acceptodds
Under review as a conference paper at ICLR 2027

Venus: Compiling an Aesthetic Standard into Supervision for Web Coding

Abstract

Web coding models now meet basic functional requirements, yet still struggle to turn a user's subjective aesthetic wishes, a premium look or visual impact, into concrete design. Improving design ability with high-quality web data is bottlenecked by obtaining professional aesthetic supervision at scale: the criteria for design quality are usually implicit in a designer's experience and resist conversion into executable rules, while page-by-page human review is reliable but too costly for the data scale training demands. We present Venus, a pipeline for constructing web-coding training data driven by professional design knowledge. Rather than convert aesthetic intuition fully into automatic rules, Venus decomposes it into a multi-dimensional, graded, verifiable design standard and uses that standard to drive a two-stage process of data expansion and supervision construction, so that limited human effort supports building professional aesthetic data at scale. First, high-quality seed pages are expanded along each dimension and a small amount of human review keeps the good ones. Then, guided by the same standard, these pages are degraded in a controlled way, each degradation along one aesthetic dimension, which yields large-scale aesthetic comparison data without any new pairwise human annotation. From this data we distill a local aesthetic evaluation model whose automated review builds a larger high-quality corpus, on which we train a web coding model. On WebDevJudge, a public benchmark of real human preferences over web pages, the Venus evaluation model attains the highest order consistency in the panel and a strict accuracy within the band of the frontier closed-source judges we compare against, and in a blind study it agrees with independent designers at a rate approaching their agreement with one another. An open Qwen-family base trained on the Venus corpus produces web pages of substantially higher design quality than the same base and competitive with frontier systems on public benchmarks. We will open-source the ten-dimension web design standard and the trained models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
web coding, llm

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3QExOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The proposed VENUS improves generated websites by teaching models what good design looks like. It defines a ten-dimension design standard, starts with designer-approved pages, and deliberately introduces visual defects to create comparison pairs. These pairs train a small visual judge, which filters web-code training data for a generation model. The resulting VENUS-CODER improves its base model’s ArtifactsBench score from 53.50 to 64.52; designers prefer its pages in 77.6% of majority-decided comparisons.

Critical strengths

This work has low labeling effort. Designers review only a small set of seed pages. Making each one deliberately worse in a single design dimension produces "good vs. worse" pairs automatically. The pair also records which dimension is worse and how severe the defect is.

Across all three tables in experimental section, the results are consistent: the 7B judge matches frontier judges with the highest order consistency (Table 1), the screened corpus lifts the same base by +11.02 on ArtifactsBench to rank first among frontier systems (Table 2), and the size-matched ablation shows the gain comes from screening rather than data volume (Table 3).

Critical weaknesses

This paper used senior designers in multiple places, i.e., judge study, generator study, design standard, and seed page review. however, it doesn't specify whether they are the same designers. This matters for independence. If the evaluators also helped write the standard or select the seed pages, their agreement with the judge could be inflated.

On page 18, Fig A.2, the very different pages get the same score. Task 3 has a full sidebar and header but an empty chart. Task 9 has a laid-out page but a blank board. Task 19 is only a spinner. All three get 2/5, so the scorer doesn't separate "mostly built" from "nothing rendered." That weakens trust in the appearance metric. Is there any other examples for 0,1,3,4 scores? Please also include randomly sampled or median-gap examples in which both pages render fully.

The design standard VENUS-CANON is used at every step, i.e., to make the training data, to train the judge, and to build the pairs that designers check. So the results may only show that the system follows its own rules, not that it captures good design.

VENUS-CANON rewards tidy, conventional, mostly Western-style design, so it may push outputs toward safe, look-alike pages. It may also penalize bold or culturally different styles, such as the dense layouts common on East Asian sites. The paper doesn't measure output diversity or report the designers' backgrounds.

Questions for authors

Please clarify whether the senior designers in the two blind studies are the same individuals, and whether any of them participated in writing VENUS-CANON or screening seeds; overlap would weaken the independence of the human validation.

For Table 3, is it possible to compare with a VLM filtered subset? a stronger alternative filter - not necessarily frontier VLM as the cost is high.

It's supposed to only affect appearance, why table 2 shows it gains a lot in func. as compared to the base mode? it's mentioned in limitation as a by-product but needs more explanation.

4 comments
AuthorsOpenReview ↗

Reply to Reviewer 3QEx(1/2)

Thank you for your careful review and for recognizing the low annotation cost and the value of the corpus-size-matched ablation. We clarify the evaluators' independence, correct the screenshot presentation, and address the concerns about design standards and functional gains below.

(W1/Q1) Standard development, seed screening, and final human evaluation were conducted by distinct teams. One group of designers developed VENUS-CANON. A separate team of 1 professional designer and 6 professional evaluators familiar with VENUS-CANON screened the seeds. The two human studies—agreement with VENUS-JUDGE and comparison of VENUS-CODER's generated pages—were conducted by the same independent group of designers. None of that group participated in developing the standard or screening the seeds. We clarify these responsibilities and relationships here.

(W2) We identified an error in the displayed screenshot for Task 19 and apologize for the confusion. It has been corrected in Figure A2 of Appendix C (Page Comparisons). We also provide examples across score levels and visual styles to make the evaluator's distinctions easier to inspect.

WebGen-Bench appearance ratings use a 1–5 scale. Figure A15 in Appendix K (Additional Page Comparisons) includes:

TaskBase / VENUS appearance scores
2 — Neighborhood mapping3 / 5
4 — Basketball analytics2 / 4
11 — Strategy game3 / 2

These examples span scores from 2 to 5 and include a case where VENUS-CODER scores below the base model. Additional comparisons with equal scores and different styles appear in Figures A6–A9 of Appendix H.

(W3) We understand the concern that using the same standard to construct supervision, train the judge, and build internal evaluation pairs might demonstrate conformity to that standard rather than broader design quality. Three distinctions are relevant.

First, a shared design standard does not imply a single visual style. Style determines the visual language, such as editorial, minimalist, or neon treatments. Design criteria assess the execution within that style: hierarchy, spacing, alignment, readability, and component coherence. VENUS-CANON focuses on these criteria without requiring one fixed style. The seeds cover multiple styles so that the criteria can be exercised across different visual languages.

Second, VENUS-CANON is grounded in professional design practice. Its ten dimensions and six grade anchors were developed through repeated reviews of generated pages with professional interface designers. We organized recurring judgements into properties that can be checked from a rendered page and interpreted consistently by another reviewer (Section 3.1).

Third, external benchmarks and designer assessments examine both judge performance and generated-page quality. VENUS-JUDGE is evaluated on WebDevJudge, an independent benchmark whose preferences come from human comparisons of independently sourced pages. Those pages were not generated under VENUS-CANON, and the pairs were not constructed using our degradation operators. VENUS-JUDGE achieves Strict and Consistency scores of 55.6 and 93.2, compared with 40.8 and 71.2 for the same untrained backbone (Table 1). This supports improved agreement with external human preferences and greater stability under presentation-order swaps.

For VENUS-CODER, we evaluate generated pages on ArtifactsBench and WebGen-Bench and ask designers to compare the outputs directly. The human evaluation setup and results are provided in Appendix E (Alignment with Human Judgement). Together, these evaluations provide evidence at both the preference-judgement and page-generation stages that the design supervision can improve page quality.

We also recognize the scope of this evidence. Clear hierarchy and coherent components are widely useful design properties, but they are not the only principles under which a design can succeed. VENUS-CANON is an executable standard for screening web-design quality, rather than a universal definition of good design.

The origin of the standard is described in Section 3.1, and the distinction between WebDevJudge and the internally constructed evaluation is described in Sections 4.1–4.2. We further clarify the above limitation in this response.

AuthorsOpenReview ↗

Reply to Reviewer 3QEx(2/2)

(W4) Our presentation did not make the distinction between style choices and design criteria sufficiently clear. Tidy, conventional Western-style layouts and information-dense East Asian layouts represent different visual approaches. VENUS-CANON evaluates hierarchy, alignment, spacing, colour roles, component coherence, and readability within the chosen approach (Appendix A).

A dense page can receive a high score when its information is legible and well organized; a sparse minimalist page can still have unclear hierarchy or inconsistent components. Figure A9 in Appendix H shows an information-dense dashboard with navigation, summary cards, and tables, receiving 4 / 4 for the base and VENUS-CODER pages. This illustrates that density alone does not imply a low score.

The additional examples in Figures A6–A9 of Appendix H cover editorial layouts, portfolios, utility interfaces, and information dashboards. Figures A12–A14 of Appendix J further illustrate photographic layouts, expressive typography, geometric imagery, and product interfaces. These cases show how the criteria can be applied across different styles and information densities.

(Q2) We conducted the suggested comparison with a general-purpose VLM filter. It examines whether a VLM without VENUS-CANON training can score and rank candidates, select a corpus of the same size, and yield comparable generation quality.

We use Gemini 3.5 Flash to score each page in the same candidate pool, rank the pages, and retain the top subset at the same ratio as the VENUS-JUDGE corpus. We then fine-tune the same base model with the same training settings and evaluate the generated pages under the same protocol:

Selection methodArtifactsBench / 100WebGen functional accuracy, original protocolWebGen appearance / 5
Gemini 3.5 Flash score-and-rank selection59.4841.4%3.35

Under the same candidate-source and corpus-size conditions, this general-purpose VLM filter does not match VENUS-JUDGE's selection results. The comparison supports the value of the selection procedure built from VENUS-CANON. This is an additional experiment reported in the rebuttal.

(Q3) SFT learns complete HTML/CSS/JavaScript implementations, so the selected pages carry supervision beyond isolated visual styling. A page with complete content, usable controls, and working interactions is also a more complete code example. Conversely, empty charts, blank dashboards, or persistent loading spinners often reflect incomplete implementations.

Although VENUS-CANON primarily addresses visual quality, its interaction and motion dimensions require appropriate responses to user actions, and its component dimension requires key controls such as buttons and navigation to render correctly. Candidates must also pass the interaction and motion checks before entering the corpus (Section 3.5). These requirements help explain why the selected training examples can improve functional performance as well as appearance.

Thank you for the clarifications and the additional exp:

W1 (Resolved)

W2 (Partially resolved): The corrected screenshot and additional examples help, but I would still like to see randomly sampled, fully rendered pages with the same score.

W3 (Largely resolved): The external WebDevJudge results show that the judge learns more than its own rubric. The matched Gemini baseline also strengthens the case for screening, though the contribution of the interaction checks to the functional gains remains unclear.

W4 (Still open): The examples show that VENUS can produce different styles, but not whether training makes less conventional styles less common.

I will keep my original score given the open W4

AuthorsOpenReview ↗

Reply to Reviewer 3QEx

Thank you for carefully reviewing the clarifications and additional experiments, and for recognizing W1 as resolved and W3 as largely resolved. We further clarify the design differences among equally scored pages, the role of design judgement, and the evidence for stylistic diversity.

(W2) Randomly sampled pages with the same score. The new Appendix N presents four randomly sampled, previously unshown WebGen-Bench tasks: a student portal, a children's recycling game, online polls, and stock reports. Both the base and VENUS-CODER receive an appearance score of 5/5. The screenshots allow comparison of composition and information organization within the same score category.

(W3) Design judgement and the dynamic check. The dynamic check evaluates interaction and motion, while VENUS-JUDGE assesses design quality across eight static dimensions, including layout, typography, colour, and components. A page responding to user actions does not ensure well-aligned navigation, clear hierarchy, or coherent components. The designer study in Appendix E also supports this design judgement capability: across pairs spanning the eight static dimensions, the judge agrees with independent designers on 77.2% of their decisive judgements on average.

The generation results provide complementary evidence. WebGen-Bench reports appearance separately, and Tables 2 and 3 show improved visual presentation. ArtifactsBench's per-task checklists provide additional evidence about implementation and presentation quality, while the Coder study in Appendix E confirms that professional designers see benefits in the generated pages. The existing screening comparisons evaluate the combined pipeline and do not separately quantify each check's contribution to generation gains. Our response therefore rests on the independent evaluation of design judgement and the visual improvement of the resulting pages.

(W4) Style and design constraints. We clarify two aspects.

(1) Style and design constraints serve different purposes. Style defines a page's visual language; VENUS-JUDGE focuses on hierarchy, alignment, spacing, readability, and component coherence within that style. Candidates are compared with their respective seeds, without requiring pages in different styles to share a palette or composition.

(2) Style diversity was an important consideration in seed construction. Seed-page production included minimalist, hand-drawn illustrative, soft-gradient, expressive, technology-oriented, and nature-inspired styles. The new Appendix M (Figure A22) shows style examples from a subset of seed pages. For copyright and privacy reasons, it presents reduced views of a subset of seed pages. This shows that stylistic coverage was an explicit choice at the seed-construction stage, with candidates subsequently screened for design quality against their respective seeds. The photography, expressive typography, and geometric compositions in Appendix J further illustrate the model's use of different visual languages.

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

This paper proposed Venus, including a judge and a coder. The judge can screen the aesthetic quality of a page and is used to filter datasets to train the coder. The coder trained on this dataset can produce higher aesthetics quality webpages.

Critical strengths
  • The motivation of this paper is clear: now models generate universal, templated webpages when prompted with generic keywords.
  • The Venus-Coder performs well on the coding task; it even surplasses some closed-source models.
  • The separation of what dimensions are needed in a page is a valid design.
  • The expert input in this work is valuable.
Critical weaknesses
  • The author criticized the model for producing generic output when prompted with vague words (S1). However, Venus-Coder seems to fall into the same pitfall. In Figure A4, the UI generated by Venus-Coder exhibits the strong, neon colored problem for the two mini-programs. This is similar to the classic “purple problem” in “vibe coding.” I am concerned that, as the model is optimized for a single standard, it reinforces this mode collapse style failure and produces websites that are generic.
  • The paper’s approach of filtering the corpus is very similar to rejection sampling SFT; the authors should consider citing a few papers on this.

Minor:

  • The images in the text are too small; the authors can use wrapfigure and enlarge them. It will be easier to understand the paper if at least one visual example can be included in the main text.
  • I am not sure if this is just a style choice, but in a lot of places the language is hard to follow for me:

Screening is the one stage whose cost grows with the pool rather than with the corpus, and it is where the corpus is made or lost. Where a dimension needs vocabulary, the standard supplies it, defining the colour roles by area and salience so that a single decisive accent becomes a question about regions of the screenshot. A derived page thus enters the corpus only when it improves on a reference already known to clear VENUS-CANON, not merely when it outranks other generated pages, and each page costs one order-consistent comparison

Questions for authors

For W2, could the author provide more examples, especially cases where it does not fall into this generic “vibe-coded” style or correct it when the base model has this style?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer kukS

Thank you for recognizing the paper's motivation, model performance, and the value of professional designer involvement. We address your concerns about style diversity, related work, and presentation below.

(W1) Style diversity is considered when constructing the seed set, before candidate screening. The seeds cover the six page genres described in Section 3.2: landing pages, slide decks, interactive tools, 3D scenes, games, and data visualizations. With guidance from professional designers and the design standard, we structure colour roles, visual hierarchy, composition, and other dimensions across these genres, then vary the requests while checking functional completeness during candidate screening.

This serves two purposes: it exercises VENUS-CANON across different page types and gives the local visual judge diverse, comparable design references. Each candidate is compared with its own seed rather than a shared style template. This preserves distinct references during screening and helps reduce pressure toward a single visual style. Derivation retains the seed's design intent while varying the scene, content, and surface constraints, supporting generalization across requests.

The additional cases illustrate editorial layouts, personal portfolios, and utility interfaces. Please see Figures A6–A9 in Appendix H (Additional Style and Composition Comparisons):

TaskDesign treatmentBase / VENUS appearance scores
70 — Poetry blogOlive/gold editorial layout with search and category organization5 / 5
71 — Personal portfolioWarm peach/red palette and alternative first-screen composition5 / 5
79 — Diagram editorLight workspace with separate toolbar, canvas, shape palette, and properties panel3 / 5
83 — Case-management dashboardBeige/brown navigation, summary cards, and tables4 / 4

Figures A12–A14 in Appendix J (Additional Pages Generated by VENUS-CODER) show six further generated pages, including light photographic layouts, expressive typography, geometric imagery, and product interfaces. Together, these cases illustrate a broader range of visual expression.

(W2) We agree that our corpus construction is related to rejection fine-tuning (RFT), as illustrated by Yuan et al. (2023) and STaR (Zelikman et al., 2022): generate candidates, select examples, and fine-tune on the retained data. VENUS differs in two aspects.

First, the source of candidates differs. In these approaches, the model being fine-tuned generates its own candidates. Selection may therefore favour output patterns that the model already produces frequently. VENUS-CODER does not generate its training candidates. Instead, external frontier systems generate pages from instructions derived from designer-reviewed seeds under VENUS-CANON (Section 3.3). The candidate distribution is thus guided by the design standard and seed coverage rather than VENUS-CODER's own generation preferences.

Second, the selection criterion differs. RFT commonly selects examples using task correctness or a scalar reward. VENUS compares each candidate with its designer-approved seed in both presentation orders and retains it only when VENUS-JUDGE prefers the candidate in both comparisons (Section 3.5). The selection target is design quality relative to a reviewed reference.

We have added these references and the methodological distinction to Section 2, Related Work. Your comment also helps us clarify the potential risk of selection narrowing the training distribution, which we discuss here.

(W3) We apologize that the original figure sizes made the page details difficult to inspect. We have adjusted the main-text figure and table layout using wrapfigure and enlarged the relevant images to improve readability.

(W4) We have revised the passages you identified in the method section. They now state directly how colour roles are defined, why screening cost scales with the candidate pool, and why a candidate must outperform its seed in both presentation orders to be retained. We will continue to refine the writing for clarity.

Thanks to the authors for the rebuttal. For my W1, the authors did not address my concerns but the authors addressed my concerns for other parts. My concerns were mostly that if the training will lead the model to generate a uniform taste. The examples in the appendix also did not fully show whether the training will correct the model from generating generic content. Given my score was already leaning toward accept (3), I will maintain my score.

AuthorsOpenReview ↗

Reply to Reviewer kukS

Thank you for acknowledging the other responses and clarifying W1. We understand that the key questions concern whether the trained model organizes its design around the task and whether it develops one taste or a generic template.

VENUS-CANON specifies relationships within a page without requiring a shared appearance across pages. Component consistency, clear hierarchy, and coherent spacing apply to blogs, portfolios, and tools, but do not require the same palette or composition. Controlled degradation introduces specific defects within an existing design, and candidates are compared with their respective seeds, directing supervision towards design quality. The new Appendix L makes this concrete: the portfolio comparison selects A for navigation aligned with the page centre and discards B for its offset navigation; the festival comparison selects A for separating the scroll cue from the buttons and discards B for placing them too close together. The pairs retain their beige-paper and deep-purple styles, respectively, while the preferences concern their own spatial organization.

Task-specific information structure is more informative about generic templates than palette changes alone. Appendix H provides base-model and VENUS-CODER comparisons for the same requests. In the diagram editor in Figure A8, VENUS-CODER places the toolbar above the diagram and separates shape selection, the canvas, and properties into a clearer workspace. The poetry blog organizes reading, search, and categories, while the management dashboard organizes navigation, summaries, and tables, reflecting different information needs. Appendix J adds photography-led pages, a typographically expressive design studio, a geometric showcase, and product interfaces, illustrating further visual treatments.

These examples show that following a design standard can coexist with task adaptation and varied visual expression. We ground the argument in inspectable structural improvements and the range of outputs, without treating selected examples as comprehensive correction of all generic content or a complete measurement of the style distribution.

AuthorsOpenReview ↗

Reply to Reviewer kukS

Thank you for highlighting the concern about a uniform taste. Style was an important consideration when selecting seed pages. The new Appendix M (Figure A22) shows a subset of seed pages spanning minimalist, hand-drawn illustrative, soft-gradient, expressive, technology-oriented, and nature-inspired styles. For copyright and privacy reasons, the figure presents reduced views of a subset of seed pages.

Seed selection provides different visual languages, while VENUS-JUDGE compares candidates with their own seeds and evaluates hierarchy, alignment, spacing, readability, and component coherence within each style. Appendix H shows responses to different task and style requirements: an olive-toned editorial layout for a poetry blog, a warm composition for a portfolio, and a light workspace for a diagram editor. Appendix J adds photography, expressive typography, and geometric compositions. Together with the style examples from a subset of seeds in Appendix M, these additions demonstrate our consideration of both stylistic diversity and design quality at the seed-construction and generation stages.

Reviewer_PAJxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VENUS converts a structured web-design standard into supervision through controlled degradation of designer-approved pages. It trains a local aesthetic judge to filter generated webpages, then fine-tunes a web-coding model on the selected corpus, reporting improvements in visual quality and functional performance.

Critical strengths
  1. The paper addresses a practical problem and presents a coherent pipeline connecting professional design criteria, synthetic preference supervision, data filtering, and web generation.

  2. The size-matched corpus ablation and blinded designer evaluation provide useful evidence that the selected data improves generation beyond simply adding more training examples.

Critical weaknesses
  1. Several details needed to assess and reproduce the pipeline are missing: the numbers of seeds and designers involved in seed screening; the number of degradation pairs and arena preferences used to train VENUS-JUDGE; the absolute sizes of the candidate pool and final corpus (only the 68.8% retention rate is given); the specific frontier models used for candidate generation; and the training hyperparameters for VENUS-CODER.

  2. While Table 3 supports the benefit of screening, it does not disentangle the effect of the judge from that of the dynamic check, leaving their individual contributions unclear. Furthermore, it reports no variance across random seeds or subset draws, making it difficult to assess the statistical robustness of the observed gains.

  3. The designer agreement study uses operator-generated pairs, which share the same type of synthetic degradation as the training data. During screening, however, the judge compares a derived page against its seed, and the two differ in scene and content. It would help to validate the judge on such pairs, for example with a designer spot check of screening decisions.

  4. Both ArtifactsBench scores and WebGen-Bench appearance ratings rely solely on Gemini 3.1 Pro as the evaluator. While the designer study supports gains over the base model, the comparisons against frontier baselines lack independent human evaluation or cross-judge validation. Reporting confidence intervals and evaluating with a secondary judge would help establish whether the reported rankings are robust.

Questions for authors
  1. Do the candidate generators include any of the models compared in Table 2, or Gemini 3.1 Pro, which is used as the evaluation judge?

  2. How much of VENUS-JUDGE’s performance comes from the operator pairs versus the arena preference data? Training the judge on each source separately would clarify this.

  3. The Font grade-5 criterion favors rendering display type as SVG or an image. How does this criterion account for accessibility, given WCAG’s preference for actual text when an equivalent presentation is achievable? Does it distinguish accessible SVG text from outlined or rasterized text?

5 comments
AuthorsOpenReview ↗

Reply to Reviewer PAJx(1/2)

Thank you for your careful review and for recognizing the value of connecting professional design standards with data screening. We address your concerns about reproducibility, the contribution of screening, and evaluation robustness below, using additional explanations and experiments.

(W1) We provide the following implementation details. Seed pages were screened jointly by 1 professional designer and 6 professional evaluators familiar with VENUS-CANON. Depending on the derivation dimensions, each seed produced approximately 50–100 candidate pages. The training data for VENUS-JUDGE comprise approximately 4k pairs, with a ratio of roughly 4:1 between synthetic degradation pairs and arena preference pairs. The candidate pool contains approximately 40k pages, of which approximately 28k were retained, corresponding to a 68.8% retention rate. Candidates were generated using a mixture of Gemini 3 Pro and Claude Opus 4.6.

VENUS-CODER uses Qwen3.6-35B-A3B as its base model and is trained through full-parameter fine-tuning in BF16, with a learning rate of 1 × 10⁻⁵, a global batch size of 256, a maximum sequence length of 32,768 tokens, and 3 epochs per training run. These settings are provided in the Coder training paragraph of Appendix F (Implementation Details).

(W2) We conducted three additional runs using randomly selected subsets of the candidate pool to examine variation across subset draws:

RunArtifactsBench / 100WebGen functional accuracyWebGen appearance / 5
random159.2136.6%3.09
random258.6838.6%3.48
random358.6243.0%3.22

ArtifactsBench scores range from 58.62 to 59.21, remaining below VENUS-CODER's 64.52 across these runs. We also report the WebGen results so that variation on both benchmarks can be assessed.

ArtifactsBench and WebGen-Bench originally use Gemini 2.5 Pro and GPT-4o, respectively. In our experiments, we use Gemini 3.1 Pro as a strong evaluator available at the time. We further examined an alternative selection method: Gemini 3.5 Flash scores and ranks the candidate pages, and a subset is selected at the same retention ratio. The results are:

Selection methodArtifactsBench / 100WebGen functional accuracy, original protocolWebGen appearance / 5
Gemini 3.5 Flash score-and-rank selection59.4841.4%3.35

The resulting model scores below VENUS-CODER on all three metrics (64.52, 57.0%, and 4.26, respectively). This provides an additional comparison supporting the value of professional design supervision and screening against a reviewed reference.

(W3) Seed derivation and controlled degradation are both guided by the design standard, but serve different purposes. Derivation varies the scene, content, and surface constraints while preserving the design intent; degradation deliberately introduces design defects to construct comparison supervision.

Appendix I (A Seed-to-Derivative Screening Example) illustrates the actual seed–derivation–selection workflow. Figure A10 shows Kaito Morita (Game Audio Director), the original seed, together with two pages derived from that seed: Mina Sato (Spatial Audio Systems Designer), which is retained, and Mina Vega (Interactive Foley Systems Designer), which is rejected. Figure A11 provides enlarged views of the retained and rejected pages' components. The appendix includes the medium-length seed and derived requests alongside their corresponding HTML pages. This example illustrates that preserving the seed's style is insufficient on its own: the derived page must also present its main components completely.

AuthorsOpenReview ↗

Reply to Reviewer PAJx(2/2)

(W4) We reevaluated the previously generated pages with three additional judges—Gemini 3.8 Flash, Gemini 3.5 Flash Lite, and JEV—using the original evaluation criteria. While ArtifactsBench and WebGen-Bench originally use Gemini 2.5 Pro and GPT-4o, respectively, our original experiments use Gemini 3.1 Pro for both benchmarks. We report its results alongside those of the additional judges below. Parentheses indicate each generator's rank within the corresponding judge column, with ties determined from scores rounded to two decimal places.

ArtifactsBench (0–100)

GeneratorGemini 3.1 ProGemini 3.8 FlashGemini 3.5 Flash LiteJEV
VENUS-CODER(1) 64.52(3) 62.99(1) 85.41(1) 53.79
DeepSeek-V4-Pro(2) 63.91(1) 63.31(5) 79.77(2) 53.04
GLM-5.2(3) 62.60(2) 63.27(4) 79.84(4) 51.43
GPT-5.5(4) 62.55(4) 62.19(3) 80.35(3) 52.93
Gemini-3.6-Flash(5) 60.26(5) 62.03(2) 82.33(5) 50.45
Claude-Opus-4.8(6) 56.88(6) 57.25(6) 79.09(6) 49.65
Qwen3.6-35B-A3B base(7) 53.50(7) 52.65(7) 75.60(7) 48.92
AesCoder-4B(8) 38.87(8) 38.91(8) 64.76(8) 42.51

VENUS-CODER ranks first under Gemini 3.1 Pro, Gemini 3.5 Flash Lite, and JEV, and third under Gemini 3.8 Flash, only 0.32 points below the highest mean score. Its gains over the base model are 11.02, 10.34, 9.81, and 4.87 points, respectively, in the order of the judge columns.

WebGen-Bench appearance (1–5)

GeneratorGemini 3.1 ProGemini 3.8 FlashGemini 3.5 Flash LiteJEV
VENUS-CODER(2) 4.26(3) 4.03(2) 4.91(4) 3.97
DeepSeek-V4-Pro(4) 4.15(4) 3.98(5) 4.84(5) 3.95
GLM-5.2(7) 3.29(5) 3.95(3) 4.86(3) 4.12
GPT-5.5(3) 4.22(1) 4.27(1) 4.92(1) 4.47
Gemini-3.6-Flash(1) 4.33(2) 4.22(3) 4.86(2) 4.13
Claude-Opus-4.8(4) 4.15(6) 3.71(6) 4.82(6) 3.94
Qwen3.6-35B-A3B base(6) 3.55(7) 3.64(7) 4.59(7) 3.59
AesCoder-4B(8) 3.13(8) 3.46(8) 4.48(8) 3.53

VENUS-CODER receives higher appearance scores than the base model under all four judges, with gains of 0.71, 0.39, 0.32, and 0.38 points, respectively. These results support its competitiveness and show that the direction of improvement over the base model is consistent across the judges.

(Q1) Candidates were generated using Gemini 3 Pro and Claude Opus 4.6. Neither the specific model versions compared in Table 2 nor the evaluation judges were used as candidate generators. We distinguish these versions explicitly to avoid conflating different releases within the same model family.

(Q2) We compared training on synthetic degradation pairs alone, natural preferences alone, and their combination:

Training dataStrictConsistency
Synthetic degradation only52.794.1
Natural preferences only54.192.2
Full mixture56.595.2

Metric definitions: Each test pair is judged in both A/B and B/A orders, and the reverse-order prediction is mapped back to the original page identities. Strict accuracy counts a pair as correct only when both verdicts match the human preference label. Order consistency counts a pair as consistent when both non-missing verdicts agree after this mapping, regardless of whether they match the human label; two tie verdicts also count as consistent. Missing predictions receive no credit under either metric but remain in the denominator. Strict therefore measures correctness in both orders, while Consistency measures stability under swapping.

Among the reported runs, the full mixture achieves the highest Strict and Consistency scores, providing additional evidence that the two supervision sources are complementary.

(Q3) The original Font grade-5 wording did not sufficiently delimit the use of SVG or images for display lettering, and we apologize for the ambiguity. Its intended scope is decorative lettering, rather than converting all text into graphics. Informational content remains real text; vector or image rendering is intended for display or decorative elements whose removal would not affect understanding of the page's information.

We have clarified that display lettering may be rendered graphically only when real text cannot achieve an equivalent visual effect. The revision appears in the Font grade-5 entry in Appendix A (The Full VENUS-CANON).

Thank you for the additional runs and the revised appendix. Two concerns remain.

First, the new W2 results still do not separate the judge from the dynamic check, so I cannot rule out that the gain comes from removing pages with broken interactions rather than from the aesthetic judge.

Second, the only screening example (Appendix I) shows a derivative rejected for incomplete rendering, not for a difference in design quality. The same appendix shows that derivatives preserve the seed’s design, so the corpus may be stylistically narrow.

I will therefore maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.