Mental simulation is a powerful cognitive capacity that underlies people's ability to draw inferences about what happened in the past from the present. Recent work suggests that eyetracking can be used as a window through which one can study the process of mental simulation in intuitive physics tasks. In our experiment, participants have to figure out in which of three holes a ball was dropped in a virtual Plinko box. We develop a computational model of human intuitive physical reasoning in Plinko that runs repeated simulations in a noisy physics simulator in order to infer in which hole the ball was dropped. We evaluate our model's behavior against multiple human data signals: trial judgments, response times, and eye-movement data. We find that a model that sequentially samples simulations while balancing uncertainty and reward best explains the patterns of participant behavior we observe in these three signals.

The words we use to describe what happened shape the story a listener imagines. How do speakers choose what causal expression to use? How does that impact what listeners infer about what happened? In this paper, we develop a computational model of how people use the causal expressions "caused", "enabled", "affected", and "made no difference". The model first builds a causal representation of what happened. By running counterfactual simulations, the model computes causal aspects that capture the different ways in which a candidate cause made a difference to the outcome. Logical combinations of these aspects define a semantics for the different causal expressions. The model then uses pragmatic inference favoring informative utterances to decide what word to use in context. We test our model in a series of experiments. In a set of psycholinguistic studies, we verify semantic and pragmatic assumptions of our model. We show that the causal expressions exist on a hierarchy of informativeness, and that participants draw informative pragmatic inferences in line with this scale. In the next two studies, we demonstrate that our model quantitatively fits participant behavior in a speaker task and a listener task involving dynamic physical scenarios. We compare our model to two lesioned alternatives, one which removes the pragmatic inference component, and another which additionally removes the semantics of the causal expressions. Our full model better accounts for participants' behavior than both alternatives, suggesting that causal knowledge, semantics, and pragmatics are all important for understanding how people produce and comprehend causal language.