For some time, it has been clear that students who are tutored generally leam more than students who experience classroom instruction (e.g.. Bloom, 1984). Much research has been devoted to identifying features of tutorial dialogue that can explain its effectiveness, so that these features can be simulated in natural-language tutoring systems. One hypothesis is that the highly interactive nature of tutoring itself promotes leaming-that is, the interaction hypothesis. Although reasonable and agreeing with much research, the interaction hypothesis raises the question of what linguistic mechanisms are involved: that is, which features of "highly interactive" dialogues trigger what processes that are conducive to learning? Our overall strategy in the research described in this article was to inform this question by identifying co-constmcted discourse relations in tutorial dialogues whose frequency of occurrence predicts leaming, identify the context in which these relations occur, and use this knowledge to formulate decision rules to guide automated dialogues. We used Rhetorical Structure Theory to identify and tag co-constmcted discourse relations in a large corpus of physics tutoring dialogues. Our analyses suggest that the effectiveness of human tutoring might well lie in the language of tutoring itself Moreover, the types of co-constructed discourse relations that predict leaming seem to vary based on students' ability level. We describe Rimac, a natural-language tutoring system that implements an initial set of decision rules based on these analyses. These rules guide reflective dialogues about the concepts associated with physics problems. Rimac is being pilot tested in high school physics classes.Educators and policy makers in the United States have looked to educational technology as a tool to increase students' proficiency in math, science, reading, and other subject matter domains. For example, early in his administration. President Obama (2009) challenged developers of intelligent tutoring systems (ITSs) to develop "leaming software as effective as a personal tutor" (para. 19). Apparently, Obama cast this challenge a bit too late. A recent meta-analysis of research comparing the effectiveness of human tutors with state-of-the-art ITSs showed that ITSs have already nearly caught up with human tutors (VanLehn, 2011), with effect sizes {d) of 0.76 for human tutoring and 0.79 for ITSs relative to