Whether animals use relational cues in transposition tests has long been considered a controversial issue. In the present study, we examined whether common marmosets could generalize relational responses to untrained stimulus pairs and further apply these generalizations to unknown shapes. The subjects were trained to perform simple discrimination tasks using a pair of stimuli. The stimuli differed in size, and the subjects were required to select the larger or smaller of the 2 sizes, depending on the given contingencies. After experiencing several reversals, the subjects were examined using 2 different tests: transposition and shape generalization. In the transposition test trials, in which squares of different sizes than those used in the training trials were presented, the subjects selected the stimulus based on the relative size of the stimulus. In the shape generalization tests, sets of 5 novel shapes with the same relative sizes were presented with the training stimuli. The subjects' performance indicated successful transposition to the novel stimulus pairs, and further analysis showed that transposition was more likely to occur when the test stimuli shared physical features, such as the outer length and the number of line segments, with the trained stimuli. Thus, the present study demonstrated the robust ability of transposition in common marmosets based on relative size, both with and without common shape features, and offered a possible method for specifying the critical stimulus features through which transposition can be more readily observed.