Researchers analyzed functional MRI data from people who learned the relative rankings of 16 faces on competence and popularity. Participants later had to infer rankings they had not been directly taught, switching between the two dimensions according to a context cue.

A medial temporal and orbitofrontal network represented both dimensions together in a stable map-like pattern. A frontoparietal network instead formed separate, context-dependent representations that emphasized the relevant dimension. Recurrent neural networks trained on the same task developed similar patterns, with the broad map-like representation appearing earlier in training.