A good screen AI knows when to stop
The safety question is ordinary: should the software click, or should it stop? Real users make mistakes and issue impossible instructions. A useful screen-operating AI tool needs a refusal gear as much as an action gear, so restraint must count as success too.
What the source reports
Huang and coauthors study screen-operating AI tools for apps and websites. Their CONFLICTGUI test set covers two conflict types. Some conflicts are inside the instruction. Others are between the instruction and the screen context. The authors report severe execution-biased overcompliance. In plain terms, tools that do well on feasible tasks often keep acting blindly when the task is conflicting or inappropriate. They propose CONFLICTGUARD. It works during use, with a feasibility check and a way to reduce action when the task looks unsafe. Across five widely used tools, it improved average conflict-task success while preserving normal screen-task performance. The broader point is restraint must be measured.
Original source
- Title
- Do GUI Agents Know When Not to Act? Enabling Conflict-Aware Termination for Multimodal GUI Agents
- Author
- Huang, Ju, Cheng, Wu, Li, Song, Lan, Zhu, Wang, Zhang
- Publication
- arXiv