We studied transposition in pigeons, using multiple-pair discrimination training. Four birds discriminated two pairs of circles: 1+ 2- and 5+ 6- or 1 - 2 + and 5- 6 + (digits denote circle diameters and plus and minus signs denote reward and nonreward, respectively). Four other birds discriminated four pairs of circles: 1+ 2-, 1+ 3-, 4+ 6-, and 5+ 6- or 1- 2 +, 1- 3 +, 4- 6+, and 5- 6+. Finally, 4 birds discriminated only one pair of circles: 1+ 2-, 1- 2+, 5+ 6-, or 5- 6+. Testing included five new pairs--1/5, 2/3,2/6,3/4, and 4/5--that distinguished absolute from relational accounts of transposition. The pigeons' relational responding rose from one- to two- to four-pair training. The similarity of the testing stimuli to one another also affected relational responding: Transposition increased with highly dissimilar stimuli. Neither Spence's (1937) theory nor existing relational accounts could predict the obtained pattern of relational responding.