A11yn: Aligning LLMs for Accessible Web UI Generation

ACL ARR 2026 January Submission1391 Authors

Keywords: AI for Accessibility, Generative UI, Post-Training Alignment

Abstract: Large language models can generate functional and visually appealing web UIs from natural language instructions. However, these models frequently produce interfaces that fail to meet accessibility requirements, excluding users with diverse needs and contexts. We address this by introducing A11yn, a policy alignment framework for accessibility-aware UI code generation. A11yn utilizes a severity-aware WCAG reward as a training signal, prioritizing the correction of high-impact accessibility failures. To support model training and evaluation, we contribute two resources: UIReq-6.8K, a dataset of 6,800 diverse UI generation instructions, and RealUIReq-300, a benchmark of 300 real-world tasks grounded in public web pages. Experimental results show that A11yn significantly improves accessibility compliance over strong baselines. Critically, these gains are achieved while maintaining the semantic fidelity and visual quality of the generated UIs, demonstrating that code-generating LLMs can be successfully aligned with objective accessibility requirements.

Paper Type: Long

Research Area: Human-AI Interaction/Cooperation and Human-Centric NLP

Research Area Keywords: human-centered evaluation, user-centered design, value-centered design, code generation and understanding

Contribution Types: NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 1391