Infants learn better following expectancy violations. Yet it is unknown whether this surprise-induced learning operates across development, is all-or-none or graded, and whether surprise directly mediates it. We addressed these questions by showing adults events depicting varying numbers of violations. In Experiments 1 and 2, adults saw events with 0 to 3 physical violations, then heard a novel verb for the presented action. Adults learned better after observing violations; notably, their learning exhibited a Goldilocks pattern—initially increasing with number of observed violations, then declining. Experiment 3 asked whether this learning enhancement was driven by surprise itself, or by the search for explanations for the surprising events. Adults saw events with different numbers of violations, then rated their surprise and generated candidate explanations. Whereas surprise increased monotonically with violations, explanation-generation exhibited a Goldilocks pattern like that in Experiments 1-2. This suggests that surprise-induced learning may reflect the search for explanations.