Researchui generationdesigner feedbackqwenreward model
Apple Trains UI Models With Designer Feedback
9.3
Apple researchers publish a new paper showing designer-native feedback improves automated UI generation. They collected 1,460 annotations from 21 professional designers and used comments, sketches, and edits to train a reward model and fine-tune Qwen-based generators; 181 sketch annotations enabled a Qwen3-Coder variant to outperform GPT-5 in UI tasks. Authors note subjectivity causes high variance, with sketches and edits yielding higher agreement.
Key Points
- 1Collected 1,460 designer annotations and 181 sketch edits to create paired preference training examples
- 2Trained a reward model on rendered screenshots plus natural-language prompts to score visual design quality
- 3Showed smaller Qwen variants fine-tuned with designer feedback can outperform larger proprietary LLMs in UI generation
Scoring Rationale
Strong novelty and practical methods from credible Apple researchers, limited by narrow focus on UI generation
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
