Models & Capabilities

Do GUI Agents Know When Not to Act? Enabling Conflict-Aware Termination for Multimodal GUI Agents

The agent side (2609.03438, "Do GUI Agents Know When Not to Act?"). CONFLICTGUI, a benchmark of instruction-internal and instruction-context conflicts, reveals execution-biased overcompliance: agents that perform well on feasible tasks continue to execute blindly under conflicting instructions. CONFLICTGUARD, an inference-time framework, is the repair.

AI Agency

What the source reports

The agent side (2609.03438, "Do GUI Agents Know When Not to Act?"). CONFLICTGUI, a benchmark of instruction-internal and instruction-context conflicts, reveals execution-biased overcompliance: agents that perform well on feasible tasks continue to execute blindly under conflicting instructions. CONFLICTGUARD, an inference-time framework, is the repair. The evaluation reading is the sharp part: no mainstream agent benchmark rewards not acting - so restraint never develops as a trained property.

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