This study investigated the role of generative strategies and working memory capacity on word problem solving accuracy in children with math difficulties (MD). Within classrooms, children in Grade 3 with MD ("n" = 69) were randomly assigned to one of three treatment conditions: paraphrase question propositions (Restate), paraphrase relevant propositions (Relevant), and paraphrase all propositions (Complete), or to an untreated control. An additional control group included children without MD ("n" = 22). Mixed regression modeling showed that generative strategies significantly improved posttest scores for children with MD compared with the control condition, but outcomes were related to the type of dependent measures. The Relevant and Complete treatment conditions improved problem-solving accuracy, the Complete condition improved problem component identification, and the Restate and Relevant conditions improved operation span performance when compared with the control conditions. Only the Relevant and Complete generative learning treatments allowed children with MD to catch up to children without MD, but the results were moderated by working memory capacity.