Previous research suggests that 3-year-olds fail to learn from statistical data when their prior beliefs conflict with evidence. Are children's beliefs entrenched in their folk theories, or can preschoolers rationally update their beliefs? Motivated by a Bayesian account, we conducted a training study to investigate this question. Children (45 months) who failed to endorse a statistically more probable (but a-priori unlikely) cause following ambiguous evidence were assigned to a "Statistical Reasoning" training, one of two "Prior Belief" trainings ("Base Rates", "Mechanisms"), or a "Control" condition. Relative to the Control, children in the trainings were more likely to endorse the a-priori unlikely variable on a free-explanation task. Critically, children in the "Statistical Reasoning" condition passed this task, even though their only information about the belief-violating variable came from ambiguous evidence. This suggests that statistical reasoning training improves preschoolers' ability to learn even from data inconsistent with their prior beliefs. (Contains 2 figures.)