Graphical user interface (GUI) agents can follow natural‑language instructions to manipulate screens, yet real users often issue infeasible commands due to simple mistakes. A trustworthy agent must not only know how to act but also when to refrain. To study this, we introduce the CONFLICTGUI benchmark, which contains instruction‑internal conflicts and instruction‑GUI context conflicts for evaluating conflict‑aware termination. Our evaluation shows that many agents that excel on feasible tasks become over‑compliant under conflicting instructions, blindly continuing execution. To address this, we propose the CONFLICTGUARD inference‑time framework that aligns feasibility awareness with action generation. CONFLICTGUARD consists of (1) a feasibility verification protocol that makes the agent assess instruction logic and GUI evidence before acting, and (2) a conditional action modulation mechanism that redirects the agent from blind execution to termination when a conflict is detected. Experiments on five widely used GUI agents demonstrate a significant boost in conflict‑task success rate while preserving normal task performance. These results confirm that a lightweight inference‑time intervention can markedly improve an agent’s ability to recognize inappropriate execution scenarios and avoid unnecessary actions.
Review