Representation learning facilitates different levels of generalization

: Cognitive maps represent relational structures and are taken to be important for generalization and optimal decision making in spatial as well as non-spatial domains. While many studies have investigated the benefits of cognitive maps, how these maps are learned from experience has remained less clear. We introduce a new graph-structured sequence task to better understand how cognitive maps are learned. Participants observed sequences of episodes followed by a reward, thereby learning about the underlying transition structure and fluctuating reward contingencies. Importantly, the task structure allowed participants to generalize value from some episode sequences to others, and generalizability was either signaled by episode similarity or had to be inferred more indirectly. Behavioral data demonstrated participants ` ability to learn about signaled and unsignaled generalizability with different speed, indicating that the formation of cognitive maps partially relies on exploiting observable similarities across episodes. We hypothesize that a possible neural mechanism involved in learning cognitive maps as described here is experience replay.