<p>Anhedonia, which is characterized by a loss of pleasure and engagement in rewarding activities, is prevalent across diagnostic groups and is associated with worse clinical and treatment outcomes. Existing evidence-based treatments have only modest effects on improving anhedonia. Reward learning (RL) is an important driver of reward-responsive behavior, which could be leveraged to improve reward sensitivity - a candidate mechanism underlying anhedonia. Individuals (N = 50) with at least moderate anhedonia and clinical levels of depression and/or anxiety were randomized to either an active RL training (high variance of reward rates to promote reward exploration and maximization) or a sham version. Before and after training, they completed an assessment RL task while undergoing functional neuroimaging as well as ratings of current affect. Computational modeling was used to assess change in reward learning and sensitivity, as well as associated neural activity. Participants assigned to the active condition exhibited greater increases in learning-based reward maximization, i.e., the propensity to choose options predicted to be most rewarding, and reductions in anterior cingulate cortex activation to reward prediction errors (discrepancies between expected and observed rewards), consistent with optimization of reward learning and performance. In exploratory analyses, learning-based reward maximization during training mediated increase in positive affect associated with the active condition. Our results underscore the potential of RL-informed computerized trainings to improve RL and associated goal-directed reward behavior in anhedonic individuals. Future research is needed to determine whether the proposed training can successfully transfer to real-world settings and produce long-term affective and clinical improvements (<b>ClinicalTrials.gov</b>: NCT05383248).</p>