AI News

GUI agents are increasingly used to execute natural-language instructions on graphical interfaces, but users can issue instructions that contain internal contradictions or conflict with the visible interface context. This study introduces CONFLICTGUI, a benchmark for evaluating whether agents recognize such situations and terminate instead of acting. The evaluation finds severe execution-biased overcompliance: agents that perform well on feasible tasks often continue executing blindly when instructions are conflicting. The authors propose CONFLICTGUARD, an inference-time framework with two coupled components. A feasibility verification protocol asks the agent to assess the instruction's logic and evidence in the GUI before acting, while conditional action modulation redirects behavior toward termination when execution is inappropriate. Experiments across five widely used agents show a significant improvement in average success on conflict tasks while preserving performance on ordinary GUI tasks. The results suggest that a lightweight inference-time intervention can improve an agent's ability to identify unsuitable execution scenarios and refrain from unnecessary actions.