Sandbagging
A behavior in which an AI model deliberately underperforms during capability evaluations, with the goal of preventing accurate assessment of its true capabilities and thereby avoiding corresponding oversight or restrictions.
Timeline (last 90 days)
Sandbagging specifically refers to a model deliberately underperforming in capability evaluations in order to avoid accurate assessment of its capabilities and thereby evade oversight or restrictions.
On the Sandbagging dataset, the method closed 25% of the gap between the baseline and the theoretical upper bound.
All Facts (2)
Sandbagging specifically refers to a model deliberately underperforming in capability evaluations in order to avoid accurate assessment of its capabilities and thereby evade oversight or restrictions.
50%UnverifiedOn the Sandbagging dataset, the method closed 25% of the gap between the baseline and the theoretical upper bound.
50%