I investigated the neural mechanisms for learning in dynamic environments. In dynamic environments, where the underlying state of the environment can change, inferring the current state of the environment is important to guide adaptive behavior. People should maintain their beliefs about the environmental state when it is stable and they should quickly update their beliefs when the environmental state changes. Belief updating can be guided by prediction errors, which are the difference between the expected observation and the actual observation. How people use prediction errors to update their beliefs is determined by their learning rate. Learning rates can be influenced by belief surprise and belief uncertainty. Belief surprise reflects how unlikely an observation is, given the person’s belief about the current state. Belief uncertainty reflects how imprecise a person’s belief is about the current state. In three studies, I investigated how the brain detects state changes and guides subsequent behavioral adaptation. In the first study, I examine the roles of physiological arousal during learning in two kinds of dynamic environments. In both environments, prediction errors enhanced learning rates and induced pupil dilation. Among different measures of physiological arousal (pupil dilation, skin conductance, heart rate and respiration rate), only pupil dilation consistently predicted trial-by-trial learning rates in both environments. Furthermore, pupil dilation mediated the relationship between prediction errors and learning rates and predicted variance in learning rates that could not be accounted for by prediction errors. In the second study, I investigated how whole-brain functional networks reconfigure for the adjustment of learning. Learning rates were influenced by belief surprise and belief uncertainty, and these two variables also modulated the integration between fronto-parietal and other brain networks. This modulation of functional networks was also associated with behavioral adaptation across individuals. In the third study, I further distinguished the functional roles of frontal and parietal regions during learning. Using multi-voxel pattern classification, I showed that posterior parietal cortex encoded prediction errors in a task-dependent manner while frontal cortex predicted the subsequent behavioral shifts in response to errors. From these studies, I demonstrated how different neural systems contribute to learning in dynamic environments.