New Framework Trains AI Models to Generate Web Interfaces That Follow Accessibility Standards

Large language models have become capable of producing functional web user interfaces from simple text descriptions—but they often overlook critical accessibility requirements that allow people with disabilities to navigate and use digital products effectively. A new framework aims to close that gap.

The Accessibility Problem#

Despite their sophistication, current LLMs frequently generate web UIs that violate Web Content Accessibility Guidelines (WCAG), the internationally recognized standard for digital accessibility. This creates barriers for users with diverse abilities and contexts, limiting the reach and inclusivity of generated interfaces. Traditional approaches to fixing accessibility issues rely on manual review and correction after generation—an inefficient and resource-intensive process.

Introducing A11yn#

Researchers have developed A11yn, a post-training framework that embeds accessibility awareness directly into how language models generate web UIs. Rather than depending on labor-intensive human feedback, the system transforms automated WCAG audits into a measurable accessibility reward signal. This allows models to learn accessibility principles during training without needing extensive annotated examples.

Datasets and Evaluation#

To enable this research, the team released two resources: UIReq-6.8K, a collection of 6,800 diverse UI generation instructions, and RealUIReq-300, a benchmark of 300 real-world web UI generation tasks. These datasets support both training and rigorous evaluation of accessibility-aligned models.

Results and Performance#

Testing shows that A11yn substantially outperforms baseline approaches. Compared to unaligned models, the accessibility-trained version reduces accessibility violations by 87.5% under one independent auditor and 58.1% under another. Critically, A11yn achieves these improvements in a single-pass generation—meaning users get accessible interfaces immediately, without requiring a post-processing step. The model maintains comparable visual quality and semantic accuracy to standard approaches.

Code and datasets have been made available for further research.