DECKBench v1.0.0

Official benchmark release accompanying the KDD 2026 paper: DECKBench: Benchmarking Multi-Agent Frameworks for Academic Slide Generation and Editing

What's Included

• 294 paper–slide pair URLs spanning CVPR, ECCV, ICLR, and ICML across multiple years
• Evaluation code for both tasks: paper-to-slide generation (Task 1) and multi-turn iterative editing (Task 2)
• User simulation pipeline with three personas (Granular Analyst, Balanced Editor, Executive) for generating realistic multi-turn editing instructions
• Baseline editor agent built on the OpenAI Agents framework, with an abstract base class for plug-in custom agents
• Full metric suite: perplexity, faithfulness, text/figure similarity, layout quality, DTW, transition similarity, ΔDTW, ΔTransSim
• Tutorial notebook for the user simulator (user_simulator_tutorial.ipynb)
• Pre-generated baseline decks available on HuggingFace

Citation

If you use DECKBench in your research, please cite:

@misc{jang2026deckbench,
  title        = {DECKBench: Benchmarking Multi-Agent Frameworks for Academic Slide Generation and Editing},
  author       = {Daesik Jang and Morgan Lindsay Heisler and Linzi Xing and Yifei Li
                  and Edward Wang and Ying Xiong and Yong Zhang and Zhenan Fan},
  year         = {2026},
  eprint       = {2602.13318},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2602.13318}
}