The prevailing model of cerebellar learning states that climbing fibers (CFs) are both driven by, and serve to correct, erroneous motor output. However, this model is grounded largely in studies of behaviors that utilize hardwired neural pathways to link sensory input to motor output. To test whether this model applies to more flexible learning regimes that require arbitrary sensorimotor associations, we developed a cerebellar-dependent motor learning task that is compatible with both mesoscale and single-dendrite-resolution calcium imaging in mice. We found that CFs were preferentially driven by and more time-locked to correctly executed movements and other task parameters that predict reward outcome, exhibiting widespread correlated activity in parasagittal processing zones that was governed by these predictions. Together, our data suggest that such CF activity patterns are well-suited to drive learning by providing predictive instructional input that is consistent with an unsigned reinforcement learning signal but does not rely exclusively on motor errors.