ABSTRACT Individuals with Substance Use Disorder (SUD) often exhibit increased stochasticity in their choice behavior, indicating an inability to adopt a consistent internal model and/or an appropriate rate of learning. Tackling uncertainty related to model adoption requires the brain to estimate the reliability of different internal models and arbitrate between these models. These processes, however, are not well understood mainly because in most experimental paradigms only a single type of reward contingency is tested such that either stimuli or actions predicts reward, not both. Additionally, the rate of learning is affected by other types of uncertainty in the environment: (1) expected uncertainty due to the probabilistic nature of reward outcomes, and (2) unexpected uncertainty due to actual changes in the reward environment. Interestingly, subregions of prefrontal cortex (PFC) are involved in both stimulus- and action-based learning, and in learning under uncertainty. For example, both orbitofrontal cortex (OFC) and anterior cingulate cortex (ACC) seem to be similarly engaged in learning under uncertainty while their activity reflects stimulus and action values, respectively. Although ACC and OFC subregions of PFC are both involved in adaptive behavior, they may serve different functions. Densely interconnected with both OFC and ACC, the basolateral amygdala (BLA) is also involved in adjusting behavior when the environment changes. However, it is unclear how BLA and PFC subregions interact to control these behavioral adjustments. Based on our experimental and computational work in the prior funding period, we hypothesize that BLA provides information to different subregions of PFC to adjust learning under different forms of uncertainty. Our central hypothesis is that BLA contributes to the arbitration between internal models (stimulus vs. action-based) and between different forms of uncertainty (expected vs. unexpected uncertainty) through its connections to OFC and ACC. Using a combination of computational methods, pathway-specific chemogenetics, in vivo 1-photon (1P) single-cell calcium imaging, and closed-loop manipulation, we will test how adopting an appropriate internal model (Aim 1) and adjustments to changes in the environment (Aim 2) rely on BLA-to-OFC (BLAOFC) vs. BLA-to-ACC (BLAACC) pathways. We further causally test these circuits by closed-loop control of learning based on both reward-expectation signals in ACC and OFC at the neuronal ensemble level, which we can decode in real time, and behaviorally using our novel information- theoretic metrics and postural measures (Aim 3).