On mental models theories, reasoners create mental representations of information which they manipulate in order toderive new conclusions. These theories have been uniquely successful at explaining a class of attractive fallacies involvingdisjunctions. The original theories have appealed to low-level matching mechanisms (Walsh & Johnson-Laird, 2004;Koralus & Mascarenhas, 2013) to compare the models of the premises and the models of the conclusion and predict ananswer. In three experiments, we show that the check for overlap in content involved in these accounts must take place ata high level of cognition in order to incorporate complex world knowledge. We introduce variants of illusory inferencesfrom disjunction whose acceptance is accurately predicted by independant measures of confidence in causal connections.We conclude that the Revised Mental Model Theory of Khemlani et al. (2018) holds promise, but cannot account for ourdata out of the box.

Studies of human expertise suggest that experts and novices “see“ problems differently. Experts not only acquire a bodyof domain-specific strategies and knowledge, but also learn to quickly identify when those concepts apply to problemswithin the domain. We propose modeling these elements as an iterative process of domain-specific language (DSL)learning, while jointly training a neural network to recognize when learned concepts apply to new problems. We showthat the algorithm solves problems more accurately and quickly than either a neural network alone, or a model that simplyacquires new concepts without learning when to use them. We also examine the implicit problem representations learnedby the neural network recognition model, and find that they increasingly come to reflect abstract relationships betweenproblems, rather than surface features, as the model acquires domain expertise. A full paper and additional details areavailable at: https://sites.google.com/view/neurally-guided-expertise-mit

How do people learn complex rules? We introduce a novel paradigm called ”Track-A-Mole”, in which participants have tolearn about and predict the moves of a cartoon mole, whose movements are generated by graphical programs. Our resultsshow that participants can learn to predict richly structured programs, and often require only few observations to do so,showing rapid learning and early insights about the underlying patterns. Moreover, we found that how learnable a programis can be predicted by features related to its complexity and compressibility. Finally, participants also show interestingpatterns of generalizations, assuming more parsimonious rules first and then gradually adjusting their predictions to morecomplex regularities, as well as matching their predictions to the general direction of movements and producing sensi-ble errors. These results extend our understanding of complex rule learning and open up future opportunities to modelsequential pattern predictions as graphical program induction.