Reproduce “AeSlides: Incentivizing Aesthetic Layout in LLM-Based Slide Generation via Verifiable Rewards” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yiming Pan, Chengwei Hu, Xuancheng Huang, Can Huang, Mingming Zhao, Yuean Bi, Xiaohan Zhang, Aohan Zeng and 1 more
The AeSlides framework introduces a reinforcement learning approach that uses verifiable, rule-based metrics to directly optimize the aesthetic layout quality of LLM-generated slides. This method achieved substantial improvements in objective aesthetic metrics and human evaluation scores, matching leading proprietary models, while demonstrating the limitations of VLM-based aesthetic feedback for this task.
Key insights
- Verifiable, rule-based metrics are highly effective, accurate, and efficient for quantifying fundamental slide layout aesthetics, outperforming current VLMs which exhibited "aesthetic blind spots" for subtle details.
- Direct aesthetic supervision through verifiable rewards in RL is more effective than iterative agentic refinement with VLMs, which often provides unreliable feedback and incurs high computational costs.
- Maintaining KL divergence regularization during RL training is crucial to prevent policy collapse, ensuring aesthetic improvements are achieved without sacrificing the model's ability to generate diverse and semantically coherent content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.