Learning shapes neural geometry in the prefrontal cortex

The relationship between the geometry of neural representations and the task being performed is a central question in neuroscience ^1–6 . The primate prefrontal cortex (PFC) is a primary focus of inquiry, as it can encode information with geometries that either rely on past experience ^7–13 or are experience agnostic ^3,14–16 . One hypothesis is that PFC representations should evolve with learning ^4,17,18 , from a format that supports exploration of all possible task rules to a format that minimises the encoding of task-irrelevant features ^4,17,18 and supports generalisation ^7,8 . Here we test this idea by recording neural activity from PFC when learning a new rule (‘XOR rule’) from scratch. We show that PFC representations progress from being high dimensional, nonlinear and randomly mixed to low dimensional and rule selective. Upon generalising the rule to novel stimuli, these representations further evolve into an abstract, stimulus-invariant geometry. These findings reconcile previously conflicting accounts of PFC function by demonstrating how neural representations adapt across distinct stages of learning.