DeepSlides: AI generates beautiful slides without templates using multi-agent RL

Researchers Zhiyao Cui and colleagues have unveiled DeepSlides, a novel framework for automatic slides generation that abandons traditional template-based approaches. Instead of relying on fixed layouts or directly emitting executable code, DeepSlides introduces a hierarchical workflow that first designs the slide page's visual structure and narrative flow, then generates the implementation code. This design-first, code-later approach allows large language models to exercise creative layout design without the constraints of predefined templates.

To train the system, the team built SlideDesign, a dataset specifically curated for slide generation tasks. They then employed a multi-agent reinforcement learning paradigm to train two SlideQwens models — one for slide design and one for implementation. Experimental results show DeepSlides outperforms baseline methods on evaluated metrics and achieves superior performance in human preference evaluations. The dataset and code are publicly available, enabling further research and practical adoption.