Project Summary Reinforcement learning (RL) is a powerful framework for understanding outcome-driven behavior and its neural basis. A key issue for RL is how to incorporate an ability to make inferences that allow appropriate responding to novel conditions not encountered during training. Inference relies on the detection of latent relationships that predict rewards but are not signaled by explicit cues. In the previous funding period, we developed behavioral and computational approaches to rigorously address this problem, and made single-cell neurophysiological observations that began to reveal how the brain represents latent structures. Here, we re-conceptualize this problem as one of using a cognitive map to represent latent order and support inference. We propose to test this idea with a newly developed behavioral approach, which allows unprecedented insight into the role of cortico-striatal neural circuits and ascending neuromodulatory systems in model-based RL. These investigations will focus on brain regions that neurophysiology, neuroimaging, and lesion studies suggest have important roles in implicit serial learning, specifically dorsolateral (dlPFC) and ventromedial prefrontal cortex (vmPFC), and dorsal striatum. Three aims will 1. Test NHPs ability to make model-based inferences in the presence of countervailing reward incentives, 2. Identify neural circuitry of serial learning in the dlPFC, vmPFC, and caudate nucleus, and 3. Test if changes in Ach and DA concentrations are correlated with serial learning, particularly during transfer. Inferential reasoning is impaired in many psychiatric illnesses including schizophrenia and bipolar disorder. This impairment may underlie thought disorders such as delusions and paranoia.