This paper examines the nature of stimulus representation in associative learning systems. Specifically, it addresses the issue of whether representation is elemental or configural in nature. We use a human causal learning paradigm, employing contingencies more commonly associated with studies of retrospective revaluation. Whereas most models of retrospective revaluation view it as an entirely elemental process, our results show that it has a configural component. However, the results also prove troublesome for simple configural theories employing fixed generalisation coefficients. It is possible to explain the data using an elemental theory employing configural representation. Our favoured explanation, however, involves a configural theory employing adaptive generalisation. We present such a theory, APECS, and show through simulation that it is well-equipped to deal with our findings.