SeerGuard uses a world model to screen risky actions in mobile GUI agents

Xue Yu · hf · 2026-07-22

The paper proposes SeerGuard, a consequence-aware safety framework for mobile GUI agents.
It combines pre-execution instruction screening with action-level risk assessment so the system can predict likely outcomes before the agent acts.
The authors build a unified safety-augmented world model (SAWM) via multi-task learning, mixing next-state prediction and safety risk assessment.
On Qwen3-VL-8B-Instruct, the safety-utility score rises from 0.191 to 0.596 at ω=0.8, while risk-cost drops from 0.347 to 0.130 at α=0.8.
The paper argues the model generalizes across diverse mobile GUI agents and shows both screening and action-risk prediction contribute to the gains.

More from coding & agent