In Experiment 1, brightness discriminations were established under conditions in which the stimulus to which "same" responses were required (S+) was presented more frequently than the stimulus to which "different" responses were made (S-), S+ and S-were presented equally often, or S-was more frequent than S+. Subsequent generalization gradients along the brightness dimension were shifted away from the stimulus value more frequently presented during training. In Experiment 2, brightness discriminations were established under conditions of equally frequent presentations of S+ and S-stimuli in training, but in subsequent generalization testing S+ and S-values were equally represented or overrepresented relative to presentations of other test stimuli. Generalization gradients shifted towards the overrepresented stimulus value, relative to gradients obtained for equal presentation-frequency conditions. The shift in gradients away from the more frequently presented stimulus during training in Experiment 1 was consistent with accounts of stimulus generalization based on both adaptation-level theory and signal-detection theory. The shift in gradients towards the more frequently presented stimulus during testing in Experiment 2, however, was consistent with adaptation-level theory, whereas the signaldetection account had predicted a shift in the opposite direction. These results attest to the importance of relative stimulus presentation frequency as a determinant of postdiscrimination stimulus generalization performance. Because feedback was given only during training, the opposite effects of presentation probability in training and testing may reflect an interaction between presentation probability and feedback.