Both the American Heart Association and the VA/DoD endorse upper-extremity robot-mediated rehabilitation therapy for stroke care. However, we do not know yet how to optimize therapy for a particular patient's needs. Here, we explore whether we must train patients for each functional task that they must perform during their activities of daily living or alternatively capacitate patients to perform a class of tasks and have therapists assist them later in translating the observed gains into activities of daily living. The former implies that motor adaptation is a better model for motor recovery. The latter implies that motor learning (which allows for generalization) is a better model for motor recovery. We quantified trained and untrained movements performed by 158 recovering stroke patients via 13 metrics, including movement smoothness and submovements. Improvements were observed both in trained and untrained movements suggesting that generalization occurred. Our findings suggest that, as motor recovery progresses, an internal representation of the task is rebuilt by the brain in a process that better resembles motor learning than motor adaptation. Our findings highlight possible improvements for therapeutic algorithms design, suggesting sparse-activity-set training should suffice over exhaustive sets of task specific training.