WebGen-R1: 7B RL model generates multi-page websites, beats 72B models and rivals 671B DeepSeek-R1

WebGen-R1, an end-to-end RL framework from Juyong Jiang, trains a 7B model to generate functional and aesthetically aligned multi-page websites. It uses a scaffold-driven structured generation paradigm and a cascaded multimodal reward combining structural, functional, and aesthetic feedback. The model outperforms open-source models up to 72B and rivals DeepSeek-R1 (671B) in functional success, while exceeding it in valid rendering and aesthetic alignment.

• #reinforcement-learning
• #code-generation
• #webgen-r1
• #llm