This paper aims to evaluate the utility of disentangled representations fo abstract visual reasoning tasks. This is important w.r.t. motivating continued work on learning such representations; the `downstream' utility of these has thus far mostly been taken for granted. This effort aims to put the assumption on steadier footing, via a large-scale empirical analysis over two new tasks that the work introduces. The authors establish that disentanglement empirically leads to sample efficiency, as one would hope. This is an important result to establish because it motivates continued work on learning disentangled representations.