Deep neural networks have achieved remarkable advancements in various areas, such as image processing [18, 6] and natural language processing [16, 21].

While standard recurrent neural networks explicitly impose achain structure on different forms of data, they do not have an explicit bias towards recursive selfinstantiation where the extent of recursion is dynamic.

The Gumbel-Max trick is the basis of many relaxed gradient estimators. These estimators areeasy toimplement and lowvariance, butthegoal ofscaling them comprehensively to large combinatorial distributions is still outstanding. Working within the perturbation model framework, we introduce stochastic softmax tricks, which generalizetheGumbel-Softmax tricktocombinatorial spaces.