One powerful way children can learn word meanings is via cross-situational learning, the ability to discern consistent word-referent mappings from a series of ambiguous scenes and utterances. Various computational accounts of word learning have been proposed, with mechanisms ranging from storing and testing a single hypothesized referent for each word, to tracking multiple graded associations and selectively strengthening some of them. Nearly all word learning models assume storage of some feasible word-referent mappings from each situation, resulting in a degree of learning proportional to the number of co-occurrences. While these accumulative models would generally predict that incorrect co-occurrences would slow learning, recent empirical work suggests these accounts are incomplete: paradoxically, giving learners incorrect mappings early in training was found to boost performance (Fitneva and Christiansen, 2015). We test this finding’s generality in a new experiment with more items, consider system- and item-level explanations, and find that a model with error-driven learning best accounts for this benefit of initially-inaccurate pairings.