Improving AI Web Page Generation with Rubrics

Summary & Methodology Analysis

The RubSE method addresses the challenge of visual repair coupling, where local code modifications often trigger unintended regressions in complex web layouts. The architecture operates through a cycle of rubric-based evaluation and targeted revision. It generates a candidate set of failure descriptors across five specific aspects: layout geometry, spacing density, typography text, styling visual, and completeness. By selecting a prioritized rubric and maintaining a per-instance history of these choices, the agent avoids redundant or regressive code edits during the iterative generation process. This history effectively serves as an avoid list for the model as it traverses the solution space.

Performance was evaluated across a diverse set of models, including proprietary frontier VLMs like GPT-5.4, GPT-5.2, and Claude-Sonnet-4.5, as well as open-source variants such as Qwen3-VL-32B-Instruct, Qwen3.5-9B, and Qwen-3.6-35B-A3B. The system demonstrates a higher trajectory-level ceiling for code quality compared to naive approaches. The evaluation relies on benchmarks like Design2Code, which provides model-based metrics, and UI2Code-Real, which contains 115 real-world web pages designed to challenge existing models with realistic UI complexity.

Despite these gains, the research has notable limitations. The paper does not explicitly guarantee code correctness, and the refinement process occasionally introduces syntax or runtime errors that lead to rendering failures. Additionally, evaluations performed by VLMs can remain unstable, potentially diverging from human judgment in later refinement rounds when assessing subtle visual differences. The current research scope is limited to HTML and CSS, leaving the performance of RubSE on other UI languages and a broader range of model families underexplored.