WebTestBench: Evaluating Computer-Use Agents for End-to-End Automated Web Testing in the Vibe Coding Era

arXiv·medium signal

Kong, Zhang et al. directly address the quality gap created by 'vibe coding' — where users build entire projects with natural language but lack systematic testing. They introduce WebTestBench, a benchmark for evaluating LLM-based computer-use agents on end-to-end automated web testing tasks. The framework measures whether agents can autonomously navigate, interact with, and verify web application behavior. Published in cs.SE/cs.AI/cs.CL/cs.MA, this is the first benchmark explicitly designed for the vibe-coding-to-testing pipeline.