Randomized testing is a popular approach for checking properties of large embedded system designs. It is well known that a uniform random choice of test inputs is often sub-optimal. Ideally, the choice of inputs has to be guided by choosing the right input distributions in order to expose corner-case violations. However, this is also known to be a hard problem, in practice. In this paper, we present an application of the cross-entropy method for adaptively choosing input distributions for falsifying temporal logic properties of hybrid systems. We present various choices for representing input distribution families for the cross-entropy method, ranging from a complete partitioning of the input space into cells to a factored distribution of the input using graphical models. Finally, we experimentally compare the falsification approach using the cross-entropy method to other stochastic and heuristic optimization techniques implemented inside the tool S-Taliro over a set of benchmark systems. The performance of the cross entropy method is quite promising. We find that sampling inputs using the cross-entropy method guided by trace robustness can discover violations faster, and more consistently than the other competing methods considered.