DragonCrawl paper reports LLM-driven mobile E2E tests at 91%+ pass rates in CI

An arXiv preprint introduces DragonCrawl, an AI-driven mobile end-to-end testing system aimed at continuous regression testing rather than only exploratory crash hunting. The authors say traditional E2E frameworks struggle with UI volatility, maintenance overhead, and cross-platform scale as apps grow more complex. DragonCrawl is described as evolving from embedding-based similarity matching to generative intent-based reasoning with large language models. Using GPT-4o's multimodal capabilities, the system reports a 91.6% pass rate on iOS and 92.2% on Android across 1,013 automated tests running continuously in CI/CD pipelines, and it blocks commits that break critical user flows. The paper states test onboarding time falls from 96-120 hours to under 4 hours and estimates 27 developer years saved in test maintenance. Multimodal vision for end-state detection and tool calling for backend state transitions are cited as ways to connect UI steps with system state. Authors discuss production deployment challenges including token explosion and memory constraints and argue AI-driven testing can cut brittleness while supporting continuous quality assurance at scale.