Abstract

Multi-agent systems built on large language models (LLMs) and multimodal large language models (MLLMs) have recently shown strong capabilities on complex compositional tasks. We apply this paradigm to paper-to-poster generation, a practical but time-consuming step in research communication. Most of existing methods ignore core design and aesthetic principles, and often produce posters that still require substantial manual refinement. To address these limitations, we propose PosterGen, a multi-agent framework that embeds fundamental design principles into a specialized agent workflow that mirrors professional designers. Our system incorporates collaborative agents to distill a paper's narrative (CuratorAgent), map content to a balanced spatial layout (LayoutAgent), apply a cohesive visual system of color and typography (StylistAgents), and render the final poster in PPTX format (Renderer). This design-centric approach produces posters that are both semantically grounded and visually compelling. To systematically evaluate visual quality, we introduce a comprehensive VLM-based rubric measuring information fidelity and aesthetic design, which we validate against human preferences in a user study. Experimental results show that PosterGen achieves both content fidelity and design aesthetics comparable to human-designed posters while significantly and consistently outperforming state-of-the-art methods.