Research in the Learning & Decision Neuroscience Lab addresses how intelligent systems construct and deploy internal models of the world, with a focus on the roles of intrinsic motivation and compositional generalization. Our theoretical perspective is that high-level animal cognition is underpinned by relational knowledge representations, and that relational constructs that promote adaptive behavior, such as causality and controllability, have intrinsic incentive salience. Our approach is multidisciplinary, drawing on a wide range of methods from psychology, neuroscience, statistics, and computer science. In particular, we combine innovative behavioral experiments with computational cognitive models, neuroimaging, and (sub)clinical assays, in order to characterize the nature of valenced and structured mental representations.