In recent years, demand for robot-assisted rehabilitation has increased due to the rising number of elderly and disabled people. Rehabilitation robots help patients to enhance muscle strength and recover motor functions, typically through practicing reaching movements. In this letter, we are interested in robot-assisted rehabilitation not only for simple trajectory following tasks but also for co-operative therapy tasks that elicit a force-based or an impedance-based behavior from the patient. When a patient is unable to complete the task, determining the minimum-required assistance to be provided to the patient such that the task is accomplished is of interest. In this letter, we develop a learning from demonstration (LfD) framework to compare the performance of a therapist in multiple trials of the task (demonstrations in LfD terms) carried out previously with that of the patient in live performance of the task. Based on this performance differential, the LfD framework helps to determine the minimum required adjustment in the task's difficulty level on a patient-specific basis for the task to be completed. To encourage active and free participation of the patient, a dynamic bicycle cranking model is used such that provision of assistance (reproduction in LfD terms) is coupled to the variability observed in the therapist's behavior across various trials of the task. Experimental results show that the proposed framework effectively provides the patient with assistance as needed during a cooperative therapy task.