Rubrics as Visual-Repair Context for Self-Evolving UI-to-Code Generation
RubSE tries to stop UI-to-code self-repair from fixing one visual mismatch while breaking another.
The paper names that failure mode “visual repair coupling,” where a local code edit can ripple through layout, style, and component dependencies. RubSE turns visual feedback into typed rubrics, picks one repair target per refinement round, and keeps prior rubrics as history. Across six vision-language models and three UI-to-code benchmarks, the authors report stronger final- and best-round results than naive self-evolution. They also say the method helps recover from severe regressions and lets stronger rubric generators guide weaker code improvers. HF Daily Papers' note
The paper names that failure mode “visual repair coupling,” where a local code edit can ripple through layout, style, and component dependencies. RubSE turns visual feedback into typed rubrics, picks one repair target per refinement round, and keeps prior rubrics as history. Across six vision-language models and three UI-to-code benchmarks, the authors report stronger final- and best-round results than naive self-evolution. They also say the method helps recover from severe regressions and lets stronger rubric generators guide weaker code improvers. HF Daily Papers' note
score 4