Complex decisions require an adequate model of the world that needs to be updated in response to changes in the environment. In the first part of the talk, I will present data illustrating how humans form models of other people’s preferences and integrate this information into an own mPFC value computation. By combining computational and representational fMRI techniques I show that prediction errors caused by learning about the preferences of another individual drive changes in local cortical representations, resulting in a change in subjects´ own preference. In the second part of the talk, I demonstrate that a newly learnt model of the statistical relationships between objects is stored as an abstract relational map in the entorhinal cortex. This suggests that the hippocampal-entorhinal system creates metric representations based on associative strength when relationships are non-spatial rather than spatial, discrete rather than continuous and unavailable to conscious awareness. The computation of such relationships may be directly embedded in entorhinal grid cell representations, such that inferences do not need to rely on experiences but can be computed rapidly from mapped knowledge.