This randomized control treatment study investigated the role of generative paraphrasing interventions on word problem-solving (WPS) accuracy in English language learners (ELL) who were at risk for math difficulties (MD). One hundred and forty-two third-grade monolingual and English language learners were randomly assigned to either an untreated control group or one of three treatment conditions: paraphrase question ("Restate"), paraphrase relevant information ("Relevant"), and paraphrase all propositions ("Complete"). Results showed that the "relevant" treatment condition improved WPS accuracy for monolingual and ELL children's accuracy relative to the control and restate condition. There were, however, no positive advantages on the transfer measures (knowledge of problem-solving components, calculation) for ELL children with MD. In addition, the magnitude of posttest outcomes on the problem-solving measures was substantially smaller for ELL children with MD than for children without MD. Although the results overall support the notion that generative learning improves problem-solving accuracy, ELL children without MD are most likely to benefit from its application.