Learning from Designer Feedback: Which Human Feedback Modality Improves LLM-Based UI Generation Most Efficiently?

Human feedback for user-interface generation varies sharply in granularity: a designer can choose between two screens, write a critique, mark a region with a grounded comment, or directly modify the design. This study compares the efficiency of these four modalities as learning signals for the alignment layer of an LLM-based UI generation pipeline. The four ML-RLDF tables were harmonized into 1,499 preferred--nonpreferred pairs contributed by 21 professional designers and covering 1,792 unique screenshots. A prompt-disjoint five-fold protocol trained a symmetric pairwise reward model on 50 deterministic UI statistics and produced cross-fitted preference probabilities and a fold-calibrated UI Quality Score. Quality contrast, sample efficiency, cross-modal transfer, feature ablation, designer-level robustness, and edit-magnitude sensitivity were evaluated under the same grouped validation design. The pooled model selected the designer-preferred screen in 53.8% of held-out pairs (95% cluster-bootstrap confidence interval: 51.1%--56.5%). Comment pairs were the most learnable modality, reaching 63.2% held-out accuracy and a mean quality gain of 2.27 points. At equal budgets of 8--48 pairs, comment feedback achieved the highest sample-efficiency area under the learning curve, 0.530, compared with 0.508 for sketch, 0.501 for ranking, and 0.500 for revision; its advantage over each alternative remained significant after Holm correction. Revision produced the largest raw mean quality gain, 3.42 points, but its confidence interval crossed zero and the estimate was sensitive to a small number of large edits. Ranking provided scale but contained contradictory repeated judgments and cyclic preferences. The findings identify comment feedback as the most sample-efficient signal in this leakage-controlled reward-learning setting, while grounded sketch feedback provides the strongest off-diagonal transfer and a more stable edit-based signal than direct revision.