UI-Lens: Assessing General MLLMs' Potential to Automate UI Display Quality Assurance

The goal of multimodal large language models (MLLMs) is not simply recognition, but visual discernment. This includes the critical capability to distinguish whether content is partially rendered or merely obscured by another element. Current models, which excel at object identification, fundamentally struggle with this ability, limiting their robustness in complex, layered environments. This challenge is acutely evident in User Interface (UI) display defect detection, which requires fine-grained element boundary understanding, missing-content detection, and reasoning about sequential interface semantic consistency. However, the capabilities of multimodal large language models (MLLMs) and vision-language models (VLMs) for detecting UI defect in realistic and complex interfaces have not been systematically validated. To fill this gap, we present UI-Lens, a UI display defect detection benchmark for multilingual UI scenarios. The dataset consists of 4,759 Chinese and 3,392 English interfaces meticulously annotated by design experts, covering six display defect categories. We conduct a systematic evaluation of 9 mainstream models (7 closedsource, 2 open-source). Results show clear shortcomings in current models: for tasks requiring fine-grained element boundary understanding, performance is near-random, with task-average F1 scores of 22.19% and 33.75% on Text Overflow and Container Overlap, respectively; for sequential interface semantic consistency, the task-average F1 score is only 11.44%, indicating severe underperformance. We release UI-Lens to catalyze research toward robust UI display defect detection with fine-grained boundary awareness in realistic interfaces. The dataset is available on this link.