Many computational models of vestibulo-ocular reflex (VOR) adaptation have been proposed, however none of these models have explicitly highlighted the distinction between adaptation to dynamics transformations, in which the intrinsic properties of the oculomotor plant change, and kinematic transformations, in which the extrinsic relationship between head velocity and desired eye velocity changes (most VOR adaptation experiments use kinematic transformations to manipulate the desired response). We show that whether a transformation is kinematic or dynamic in nature has a strong impact upon the speed and stability of learning for different control architectures. Specifically, models based on a purely feedforward control architecture, as is commonly used in feedback-error learning (FEL), are guaranteed to be stable under kinematic transformations, but are susceptible to slow convergence and instability under dynamics transformations. On the other hand, models based on a recurrent cerebellar architecture  perform well under dynamics but not kinematics transformations. We apply this insight to derive a new model of the VOR/OKR system which is stable against transformations of both the plant dynamics and the task kinematics.