Neural replay is connected to latent cause inference and supports fast generalization

Dynamic environments require generalizing knowledge through two computational mechanisms: Inferring clusters of equivalent observations and propagating reward expectations within a cluster. While neural replay has been linked to both, past work has not studied these processes simultaneously, leaving their relative importance for replay unclear. We used fMRI and computational modeling to investigate latent cause inference and replay in 52 participants during a value-based task with shared rewards across bandits. Participants learned this structure and achieved one-shot generalization after reversals, consistent with a latent cause inference process. fMRI during predecision pauses revealed backwards replay of shared-reward bandits in the visual cortex and medial temporal lobe. Trial-wise fluctuations in visual replay strength and content were better explained by value updates than by structure learning. Conversely, abstract reward structure representations were localized specifically to the MTL. These results demonstrate that on-task replay serves reward propagation within learned latent structures to facilitate fast generalization.