The learning process and its outcomes depend greatly on the social interaction between teachers and students and, in particular, on the proficient and focused use of language through written text or discussions. Our overarching goal in this project is to better understand how to make automated tutorial dialogues effective and adaptive to student characteristics, such as prior knowledge. The specific goal of our current project is to develop an adaptive, natural-language tutoring system, driven by a student model, which can effectively carry out reflective conversations with students after they solve physics problems. Towards this end, we continue our work in identifying linguistic features of tutoring that predict learning gains, and extend it by characterizing the “level of support” to provide to students based on their current level of understanding particular physics concepts and principles, as dynamically captured by the student model. In this talk, I will describe the features of dialogic discourse underlying “level of support” that we have identified through the analysis of human-to-human tutorial dialogues as well as the construction and application of a coding scheme for the characterization of the “level of support”. I will present initial teacher feedback on dialogues that apply these features to coach students at different levels. I will also discuss how this line of research affects the authoring of tutorial dialogues used by an intelligent tutoring system for students who exhibit different levels of understanding.