13In novel situations, behavior necessarily reduces to latent biases. How these biases interact 14with new experiences to enable subsequent behavior remains poorly understood. We 15 exposed rats to a family of spatial alternation contingencies and developed a series of 16 reinforcement learning agents to describe the behavior. The performance of these agents 17shows that accurately describing the learning of individual animals requires accounting for 18 their individual dynamic preferences as well as general, shared, cognitive processes. Agents 19 that include only memory of past choice do not account for the behavior. Adding an explicit 20 representation of biases allows agents to perform the task as rapidly as the rats, to accurately 21 predict critical facets of their behavior on which it was not fitted, and to capture individual 22 differences quantitatively. Our results illustrate the value of making explicit models of 23 learning and highlight the importance of considering the initial state of each animal in 24 understanding behavior. 25