Stress reduces both model-based and model-free neural computations during flexible learning

Stressful events are thought to impair the flexible adaptation to changing environments, yet the underlying mechanisms are largely unknown. Here, we combined computational modeling and functional magnetic resonance imaging (fMRI) to elucidate the neurocomputational mechanisms underlying stress-induced deficits in flexible learning. Healthy participants underwent a stress or control manipulation before they completed, in the MRI scanner, a Markov decision task, frequently used to dissociate model-based and model-free contributions to choice, with repeated reversals of reward contingencies. Our results showed that stress attenuated the behavioral sensitivity to reversals in reward contingencies. Computational modeling further indicated that stress specifically affected the use of value computations for subsequent action selection. This reduced application of learned information on subsequent behavior was paralleled by a stress-induced reduction in inferolateral prefrontal cortex activity during model-free computations. For model-based learning, stress decreased specifically posterior, but not anterior, hippocampal activity, pointing to a functional segregation of model-based processing and its modulation by stress along the hippocampal longitudinal axis. Our findings shed light on the mechanisms underlying deficits in flexible learning under stress and indicate that, in highly dynamic environments, stress may hamper both model-based and model-free contributions to adaptive behavior.