Efficient Personalization of Generative User Interfaces

1. Problem Statement

Generative User Interfaces (GenUIs) offer the ability to synthesize interfaces on demand, creating new opportunities for personalization. However, personalizing these systems is challenging because:

• Subjectivity: Desirable UI properties (e.g., hierarchy, cleanliness, density) are subjective and difficult for users to articulate explicitly.
• Data Scarcity: New users provide only sparse feedback, yet the system must generalize to newly generated interfaces rather than selecting from a fixed catalog.
• Preference Divergence: Even trained designers disagree significantly on what constitutes a "good" design. A single aggregate model or a fixed rubric of design concepts fails to capture individual nuances.
• Limitations of Current Approaches: Existing methods often rely on post-hoc customization (settings) or direct prompting, which are insufficient for ephemeral, single-interaction interfaces. Furthermore, large multimodal models (LMMs) often struggle to adapt efficiently to specific user preferences with limited data.

2. Methodology

The authors propose a sample-efficient personalization framework that represents a new user not as a fixed set of rules, but as a mixture of prior designers. The approach consists of three main components:

A. Dataset Construction (The "Preference Dataset")

• Scale: 20 professional UI designers provided pairwise judgments on 600 generated mobile UIs, resulting in 12,000 preference labels.
• Generation: UIs were generated using high-performing models (GPT-5 and Gemini-2.5-Pro) based on 100 human-written screen descriptions expanded into 4 variants each.
• Annotation: Designers rated pairs on a 4-point scale (from "A much better than B" to "B much better than A") without a neutral option.
• Key Finding: Significant disagreement exists among designers ($\kappa = 0.25$). Even when designers use similar terms (e.g., "hierarchy"), they interpret and prioritize them differently.

B. The Personalization Algorithm

The method uses a retrieval-based approach combined with active learning to adapt to a new user with minimal data.

1. Data Bank Construction: A bank of labeled pairs is created where each example includes the screen pair, the designer ID, and the preference label.
2. Adaptive Onboarding (Query Selection):
   • When a new user arrives, the system selects a small number of pairwise comparisons ($k$) to ask them.
   • It uses Expected Information Gain (EIG) to select queries that maximize the reduction of uncertainty about which subset of the training designers the new user resembles.
   • After each response, the system updates a probability distribution ($\pi_t$) over the training designers using Bayes' rule.
3. Model Personalization (Reranking):
   • For a new generation task, the system retrieves the top $n$ similar pairs from the training bank using a visual encoder (UIClip).
   • Instead of using a generic model, it aggregates the labels of these retrieved pairs, weighted by the learned user distribution ($\pi_u$).
   • The final score for a candidate UI is a weighted sum of the labels from similar historical examples, where the weights reflect how much the new user aligns with the specific designers who labeled those examples.

C. Online Evaluation Pipeline

The system was tested in an end-to-end generation pipeline:

1. Onboarding: Users provide a text profile and answer 8 adaptive pairwise questions.
2. Generation: A base LLM generates a pool of UI candidates based on a prompt.
3. Selection: The personalized preference model reranks the pool to select the best candidate for that specific user.

3. Key Contributions

1. New Dataset: A dataset of 12,000 pairwise preference labels from 20 designers over 600 GenUIs, enabling the first direct analysis of preference divergence in generative UIs.
2. Analysis of Divergence: Qualitative and quantitative evidence showing that designers disagree substantially ($\kappa=0.25$) and interpret shared design concepts (e.g., "modern," "clean") differently.
3. Sample-Efficient Method: A novel personalization technique that models a new user as a mixture of prior designers. It outperforms both a pre-trained UI evaluator (UIClip) and a much larger multimodal model (GPT-5.2) in predicting preferences with limited data.
4. End-to-End Validation: An online study demonstrating that this method produces interfaces preferred by new designers over baselines, including direct user prompting and profile-conditioned generation.

4. Results

Offline Technical Evaluation

• Baseline Comparison: The proposed retrieval-based model outperformed the pre-trained UIClip (0.610 vs. 0.565 accuracy) and the zero-shot GPT-5.2 (0.591).
• Impact of Onboarding:
  • The proposed method improved with more onboarding examples, reaching 0.620 accuracy at $k=8$.
  • GPT-5.2 degraded in performance as more onboarding examples were added, likely due to "context rot" where the model struggled to integrate the few-shot examples effectively.
  • UIClip could not be effectively fine-tuned with only 8 samples.

Online Study (12 New Designers)

• Setup: Participants compared UIs generated under four conditions: Zero-shot, Profile-conditioned generation, Profile-conditioned + Judge, and Profile-conditioned + Adaptive Model (Ours).
• Performance:
  • The Adaptive Model achieved the highest win rate (60.35%) and the highest Elo rating (1054.8).
  • It significantly outperformed the Profile-conditioned Judge (50.90%) and Zero-shot generation (41.25%).
  • Crucially, even when the text profile was identical, the adaptive model (using 8 pairwise labels) selected different, preferred UIs in 96.8% of bootstrap resamples compared to the text-only baseline.

5. Significance and Implications

• Beyond Fixed Rubrics: The paper demonstrates that personalization in GenUIs cannot rely on a single "best" design standard. Instead, systems must learn a user-specific prior based on how individuals weigh conflicting design dimensions (e.g., density vs. whitespace).
• Efficiency over Scale: A lightweight, retrieval-based model adapted via active learning outperforms massive multimodal models when data is scarce. This suggests that behavioral signals (pairwise choices) are more effective for personalization than explicit text descriptions.
• Practical Application: The framework is applicable to real-world scenarios like personalized widget editors, auto-configuration of reading apps, and design suggestion tools, where user feedback is sparse but critical.
• Future Direction: The work highlights that "noise" in design preferences is actually signal representing diverse human taste. Future systems should move away from averaging preferences toward modeling the full distribution of user preferences.

In conclusion, the paper establishes that lightweight preference elicitation (approx. 8 pairwise comparisons) combined with a mixture-of-experts modeling approach is a practical and highly effective foundation for building personalized Generative UI systems.