Efficient Personalization of Generative User Interfaces: A Preference Learning Approach Based on Designers' Priors

Generative User Interfaces (Generative UI) enable instant interface prototype generation via AI, but personalization faces challenges such as subjectivity, difficulty in expression, and sparse feedback. This study finds significant disagreements among designers in judging UI quality, and proposes a preference learning approach based on designers' priors—representing new users as a combination of existing designers' styles—enabling efficient personalized generation with minimal feedback.

Generative UI can theoretically customize interfaces, but there are three obstacles:

1. Subjectivity: Good interfaces vary from person to person (concise vs. information-rich, efficiency vs. aesthetics);
2. Expression Difficulty: Users struggle to precisely describe their needs (e.g., "more professional" lacks design semantics);
3. Sparse Feedback: Users are reluctant to provide large amounts of labeled data. The root cause is that UI quality is high-dimensional, subjective, and context-dependent, making it difficult to characterize with fixed rules.

The research team constructed a dataset: 20 professional designers performed pairwise comparisons of 600 generated interfaces. The results show:

• Disagreement Quantification: The Cohen's Kappa coefficient is only 0.25 (mild agreement), indicating significant differences in judgments among designers;
• Qualitative Analysis: Even when referencing similar design concepts (e.g., "clear hierarchy"), different designers still have varying definitions, priorities, and application methods (such as size contrast, information grouping, and visual guidance paths).

New Idea: Represent new users as weighted combinations of existing designers' styles. Core Insight: Designers' styles are unique and consistent; user preferences can be viewed as combinations, and minimal feedback can infer the combination weights. Technical Implementation:

1. Designer Embedding Space: Low-dimensional vectors represent designers, where distance reflects preference similarity;
2. User Preference Inference: Determine the user's position in the embedding space through minimal feedback (e.g., pairwise comparisons);
3. Personalized Generation: Adjust the generation strategy based on the user's position, prioritizing interfaces that align with the styles of nearby designers.

Technical Evaluation: Compared with pre-trained UI evaluators and multimodal models, this method performs better in preference prediction tasks, and its performance improves significantly as feedback increases; User Study: Evaluations by 12 designers indicate that interfaces generated by this method are more favored, with statistical significance.

1. Value of Design Diversity: Individualized preferences demonstrate that one-size-fits-all design cannot meet needs;
2. Potential of Lightweight Feedback: Minimal high-quality feedback can achieve effective personalization, optimizing tool experiences;
3. Evolution of Designer Role: From rule-setters to style providers; the core value lies in creating unique design perspectives.

Limitations: The designer sample (20 individuals) may not cover the full spectrum of preferences; interface types are limited to specific generative UI; there is a lack of long-term preference adaptation mechanisms. Future Directions: Integrate implicit feedback (eye movements, operation patterns); explore multi-user collaboration scenarios; develop fine-grained design attribute control mechanisms.