Landing pages are goal-oriented web interfaces that must communicate a target-specific value proposition while organizing information flow, visual hierarchy, and calls to action (CTA). Although large language models can generate plausible webpage code from natural-language prompts, direct generation often yields generic templates and unsupported persuasive claims. We study target-grounded, reference-guided landing-page generation, where a system must create an executable page for a new target by adapting reusable patterns from real pages without copying them. We introduce LandingBench, a reference-profile dataset that abstracts real landing pages into section sequences, layout patterns, tone descriptors, visual emphasis, and CTA structure. Building on LandingBench, we propose LandingAgent, a three-phase agentic framework that profiles the target, constructs a reference-guided wireframe, and refines the page through critique-guided polishing. We evaluate LandingAgent against direct prompting on faithfulness, conciseness, readability, aesthetics, and structural diversity. Experiments show improved target grounding, presentation quality, and layout diversity. Code is available at https://github.com/IAURAI/LandingAgent.
Figures & tables
Figure 2: Overview of the LandingBench collection and filtering pipeline. Candidate pages sampled from Common Crawl are processed through rule-based filtering, LLM-based scoring and human verification.
Table 1: Core reference-profile schema of LandingBench.
Figure 3: Overview of the LandingAgent framework. Given a user-provided target specification and optional image assets, LandingAgent first performs a Profiling Phase that transforms the inputs into a Structured Page Brief. In the Wireframing Phase, a Retriever selects a structurally relevant reference from LandingBench, and a WireframeBuilder constructs a skeletal page draft. A WireframeCritic either approves the draft or triggers a Restart Loop that reselects a reference and regenerates the wireframe. In the Polishing Phase, a Polisher iteratively refines the draft with feedback from three parallel critics: AestheticCritic, MessageCritic, and FunctionalCritic.
Phase
Agent
Role
Profiling
Profiler
Transforms the target specification and image assets into a Structured Page Brief.
Wireframing
Retriever
Selects a reference profile from LandingBench that matches the Structured Page Brief along structural and visual axes. When a restart occurs, it uses the diagnostic hint to avoid repeating the same structural failure.
WireframeBuilder
Constructs a skeletal page draft guided by the selected reference profile at the level of section roles, information flow, and layout organization, without using raw HTML or verbatim source-page copy.
WireframeCritic
Evaluates whether the draft is globally aligned with the Structured Page Brief and the retrieved reference profile. It either approves the draft for polishing or triggers a restart with a diagnostic hint.
Polishing
Polisher
Refines the approved wireframe into an executable landing page by adding target-grounded copy, visual styling, asset placement, and implementation details while preserving the global structure.
AestheticCritic
Assesses visual presentation quality, including typography hierarchy, color harmony, whitespace, alignment, visual readability, modernness, brand expression and reference fidelity.
Table 2: Functional role of each agent, grouped by phase.
Method
Generator
Critic
Diversity
Content
Presentation
Faithfulness
Conciseness
Readability
Aesthetics
Direct Prompting
gemini-2.5-pro
—
0.26
3.83
4.40
3.90
3.67
gpt-5
—
0.28
4.60
4.13
3.77
3.63
One-shot Prompting
gemini-2.5-pro
—
0.29
4.05
4.67
3.95
3.67
Direct Prompting + Polisher
gemini-2.5-pro
claude-sonnet-4-6
0.25
4.33
4.71
3.95
3.62
Landing Agent
gemini-2.5-pro
claude-sonnet-4-6
0.36
4.30
4.83
4.03
4.00
Table 3: Main evaluation results with structural diversity, content quality ( Faithfulness , Conciseness ), and presentation quality ( Readability , Aesthetics ). Scores are rounded to two decimal places.
Figure 4: Qualitative comparison of structural diversity. For the same target specifications, Direct Prompting repeats hero layouts, visual assets, and section order, whereas LandingAgent produces diverse yet target-faithful layouts, visual grounding, and section flows.
Figure 5: Qualitative comparison of Direct Prompting and LandingAgent on two target specifications: Faithfulness (top) and Aesthetics (bottom). Red boxes mark failure regions; green boxes mark successful regions.
Figure 6: Pairwise user study results: Ours (LandingAgent) vs. Baseline preferences across the five criteria.
Method
Div.
Faith.
Conc.
Read.
Aesth.
Hard filter only
0.29
4.00
4.43
4.00
3.86
w/o LandingBench
0.25
4.24
4.48
4.00
3.86
w/o RAG
0.29
4.10
4.38
4.00
3.86
w/o Profiling
0.32
4.24
4.52
4.00
3.95
w/o Polishing
0.24
1.14
1.76
3.38
2.24
Ours
0.38
4.33
4.71
4.05
4.00
Table 4: Ablation results for reference filtering, LandingBench availability, reference retrieval, and pipeline phases. Scores are rounded to two decimal places.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Stage
Output
Filtered
Initial crawl
1,000
—
After Hard Filter
938
62
After Soft Filter
636
302
After Human Filter
438
198
Appendix
Table 5: Candidate counts at each stage of the LandingBench filtering pipeline. The final 438 pages are those that passed all three filtering stages and have valid extracted input images and annotation files.
Table 10: Reference profile components extracted for each LandingBench page, organized by category.
Market category
Count
%
DevTools / Infra
82
18.7
Media / Content
80
18.3
Security
61
13.9
B2B SaaS
49
11.2
Marketing / Analytics
29
6.6
Other
27
6.2
Appendix
Table 11: Market-category distribution of LandingBench pages.
Axis
Category
Count
%
Brand archetype
Sage
107
24.4
Ruler
61
13.9
Creator
59
13.5
Caregiver
50
11.4
Hero
42
9.6
Magician
27
6.2
Appendix
Table 12: Distributions of brand archetype, company stage, brand polish level, and primary color hue across the 438 LandingBench pages. Brand metadata is extracted using Claude Sonnet 4.6 from each page’s screenshot and visible text. Archetype and stage categories are normalized from free-text model output. Color hue is derived from the primary palette hex value.
Table 13: Agent prompt summary. Each agent’s prompt follows a common Role / Task / Input / Output schema structure.
Metric
Definition
Diversity
Measures whether outputs generated for the same landing target differ in section composition, information flow, and layout pattern rather than repeating the same template.
Faithfulness
Measures whether the generated page accurately reflects the target specification, including product or service description, value proposition, intended audience, and conversion goal.
Conciseness
Measures whether the page communicates essential information without unnecessary repetition, excessive explanation, or vague promotional copy.
Readability
Measures whether the page has clear section flow, legible text, identifiable CTAs, and an interpretable layout.
Aesthetics
Measures the overall visual completeness of the page, including color use, typography, spacing, alignment, visual balance, and perceived polish.
Appendix
Table 14: Evaluation metric definitions.
Item
Setting
Judge model
Claude Sonnet 4.6
Evaluation format
Pointwise scoring
Score scale
1–5
Input to judge
Target specification and generated landing page.
Aggregation
Mean score per metric and method.
Appendix
Table 15: LLM-as-a-Judge evaluation settings.
Dimension
Score
Rubric
Faithfulness
1
Misrepresents the brand or omits the value proposition; fabricated claims.
2
Major gaps — several sections the brief calls for missing, audience/tone clearly off, or only a generic brand mood with no concrete explanation of the offering.
3
Broadly on-brief but with notable omissions, vague substance (purpose/value not concretely conveyed), missing brief-supplied pricing/product detail, or one fabricated element.
4
Faithfully covers the brief; minor omissions only.
5
Fully faithful — value prop, audience, tone, the concrete substance of the offering, brief-supplied pricing/product detail, and the sections the brief calls for all accurately represented.
Conciseness
1
Severe walls of text, or near-empty of real content — either way the message fails to land.
Appendix
Table 16: LLM-as-a-Judge scoring rubric.
Category
Count
%
DevTools / Infra
3
10.0
Media / Content
7
23.3
Security
1
3.3
B2B SaaS
4
13.3
Marketing / Analytics
3
10.0
Other
1
3.3
Appendix
Table 17: Industry-category distribution of the 30 evaluation target specifications. Percentages are rounded to one decimal place.
Setting
Value
Temperature
0.7 where supported by the model API
Generation seed
No fixed seed specified
Prompt-suite version
1.0
Maximum wireframe restarts
Rmax=3
Maximum polishing rounds
Tmax=3
Appendix
Table 18: Generation and reproducibility settings.
Count per run
0
1
2
3
Wireframe restarts
71
9
4
6
Polishing-review retries
25
17
9
39
Appendix
Table 19: Observed restart and retry distributions across 90 full-pipeline runs.
Variant
Description
Hard filter only
Uses a retrieval pool constructed from pages that pass only hard filtering, without soft filtering or human verification.
w/o LandingBench
Uses 438 samples randomly selected from Common Crawl as the RAG reference pool, preserving the retrieval mechanism and reference-pool size while replacing LandingBench.
w/o RAG
Removes LandingBench-based retrieval from the generation process.
w/o Profiling
Omits the Profiling phase from the generation process.
w/o Polishing
Omits the final Polishing phase and uses the generated page draft as the final output.
Ours
Uses the full LandingAgent pipeline with LandingBench-based retrieval and full filtering.
Appendix
Table 20: Ablation variants.
Figure 7: Additional qualitative comparisons between Direct Prompting and LandingAgent. From top to bottom: Gallery — the baseline omits required content; Concord — the baseline uses text whose color is too close to the background, hurting readability; View — the attached image overflows into the right margin and copy stretches vertically, reducing readability; Library — the baseline omits the CTA and renders an empty black region.
Figure 8: Example of brand de-identification for reporting. Left: Original screenshot, with identifiable text blurred to obscure recognizable commercial brand identifiers. Right: De-identified version in which all identifiable elements — brand-specific names, UI marks, and other recognizable content — are replaced with fictional counterparts (e.g., Discord → Concord).
Web application generation uses AI to turn natural-language, visual, or multimedia specifications into functional, interactive websites and applications.
Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks. We study this crucial capability through webpage development, a practical and readily evaluable setting in which existing benchmarks remain largely idealized: they typically focus on reconstructing visual targets or generating webpages from textual instructions alone, overlooking the more common workflow of developing from visual examples. Specifically, we introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. We evaluate the resulting open-ended artifacts using independently verifiable keypoints and complementary programmatic and VLM judges. Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
Yuchi Wang, Rongyao Fang, Weihao Yu +7
MMLab, CUHK · Alibaba Token Foundry · The Chinese Univerity of Hong Kong +2
The rapid progress of Artificial Intelligence Generated Content (AIGC) tools enables images, videos, and visualizations to be created on demand for webpage design, offering a flexible and increasingly adopted paradigm for modern UI/UX. However, directly integrating such tools into automated webpage generation often leads to style inconsistency and poor global coherence, as elements are generated in isolation. We propose MM-WebAgent, a hierarchical agentic framework for multimodal webpage generation that coordinates AIGC-based element generation through hierarchical planning and iterative self-reflection. MM-WebAgent jointly optimizes global layout, local multimodal content, and their integration, producing coherent and visually consistent webpages. We further introduce a benchmark for multimodal webpage generation and a multi-level evaluation protocol for systematic assessment. Experiments demonstrate that MM-WebAgent outperforms code-generation and agent-based baselines, especially on multimodal element generation and integration. Code & Data: https://aka.ms/mm-webagent.
Yan Li, Zezi Zeng, Yifan Yang +12
Shanghai Jiao Tong University · Xi’an Jiaotong University · Microsoft Corporation +1
Frontier LLMs can navigate complex websites, but their cost and reliance on third-party APIs make local deployment impractical. We introduce Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents by analogy to human annotation roles, replacing the Task Designer, Annotator, and Supervisor with modular LLM components. Using Gemini 3 Pro as teacher, we generate 3,000 trajectories across six web environments and fine-tune a 9B-parameter student with pure supervised learning on the 2,322 that pass quality filtering. The resulting model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol, and nearly doubling the previous best open-weight result (Go-Browse, 21.7%). Capabilities transfer to unseen environments, with an 18.2 percentage point gain on WorkArena L1 (an enterprise platform never seen during training) and consistent improvements across three additional benchmarks. Ablations confirm that each pipeline component contributes meaningfully, with Judge filtering, evaluation hints, and reasoning traces each accounting for measurable gains. These results demonstrate that structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. Project page: https://agent-as-annotators.github.io