Purpose: The aim of this study was to identify variability in word-learning mechanisms used by late-talking children using a longitudinal study design, which may explain variability in late-talking children's outcomes. Method: A cohort of typically developing children (n = 40) and children who were classified as late-talking children at age 2;0 (years;months; [less than or equal to] 10th percentile on expressive vocabulary, n = 21) were followed up at ages 3;0 and 3;6. We tested the cohort across tasks designed to isolate different mechanisms involved in word learning: encoding and producing spoken forms of words (using a nonword repetition task), identifying referents for words (using a fast mapping task), and learning associations between words and referents (using a cross-situational word-learning task). Results: Late-talking children had lower accuracy on nonword repetition than typically developing children, despite most of the sample reaching typical ranges for expressive vocabulary at age 3;6. There were no between-groups differences in fast mapping and retention accuracy; however, both were predicted by concurrent expressive vocabulary. Late-talking children performed less accurately than typically developing children on cross-situational word-learning retention trials, despite showing no between-groups differences during training trials. Combining performance across all three tasks predicted approximately 45% of the variance in vocabulary outcomes at the last time point. Conclusions: Late-talking children continue to have deficits in phonological representation that impact their word-learning ability and expressive language abilities but do not show difficulties in fast mapping novel words. Late-talking children may also struggle to retain associative information about word-referent mappings. Late-talking children thus use some, but not all, word-learning mechanisms differently than typically developing children.