In recent years, significant advances have been made in intelligent tutoring systems, and these advances hold great promise for adaptively supporting computer science (CS) learning. In particular, tutorial dialogue systems that engage students in natural language dialogue can create rich, adaptive interactions. A promising approach to increasing the effectiveness of these systems is to adapt not only to problem-solving performance, but also to a student's characteristics. Self-efficacy refers to a student's view of her ability to complete learning objectives and to achieve goals; this characteristic may be particularly influential during tutorial dialogue for computer science education. This article examines a corpus of effective human tutoring for computer science to discover the extent to which considering self-efficacy as measured within a pre-survey, coupled with dialogue and task events during tutoring, improves models that predict the student's self-reported frustration and learning gains after tutoring. The analysis reveals that students with high and low self-efficacy benefit differently from tutorial dialogue. Student control, social dialogue, and tutor moves to increase efficiency, may be particularly helpful for high self-efficacy students, while for low self-efficacy students, guided experimentation may foster greater learning while at the same time potentially increasing frustration. It is hoped that this line of research will enable tutoring systems for computer science to tailor their tutorial interactions more effectively.