A key goal in educational and cognitive science is to understand how abstract concepts are best learned. In the current work, we present a low-support, interactive discovery interface for learning complex relational categories. The platform, which allows learners to make modifications to exemplars and see the corresponding effects on category membership, holds the potential to augment relational learning by facilitating self-directed comparisons that explore what the learner does not yet understand. We compared interactive training to identification training and assessed learners on their ability to generalize to novel exemplars from the learning domain. Although identification learners were provided with many more examples of the category during training, interactive learners demonstrated enhanced generalization accuracy and more specific knowledge of category membership constraints. Overall, the results show interactive training to be a powerful tool for supplementing and refining category knowledge. We conclude with implications of these findings and promising future directions.