People need to generate and test hypotheses in order to create accurate representations of their environments. But how do they know which hypotheses to consider when there are often infinitely many possibilities? Here we explore the idea that evolutionary mental representation generation and selection processes – responsible for the generation of both local (i.e., within a task) and global (i.e., about a task) representations – enable people to address this problem. We investigated this through an active learning experiment, where participants’ task was to discover a hidden rule determining the behavior of a simple physical system. Specifically, we aimed to manipulate factors that constrain this process, particularly through experimental instructions and feedback. We found that providing more opportunities for participants to recognize when their initial task conceptualization was wrong and adjust it helped them create more accurate representations about the task, which in turn led to better accuracy within the task.