To successfully navigate an uncertain world, one has to learn the relationship between cues (e.g., wind speed, atmospheric pressure) and outcomes (e.g., rain). When learning, it is possible to actively manipulate the cue values to test hypotheses about this relationship directly. Across two studies, we investigated how 5- to 7-year-olds actively learned cue-outcome relationships, and what their behavior revealed about how they represented the hypothesis space. Children learned how two cues (color and shape) predicted some monsters' relative speed, by selecting which monster pairs to see racing. We compared two computational models in their ability to capture children's behavior: a "cue-abstraction" model, which organizes the hypothesis space based on abstracted cue-outcome relationships, and a permutation-based model, which represents the hypothesis space based on the relative speed of individual monsters. The results of Study 1 (26 five-year-olds, 14 female and 25 six-year-olds, 15 female; predominantly White, fluent in English) provided the first evidence that 5- and 6-year-olds can use cue-abstraction hypothesis space representations when provided with scaffolding. However, Study 2 (65 five-year-olds, 33 female; 67 six-year-olds, 33 female; 68 seven-year-olds, 33 female; predominantly White, fluent in German) showed that young children were best described by the "permutation-based" model, and that only 7-year-olds, when provided with memory aids, were best captured by the cue-abstraction model. Overall, our results highlight the guiding role of the hypothesis space for active search and learning, suggesting that these two phases might trigger different representations, and indicating a developmental shift in how children represent the hypothesis space.