This study examined the effects of training involving spatial versus nonspatial representations of numerical magnitude for promoting arithmetic fluency. The key goal was to advance theoretical understanding of the relation between spatial and math learning, while simultaneously laying the groundwork for the development of future educational interventions. Toward this goal, the study tested the hypothesis that the use of spatial representations during training facilitates arithmetic fluency via improvements in numerical magnitude knowledge. Participants (first graders from low-income racially/ethnically diverse backgrounds, N = 205) were randomly assigned to one of four experimental conditions: spatial-continuous, spatial-discrete, nonspatial-verbal cues, and nonspatial-no verbal cues. All conditions involved eight 30-min training sessions, in which children received instruction on addition/subtraction problems with totals within 10. The key difference between conditions was the type of materials utilized during training--specifically, the type of magnitude cues they contained. The results showed that children's arithmetic skills increased from pre- to posttest in all conditions and that the increase was significantly larger in the spatial, compared to nonspatial, conditions. However, there was no effect of condition on numerical magnitude knowledge, which leaves open the question about the underlying mechanism. The findings demonstrating a causal relation between spatial and mathematical domains have both theoretical significance and practical implications for the choice of instructional materials. [This is the online first version of an article published in "Developmental Psychology" (ISSN 0012-1649).]