Ueberblick
July 2026 paper on a lightweight vision-language critic for generated web interfaces. The authors unify 19 principles spanning WCAG 2.2 accessibility, deceptive design, and cognitive usability, then train on roughly 10,000 generated pages with verified synthetic violations.
Zusammenfassung
Why This Paper Adds New Signal:
1. Quality beyond correctness: Evaluates whether generated interfaces are usable and responsible, not only whether their code runs
2. Broad design coverage: Combines accessibility, deceptive-pattern, perception, cognition, and interaction principles
3. Reproducible data pipeline: Injects and verifies known violations in diverse LLM-generated Tailwind pages
4. Efficient critic: Improves a 4B vision-language model from 36% to 84% micro-F1 across the 19-principle benchmark
5. Generation-loop utility: Positions the critic as an audit gate, training-data filter, and reward signal for better UI generators
1. Quality beyond correctness: Evaluates whether generated interfaces are usable and responsible, not only whether their code runs
2. Broad design coverage: Combines accessibility, deceptive-pattern, perception, cognition, and interaction principles
3. Reproducible data pipeline: Injects and verifies known violations in diverse LLM-generated Tailwind pages
4. Efficient critic: Improves a 4B vision-language model from 36% to 84% micro-F1 across the 19-principle benchmark
5. Generation-loop utility: Positions the critic as an audit gate, training-data filter, and reward signal for better UI generators
Tags
ui-evaluationaccessibilitydesign-qualityvision-language-modelreinforcement-learning