跳转至

文章背景与核心概要

近年来，多模态大模型（MLLMs）的快速发展使得将视觉UI设计直接转换为可执行代码成为可能。然而，真实世界的移动应用需要完整的、可构建的代码库，其中包含共享组件、合理的架构以及正常工作的跨页面导航。传统的“设计转代码”基准测试往往局限于单页面生成，无法满足这一需求。

为了填补这一空白，本文作者推出了 MobileForge，这是首个专为多屏幕移动应用生成而设计的项目级基准测试。该基准测试包含了真实的移动应用、经过人工审核的屏幕、结构化的页面关系注释以及导航测试规范。此外，它支持五维评估框架（构建、导航、视觉保真度、代码可维护性和效率），并引入了创新测试协议以确保评估的可靠性。

Looks Right, Works Right: A Project-Level Benchmark for Multi-Screen Mobile App Generation

Summary

Recent advancements in multimodal large language models (MLLMs) enable the conversion of visual UI designs directly into executable code. However, real-world mobile applications demand complete, buildable codebases featuring shared components, proper architecture, and working cross-page navigation. Traditional design-to-code benchmarks fall short by focusing solely on single-page generation.

To address this gap, the authors introduce MobileForge, the first project-level benchmark designed for multi-screen mobile app generation. MobileForge includes real mobile applications, human-reviewed screens, structured page-relationship annotations, and navigation test specifications. Furthermore, it supports a five-axis evaluation framework (build, navigation, visual fidelity, code maintainability, and efficiency) alongside novel testing protocols to ensure reliable evaluation.

Paper Metadata

Field Details
arXiv Identifier arXiv:2607.28645 [cs.HC]
Primary Subject Human-Computer Interaction (cs.HC)
Secondary Subjects Artificial Intelligence (cs.AI)
Authors Fan Wu, Cuiyun Gao, Yiming Huang, Yang Xiao, Yujia Chen, Qing Liao
Conference Venue Accepted by EMNLP 2026 Main
Submission History Submitted May 29, 2026; Last revised August 21, 2026 (v2)
Project Repository GitHub - MobileForge Evaluation

Abstract

Recent multimodal large language models can convert visual designs directly into executable code, but real mobile products require multiple screenshots to become a buildable codebase with shared components and working navigation. This project-level setting exposes three limits of existing design-to-code benchmarks:

  1. They focus on single-page generation rather than complete codebases.
  2. They cannot evaluate cross-page navigation.
  3. They do not measure project-wide maintainability.

We introduce MobileForge, the first benchmark for project-level multi-screen mobile app generation, comprising real mobile apps, human-reviewed screens, structured page-relationship annotations, and navigation test specifications. MobileForge supports five-axis evaluation of build, navigation, visual fidelity, code maintainability, and efficiency.

We also propose: * State-isolated navigation testing to avoid cascading failures in navigation evaluation. * Anchor-referenced list-wise visual evaluation protocols to improve visual-judge reliability.

Across end-to-end runs on six frontier multimodal LLMs, current models can successfully build mobile-app projects that compile and reach correct pages. However, interactive navigation remains unreliable, and visual fidelity as well as code maintainability still lag behind production standards.

全文与访问链接

Full-Text & Access Links

引用 (BibTeX)

@misc{wu2026looksrighthandworkswright,
      title={Looks Right, Works Right: A Project-Level Benchmark for Multi-Screen Mobile App Generation}, 
      author={Fan Wu and Cuiyun Gao and Yiming Huang and Yang Xiao and Yujia Chen and Qing Liao},
      year={2026},
      eprint={2607.28645},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2607.28645},
}

Citation (BibTeX)

@misc{wu2026looksrighthandworkswright,
      title={Looks Right, Works Right: A Project-Level Benchmark for Multi-Screen Mobile App Generation}, 
      author={Fan Wu and Cuiyun Gao and Yiming Huang and Yang Xiao and Yujia Chen and Qing Liao},
      year={2026},
      eprint={2607.28645},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2607.28645},
}