that captures diverse user preferences, with meticulously designed interpretable metrics for robust evaluation.
• Open Source: We release the source code and data to the community.

🔥 News

• [2025-11] Our work is accepted to AAAI 2026!
• [2025-11] We release our code and data to the community!

🛠️ Environment

1. Create a new conda environment

conda create -n slidetailor python=3.11

2. Install python dependency

pip install -r requirements.txt

3. Install other dependencies

#If using Ubuntu
sudo apt-get install libreoffice poppler-utils

4. Prepare your API key (named as api_key.json)

{
"openai_api_key": "sk-proj-1234"
}

📦 Dataset

Please prepare the data and the corresponding config files according to the instructions on the PSP Dataset page.

🤖 Inference

Remember to modify the relevant paths in the following script before running it.

sh run.sh

⚖️ Evaluation

Remember to modify the relevant paths in the following script before running it.

sh eval.sh

🙏 Acknowledgments

We would like to thank PPTAgent for their valuable contribution and for making their codebase available to the community.

📖 Citation

If you find this work helpful, please kindly cite our paper:

@inproceedings{slidetailor,
  title={SlideTailor: Personalized Presentation Slide Generation for Scientific Papers},
  author={Zeng, Wenzheng and Ouyang, Mingyu and Cui, Langyuan and Ng, Hwee Tou},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}