🤖 AI Summary
This work addresses the challenges faced by multimodal large language models (MLLMs) in automatically annotating mobile user interfaces, which are often densely populated, hierarchically nested, and visually ambiguous, leading to limited annotation accuracy and high sensitivity to prompt design and task formulation. To mitigate these issues, the authors propose a context-aware, staged annotation workflow that decomposes the overall task into multiple coordinated phases through structured prompting, schema-constrained JSON output, and element-specific instructions. Experimental results demonstrate that this approach significantly enhances the reliability of UI understanding: on the MUIAnno dataset, a two-stage pipeline achieves the highest precision, while deeper task decomposition improves recall at the cost of increased false positives, revealing a joint influence of decomposition depth and element category grouping on annotation quality.
📝 Abstract
Accurate mobile user interface annotation is important for UI understanding, accessibility tools, automated testing, dataset construction, and GUI agents. However, mobile screens are difficult to annotate because they often contain small, dense, nested, and visually ambiguous elements. Multimodal large language models can help automate this process, but their outputs are sensitive to prompt design and the organization of annotation tasks. This paper studies automated mobile UI annotation from a workflow design perspective, focusing on improving annotation precision. Rather than asking the model to annotate all UI elements in a single step, the task is divided into smaller context-aware stages, allowing related UI elements to be handled with clearer instructions and useful screen context. The proposed pipeline uses structured prompts, schema-constrained JSON outputs, and element-specific annotation instructions. Experiments are conducted on expert-annotated mobile UI screens from the MUIAnno dataset, using eight common UI element types: button, tab, clickable text, card, label, plain text, icon, and image. Four workflow strategies are evaluated: one-step, two-step, four-step, and eight-step annotation. Results show that the two-step workflow achieves the highest precision, while deeper decomposition improves recall but produces more false positives. Additional grouping experiments show that annotation quality depends on both workflow depth and element-class grouping. Overall, careful workflow design can make LLM-based mobile UI annotation more reliable for UI understanding, dataset construction, and GUI agent development.