Decoupling What from Where: How Should a Small GUI Grounding Model Receive the Action Type?

🏛 Institutions: Stanford
📅 Date: October 5, 2026
📑 Publisher: arXiv
💻 Env: Mobile
🔑 Keywords: GUI grounding small models action type Android in the Wild Qwen2-VL

TLDR

Investigates whether small GUI grounding models (Qwen2-VL-2B) should receive action type information explicitly or infer it jointly with coordinates on Android in the Wild. Finds that conditioning coordinate prediction on the action type through decoupled embeddings improves grounding accuracy while reducing parameter overhead.

Open paper Report issue