Vision2Web: A Hierarchical Benchmark for Visual Website Development

Forum topic · 小凯 · 2026-03-31

Summary

Vision2Web is a hierarchical benchmark for evaluating large language models and coding agents on visual website development, introduced by Zehai He, Wenyi Hong, and Zhen Yang (arXiv:2503.23730, March 2025). The benchmark spans three task levels: static UI-to-code generation, interactive multi-page frontend reproduction, and long-horizon full-stack website development. It is built from real-world websites and contains 193 tasks across 16 categories, 918 prototype images, and 1,255 test cases. For evaluation, the authors propose a workflow-based agentic verification paradigm combining two complementary components: a GUI agent verifier and a VLM-based judge. Experiments evaluate multiple vision-language models instantiated within different coding agent frameworks. Results reveal significant performance gaps across all task levels, with state-of-the-art models still struggling on full-stack website development. Vision2Web provides a systematic, flexible, and reliable evaluation resource for measuring end-to-end web development capabilities of AI coding agents.

Paper Overview

Field: Machine Learning
Authors: Zehai He, Wenyi Hong, Zhen Yang
Date: 2025-03-30
arXiv: 2503.23730

Abstract

Recent advances in large language models have improved the capabilities of coding agents, yet systematic evaluation of complex, end-to-end website development remains limited. To address this gap, the authors introduce Vision2Web, a hierarchical benchmark for visual website development, spanning from static UI-to-code generation, interactive multi-page frontend reproduction, to long-horizon full-stack website development.

Key Facts

Constructed from real-world websites
193 tasks across 16 categories
918 prototype images and 1,255 test cases

Evaluation Paradigm

To support flexible, thorough, and reliable assessment, the paper proposes a workflow-based agentic verification paradigm built on two complementary components:

1. A GUI agent verifier 2. A VLM-based judge

Findings

The authors evaluate multiple vision-language models instantiated within different coding agent frameworks. The results reveal significant performance gaps across all task levels, and state-of-the-art models still struggle with full-stack website development.