About
Research paper on training LLM-based UI generators with feedback that matches professional design practice. The authors compare rankings, comments, sketches, and direct revisions from 21 designers, then use the resulting data to improve code models that generate rendered interfaces.
Summary
Why This Paper Adds New Signal:
1. Designer-native feedback: Tests commenting, sketching, and direct revision instead of relying only on thumbs-up or pairwise rankings
2. Substantive dataset: Collects roughly 1,500 annotations from 21 designers with 2 to 30-plus years of experience
3. Grounded interaction matters: Finds sketch and revision feedback produces stronger training signals than comments or rankings
4. Model-level evidence: Uses the feedback to train reward and UI code-generation models, then evaluates rendered results with human judges
5. Practical GenUI lesson: Shows that improving generated interfaces depends on feedback tools that fit how designers actually critique visual work
1. Designer-native feedback: Tests commenting, sketching, and direct revision instead of relying only on thumbs-up or pairwise rankings
2. Substantive dataset: Collects roughly 1,500 annotations from 21 designers with 2 to 30-plus years of experience
3. Grounded interaction matters: Finds sketch and revision feedback produces stronger training signals than comments or rankings
4. Model-level evidence: Uses the feedback to train reward and UI code-generation models, then evaluates rendered results with human judges
5. Practical GenUI lesson: Shows that improving generated interfaces depends on feedback tools that fit how designers actually critique visual work
Tags
ui-generationdesigner-feedbackmodel-alignmentuser-study