If pigeons are presented with a compound sample stimulus consisting of values from two dimensions (e.g., vertical white lines on a red field) and then are tested with comparison stimuli that vary along only one dimension (e.g., red field vs. green field), accuracy of delayed matching has been found to be lower than in the case where the sample stimulus consisted of only a single element (i.e., red field). Two hypotheses have been advanced to account for this effect - - the shared-attention hypothesis and the generalization-decrement hypothesis. The adequacy of these hypotheses was tested in three experiments, in which a number of paradigms were tested varying the information presented on sample and comparison stimuli. Evidence for an elicitation or autoshaping process was revealed in these experiments by the findings that (a) performance was lower with compound comparison stimuli than with element comparison stimuli and (b) pigeons showed a preference for a compound comparison stimulus over an element comparison stimulus. A two-process model of delayed matching is offered that includes an elicitation process and a memory process. Within the framework of this model, it was found that both shared-attention and stimulus-generalization mechanisms adequately account for the findings of these experiments.