WebGen-R1: 7B RL-Trained Model Rivals DeepSeek-R1 671B at Multi-Page Website Generation

arXiv·high signal

WebGen-R1 applies end-to-end reinforcement learning to project-level website generation using a cascaded multimodal reward combining structural guarantees, execution-grounded functional feedback, and vision-based aesthetic supervision. A 7B parameter model trained with this framework consistently outperforms open-source models up to 72B and rivals DeepSeek-R1 (671B) in functional success while substantially exceeding it in valid rendering and aesthetic alignment. Code and data are released — this is directly applicable to anyone building AI-assisted frontend tooling.