Search papers, labs, and topics across Lattice.
This paper addresses the instability in test-time self-evolution of UI-to-code generation by identifying the issue of visual repair coupling, where local code edits can negatively impact other visual elements. The authors introduce RubSE, a Rubric-guided Self-Evolution framework that utilizes structured visual feedback to guide code revisions, improving the stability and effectiveness of the generation process. Evaluations reveal that RubSE significantly enhances performance across multiple vision-language models and benchmarks, leading to more consistent refinement outcomes and better recovery from visual regressions.
Visual repair coupling can derail UI-to-code generation, but RubSE offers a structured solution that stabilizes self-evolution and boosts performance across models.
Large vision-language models have shown strong progress in UI-to-code generation, yet their test-time self-evolution remains unstable. We first identify a fundamental obstacle, termed visual repair coupling: a local code edit may propagate through layout, style, and component dependencies, correcting one visual mismatch while degrading regions that were previously faithful. To address this issue, we present RubSE, a Rubric-guided Self-Evolution framework that uses rubrics to represent visual feedback as a structured visual-repair context. At each refinement round, RubSE generates typed candidate rubrics, selects one prioritized repair target, and stores previously selected rubrics as history, thereby steering each revision toward a well-scoped visual repair while discouraging repeated or over-broad changes. Evaluations across six VLMs and three UI-to-code benchmarks demonstrate that RubSE substantially outperforms na\"ive self-evolution in final-round and best-round settings, achieving more stable refinement trajectories and a higher trajectory-level performance ceiling. Further analysis shows that RubSE mitigates trajectory collapse by improving recovery from severe visual regressions, and that stronger rubric generators can transfer effective visual-repair guidance to weaker code improvers.
