For skill learning processes to be effective, they must encode associations that are inherent to the current task and avoid those that are spurious or particular to training conditions so that learning can transfer to novel situations. Some everyday contexts even require grouped responding to simultaneously presented stimuli. Here we test whether learning of these grouped responses depends on overlap in stimulus and/or response modality or on the conceptualization of the stimulus and response streams as belonging to a common task. In the present experiments, participants made 2 responses to 2 simultaneously presented stimuli, and learning was assessed by comparing performance on response combinations that had been practiced throughout training to performance on combinations that had been withheld. Experiments 1-4 paired the same visual-manual task with a 2nd task that differed in terms of the stimulus modality, the response modality, neither modality, or both modalities. Combination-specific learning was only observed when both the stimulus and response modalities were the same for the 2 tasks. However, Experiments 5 and 6 showed that combination-specific learning could occur with nonoverlapping stimulus modalities or response modalities if the 2 tasks were conceptually related. The results suggest that task representations provide top-down constraints on skill learning processes.