acceptodds
Under review as a conference paper at ICLR 2027

WUIDGET: Web UI Generation with Small Models via Gap-Driven Post-Training

Abstract

Web UI generation requires a model to produce executable code, coherent visual design, and working interactions in the same artifact. Small models fail differently across these axes, while aggregate benchmark scores provide little guidance about which training signal should address which failure. We propose *WUIDGET*, a gap-driven post-training framework that distills evaluation failures into a structured root-cause taxonomy and uses it to guide data and reward design across three stages. Supervised fine-tuning learns broad structural and aesthetic patterns; Rejection Fine-Tuning with Distillation (RFT-D) minimally repairs the model's own rollouts; and reinforcement learning optimizes verifier-grounded visual and interactive rewards. On OpenDesign (840 prompts), the final Qwen3-8B checkpoint improves over its base from to static ( points) and from to agentic (a % relative gain in working interactions per artifact), and improves over the comparably sized AesCoder-4B by static. The static improvement over AesCoder remains consistent on WebDev Arena () and Design Arena (). Across Qwen3-8B and SmolLM-3B, SFT supplies the largest visual-quality gain, RL the largest interactive gain, and inserting RFT-D changes the static–agentic trade-off reached by downstream RL. Matched gap/no-gap controls show that SFT guidance improves the macro-average agentic score by (95% bootstrap interval ), while gap-selected RL queries produce a smaller, primarily static shift. These results support gap analysis not only as diagnosis, but as an organizing signal for specializing small models on multi-axis generation tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
post-training, small language models, web UI generation, failure analysis, reinforcement learning, rejection fine-tuning, supervised fine-tuning, knowledge distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Q9dNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes WUIDGET, a gap-driven post-training framework for text-to-web-UI generation. It distills evaluation failures into a structured root-cause taxonomy to guide data and reward design across three stages: SFT, RFT-D, and RL. Experiments on Qwen3-8B and SmolLM-3B across different benchmarks show improvements, supporting the effectiveness of these stages. The method also shows improvements over AesCoder-4B.

Critical strengths
  1. Strong overall performance across three benchmarks and two model families, Qwen3-8B and SmolLM-3B. Qwen3-8B also improves over the comparably sized AesCoder-4B by 4.18 points in static quality.
  2. The stage contributions are carefully analyzed. Adding RFT-D changes the outcome of subsequent RL.
Critical weaknesses
  1. Support for the central gap-driven claim is not strong enough. Although the paper emphasizes gap-driven training, the contribution of gap guidance is not fully isolated and validated. The improvements may depend more on the overall training recipe.
Questions for authors

How do the existing experiments independently establish the contribution of gap guidance?

Reviewer_rkCvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

WUIDGET uses recurring failure analysis to guide the specialization of compact models for web UI generation. The pipeline combines supervised fine-tuning, teacher repair of student outputs through RFT-D, and reinforcement learning with visual and interaction rewards. Evaluations across model families and benchmarks show improvements in generation quality and different roles for the training stages. The branch comparisons further show that RFT-D changes the balance between visual quality and interactive behavior reached by subsequent RL.

Critical strengths
  1. The method has a clear rationale grounded in the task. It separates recurring failures across outputs from defects in individual implementations, then uses demonstrations, code repair, and rendered feedback accordingly. This provides a practical way to organize supervision for UI generation.

  2. The training branches provide an informative empirical finding. Including RFT-D changes the balance reached by subsequent RL, with different directions across model families. The paper reports this dependence clearly and avoids claiming that the intermediate stage improves every objective.

Critical weaknesses
  1. The novelty is limited. As discussed in Section 2, AesCoder already combines UI-specific SFT with visual and interactive rewards, while PERsD and LEMA use teacher repair. The contribution lies mainly in organizing and adapting these ingredients.

  2. The gap-driven decision process is underspecified. The main text does not concretely explain how failure prevalence determines data allocations or reward changes, making the claimed repeatability difficult to assess.

  3. The controls do not cleanly isolate the taxonomy's contribution. SFT changes both targeting and the synthetic/public data mixture, while both RL arms retain gap-informed rewards. The evidence for the central organizing principle remains incomplete.

  4. RFT-D's specific design is insufficiently justified. The branch comparison does not distinguish minimal repair and triage from additional teacher supervision, and repair locality is not measured. Changing the downstream endpoint alone does not establish the value of these constraints.

  5. The stability of the incremental gains is unclear. The gap controls use a single training run per arm, so their reported intervals do not account for training variability. This limits confidence in the reproducibility of the guidance effects.

Questions for authors
  1. How are failure frequencies converted into data allocations and reward changes, and which decisions require manual judgment?

  2. Does failure-guided SFT still help when the teacher, synthetic/public data mixture, and training budget are held fixed?

  3. What evidence isolates the contribution of gap-informed reward design, which remains present in both RL control arms?

  4. What supports minimal repair and triage over ordinary teacher correction under comparable budgets, and how is repair locality verified?

  5. How stable are the gap-guidance effects and the RFT-D tradeoffs across independent training runs?

Reviewer_dzZDOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes WUIDGET, a gap-driven post-training framework for improving Web UI generation with small language models. The framework consists of three stages: gap-guided supervised fine-tuning (SFT), rejection fine-tuning with distillation (RFT-D), and verifier-grounded reinforcement learning (RL). Experiments on three benchmarks demonstrate improvements in visual quality and interactivity.

Critical strengths
  1. The paper organizes SFT, RFT-D, and RL into a unified pipeline, using failure analysis to guide data construction and reward design.
  2. The paper evaluates the proposed framework on two model families and three benchmarks, reporting improvements over the base models and providing several ablation studies to examine the contributions of different training stages.
Critical weaknesses
  1. The paper adopts a two-stage visual reward strategy, transitioning from pointwise to pairwise evaluation, but lacks a direct ablation study to justify this curriculum. Although the existing experiments show that pairwise rewards outperform pointwise rewards, they do not demonstrate why starting with pointwise rewards and then switching to pairwise rewards is preferable to using pairwise rewards from the beginning.
  2. The Two-Turn Static Judge infers whether interactive features work based solely on screenshots and source code, without directly verifying runtime behavior. Although the paper includes a comparison with a live browser agent, it lacks an analysis of the agreement between static judgments and actual functional test results. Consequently, it remains unclear whether the RL reward accurately reflects interaction correctness.
  3. While RFT-D emphasizes localized defect repair, the RL stage assigns the same artifact-level advantage to all tokens in a generated sequence. Given that the interactivity judge can already identify component-level functional issues, the paper does not explore how such fine-grained feedback could be leveraged to improve credit assignment during RL training.
  4. The paper primarily compares its method against AesCoder-4B, which has a different parameter scale, and Claude Sonnet 4.5, which is evaluated under different reasoning settings. Stronger baselines with matched model sizes and training budgets are needed. Furthermore, the existing interactivity metrics mainly count successful interactions, providing limited evidence of correctness in complex interaction logic and end-to-end functional workflows.
Questions for authors

See the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.