Research Paper
A research paper exploring how lightweight vision-language models can be trained to identify accessibility, deceptive design, and usability violations in AI-generated interfaces.
As coding agents increasingly generate web interfaces, their outputs are often evaluated for whether they function correctly, but not whether they provide a high-quality user experience. An interface may compile successfully while still containing accessibility barriers, deceptive patterns, poor visual hierarchy, or excessive cognitive complexity.
This research investigates whether a lightweight vision-language model can be trained to act as a UI critic, automatically detecting interface-quality violations in generated web pages.
The study brings together 19 UI-quality principles across WCAG 2.2 accessibility standards, deceptive design taxonomies, and established HCI theories of perception, cognition, and interaction.
A verified dataset of approximately 10,000 generated web pages was created by introducing controlled UI violations into clean interfaces. This dataset was then used to train a 4B vision-language model to recognize the underlying design principles.
Continued reinforcement learning improved the model’s micro-F1 score from 36% to 84%, with 13 of the 19 UI principles achieving more than 80% F1.
The resulting model can help audit AI-generated interfaces, filter low-quality UI training data, and provide design-aware feedback to coding agents.