A behavior-analytic model of transitive inference (TI) as relational reasoning with derived comparative relations is outlined. Following nonarbitrary relational training and testing to establish contextual functions of " more than" (>) and " less than" (<) for two abstract stimuli, two groups of participants learned a series of contextually controlled more than or less than relations (All-More: E > D > C > B > A; All-Less: A < B < C < D < E). On meeting the training criterion, inferential tests were presented to both groups involving mutually entailed relations (All-More: A < B, B < C, C < D and D < E; All-Less: B > A, C > B, D > C and E > D) and one-step (A < C, B < D, C < E, C > A, D > B and E > C) and two-step (A < D, B < E, D > A and E > B) combinatorially entailed relations. Performance accuracy on the trained and inferential tasks was uniformly high across both groups, with no significant differences observed. In both groups, however, performance accuracy differed significantly on one-step and two-step combinatorially entailed tasks involving the same or different relation to that trained. The present findings demonstrate complex relational reasoning with derived comparative relations, replicate several key effects from the literature on TI and have potential implications for the development of a contemporary behavior-analytic account of TI.