Humans can quickly learn and update latent task structures, and use them to guide value-based decision-making. In a functional magnetic resonance imaging study, 52 participants learned a latent graph structure requiring non-local value generalization upon reward reversals. Performance was best explained by a latent cause inference model that captures structure learning and admits value generalization. The functional imaging data offered a possible substrate for this generalization by demonstrating that the hippocampus tracked the underlying task structure and exhibited non-local reactivation of unobserved sequences sharing a reward.
Renz et al. (Fri,) studied this question.