RubSE: Rubric-Guided Self-Evolution for UI-to-Code Generation

RubSE converts UI-to-code visual feedback into typed rubrics with history, selecting one repair target per round, and is reported to outperform naive self-evolution across six VLMs and three benchmarks by reducing trajectory collapse.

August 27, 20261 source55% confidence

66Strategic

Strategic score

Evidence strength35

Technical novelty78

Commercial potential55

RubSE reframes UI-to-code test-time self-evolution as rubric-scoped visual repair rather than unconstrained local code editing. The paper identifies visual repair coupling as the key failure mode, in which a local edit propagates through layout, style, and component dependencies and suppresses the very regions the previous revision had made faithful. The proposed solution operationalizes rubrics as typed visual-repair context, selecting one prioritized repair target and retaining prior rubric selections as history to bound each revision.¹

Context

In the UI-to-code setting, naive self-evolution repeatedly rewrites generated code from visual comparisons; RubSE injects a structured intermediate representation (rubrics) and a history of previously used rubrics. This moves the decisive constraint from the model's implicit ability to localize a repair in raw image-to-code feedback to an explicit, typed scope for each edit. Experiments reported in the abstract span six VLMs and three UI-to-code benchmarks, with gains in final-round and best-round performance.¹

Assessment

The abstract establishes a clear mechanism and a plausible control strategy, but the evidentiary basis is thin: no numeric results, no dataset names, no baseline models, and no ablation of the history component are given. The claim of transfer—stronger rubric generators improve weaker code improvers—is particularly underspecified; the abstract reports it as 'further analysis' without effect sizes.¹

Caveats

Because the source is a single preprint abstract, the reader cannot assess trajectory-level metrics, the stability improvements, or how rubrics are produced and evaluated. A change of mind would require inspecting the full paper's tables or code release; until then, RubSE's advantages over naive self-evolution are treated as reported but not independently verifiable.¹

Source text

Hugging Face Daily Papers — Rubrics as Visual-Repair Context for Self-Evolving UI-to-Code Generation · August 25, 2026

Computer Science > Computer Vision and Pattern Recognition

Title:Rubrics as Visual-Repair Context for Self-Evolving UI-to-Code Generation

Abstract: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 naïve 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.

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Visual repair coupling is the paper's central term: it is an unstated assumption that UI-to-code edits are mediated by visual inspection, and that code dependencies make corrections non-local. The definition in the abstract is qualitative; no causal model or quantified propagation is provided.

'Trajectory collapse' names the phenomenon where a self-evolution loop degrades across rounds. The abstract claims RubSE mitigates collapse by improving recovery from severe visual regressions, but no metric is supplied; the term itself mixes empirical observation with a collapse metaphor.

The transfer statement—'stronger rubric generators can transfer effective visual-repair guidance to weaker code improvers'—is a causal claim that the abstract-level evidence does not carry. It may be an experimental finding with unseen numbers, but as provided it conflates generator quality with guidance effectiveness.

The source text above remains the property of its publisher and is reproduced for commentary.

↑ Back to top