Asking good questions is vital for scientific learning and discovery, but improving this complex skill is a formidable challenge. Here, we show in two experiments (N = 152) that teaching others--"learning-by-teaching"--enhances one's ability to generate higher-order research questions that create new knowledge, relative to two other well-established generative learning techniques: retrieval practice and concept-mapping. Learners who taught scientific expository texts across natural and social sciences topics by delivering video-recorded lectures outperformed their peers who practiced retrieval or constructed concept maps when tested on their ability to generate "create"-level research questions based on the texts (Experiment 1). This advantage held reliably even on a delayed test 48 hr later, and when all learners similarly received and responded to post-study questions on the material (Experiment 2). Moreover, across both immediate and delayed tests, learning-by-teaching produced a recall benefit that rivaled that of the potent technique of retrieval practice. In contrast, despite recalling more than twice the study content that the concept-mapping group did, learners who practiced retrieval were unable to generate more "create"-level research questions based on that content. Three supplemental experiments (N = 168) further showed that retrieval practice consistently did not improve higher-order question generation over restudying, despite yielding superior long-term retention. Altogether, these findings reveal that simply possessing a wealth of factual knowledge is insufficient for generating higher-order research questions that create new knowledge. Rather, teaching others is a powerful strategy for producing deep and durable learning that enables research question generation. To ask better questions, teach.