WebCompass: First Multimodal Benchmark for Full-Lifecycle Web Coding Agent Evaluation
arXiv·medium signal
WebCompass provides unified evaluation of web coding agents across visual fidelity, interaction quality, and codebase-level reasoning — dimensions that existing benchmarks miss by only measuring text-conditioned generation with static-correctness metrics. The benchmark covers the full web engineering lifecycle from design-to-code to interactive behavior. Fills a critical evaluation gap as code agents increasingly handle end-to-end web development.