GUI Grounding for Computer-Use Agents: A Survey

Graphical user interface (GUI) agents automate tasks in digital environments by interpreting user goals and executing interface operations. GUI grounding serves as a critical link between high-level reasoning and low-level interaction execution in GUI agents. It maps users' operation intents to specific interaction targets and their spatial parameters, enabling agents to translate high-level decisions into executable interface actions. Advances in multimodal large language models (MLLMs) have substantially improved GUI agents' capabilities in autonomous operation. However, existing surveys primarily focus on complete GUI-agent systems, leaving grounding methods, data resources, and performance limitations insufficiently synthesized. To the best of our knowledge, this is the first comprehensive survey dedicated to GUI grounding. First, we establish a unified taxonomy along three complementary dimensions—learning paradigms, inference strategies, and target representations—to systematically characterize the core mechanisms and properties of existing methods. Second, we review training datasets, evaluation benchmarks, and metrics, tracing the evolution of data resources and evaluation tasks. Third, we conduct a comparative analysis of representative methods across evaluation benchmarks, identifying their strengths and limitations across target types, instruction complexity, and interaction requirements, while examining trade-offs between accuracy and inference cost. Our analysis reveals that GUI grounding research is expanding from explicit target localization in static interfaces toward implicit intent understanding, target resolution in dynamic interfaces, and spatial-parameter prediction for complex interactions. We discuss the resulting open challenges and future research directions.