Negative control task
A control mechanism in experimental design in which samples known to produce negative results are deliberately included to verify the reliability of the detection procedure itself. In agent evaluation, negative control tasks are designed to be underdetermined and impossible to complete honestly, so that the only correct output is to give up.
Timeline (last 90 days)
Negative control tasks originate from medical and psychological research, referring to deliberately adding samples known to produce negative results into an experimental sample, in order to verify whether the detection process itself is reliable.
In Agent evaluation, negative control tasks are designed as tasks with insufficient information that cannot be honestly completed, where the only correct output is to give up.
The refusal slice borrows from the practice of slice analysis in machine learning evaluation, specifically referring to the refusal accuracy and invalid completion rate on the subgroup of negative control tasks.
All Facts (3)
Negative control tasks originate from medical and psychological research, referring to deliberately adding samples known to produce negative results into an experimental sample, in order to verify whether the detection process itself is reliable.
50%UnverifiedIn Agent evaluation, negative control tasks are designed as tasks with insufficient information that cannot be honestly completed, where the only correct output is to give up.
50%UnverifiedThe refusal slice borrows from the practice of slice analysis in machine learning evaluation, specifically referring to the refusal accuracy and invalid completion rate on the subgroup of negative control tasks.
50%