A Detection Method for Text Input-Induced Non-Crashing Functional Bugs in Mobile GUI Testing

In mobile GUI testing, non-crashing functional bugs (NCFBs) induced by text input can lead to severe problems such as data errors, logical confusion, and display anomalies. Existing detection methods mainly rely on predefined heuristic rules or specific test oracles, which are difficult to adapt to diverse bug types, resulting in limited detection scope and low efficiency. To systematically address this challenge, this paper proposes an automated testing method based on multimodal large language models (MLLMs). We first construct a three-stage interactive "generation-exploration-verification" automated testing framework: using LLMs to deeply understand the semantics and constraints of input components for generating abnormal inputs; planning test paths through knowledge base construction and retrieval-augmented generation (RAG) technology; and utilizing MLLMs to fuse GUI screenshots and semantic information for identifying and classifying NCFBs. We comprehensively evaluated the method on a dataset consisting of 315 real bugs and 45 injected bugs. Comparison with 21 baseline methods demonstrates its overall effectiveness, and ablation experiments further confirm the contribution of each module. Additionally, 19 previously unknown NCFBs were successfully discovered in real-world application testing, 10 of which have been confirmed or fixed by developers, proving its practical value.