Notes on Iterated Rationality Models of Scalar Implicatures

In the linguistics literature, the derivation of scalar implicatures has often been handled in a relatively modular way, using computations that are sensitive to logical relations among alternatives such as entailment but are blind to other notions such as the probabilities that participants in a conversation might associate with these alternatives (or with related propositions). In recent years, a family of models that we refer to as iterated rationality models (IRMs) have offered an interestingly different perspective on such alternative-sensitive processes in terms of multiple iterations of (typically probabilistic) reasoning. Our paper investigates what at first sight seems like a very interesting argument for the adequacy of IRMs as models of scalar implicatures, coming from the conjunctive interpretation of disjunctive sentences. We then outline challenges for the argument based on a theoretical comparison with the grammatical theory of scalar implicatures. The comparison focuses on the full distribution of conjunctive interpretation, on the question of sensitivity to priors, and on other results achieved within the grammatical theory that the IRM literature has yet to engage with.