Learning requires effort, but children cannot try hard at everything. Here, we evaluated whether children use their improvement over time to decide whether to stick with a challenge. To eliminate the effect of individual differences in ability or prior knowledge, we created a novel paradigm that allowed us to surreptitiously control children's performance. Across three preregistered experiments (N = 319, ages 4 to 6 in the United States), we found that children who were given evidence that their performance was improving were more likely to persist on a challenging task than children who were given evidence that their performance was constant, even when final performance was matched. This effect was robust to differing reward contingencies, across in-person and online testing contexts, and was driven by the demotivating effect of constant performance. Our results suggest that young children will be more persistent if they are guided away from too-difficult tasks and toward opportunities that enable steady growth.