WebViSE researchers report lower retrieval cost for website design guidance

Researchers report in a preprint that WebViSE, a framework for extracting reusable design guidance from websites, outperforms non-hierarchical language-model retrieval baselines at 39.09% of their estimated retrieval cost using Qwen3.8-Max. The framework organizes that guidance into a hierarchy that it automatically revises alongside the guidance descriptions, helping match user instructions to relevant design skills. Models can use the skills as guidance when generating websites, while open-source models can incorporate them through synthesized training data. The researchers report improved visual quality and interaction fidelity across benchmarks and human evaluations, and release the extracted skills, hierarchy and code.