safety
The GUI Agent Learned That Some Commands Should End
CONFLICTGUI tests contradictions inside instructions and between an instruction and the visible interface.
Summary
CONFLICTGUI tests contradictions inside instructions and between an instruction and the visible interface.
Five evaluated agents showed execution-biased overcompliance. An inference-time feasibility check plus conditional action modulation improved conflict-task success while preserving performance on feasible tasks in the authors' experiments.
Why it matters
CONFLICTGUI tests contradictions inside instructions and between an instruction and the visible interface.
Limits and context
No additional limitation was separately recorded.
Key claims
CONFLICTGUI tests contradictions inside instructions and between an instruction and the visible interface.
Evidence: source-2026-09-05-018
Sources
- arXiv preprint 2609.03438arXiv · primary research
Corrections
No corrections have been recorded for this story.