Of all teaching methods, one-to-one human tutoring is the most powerful method for promoting learning. To achieve this aim and reduce teaching load, researchers developed intelligent tutoring systems (ITSs) to employ one-to-one tutoring (Aleven, McLaren, & Sewall, 2009; Aleven, McLaren, Sewall, & Koedinger, 2009; Anderson, Corbett, Koedinger, & Pelletier, 1995; Anderson & Reiser, 1985; Blessing, Gilbert, Ourada, & Ritter, 2009; Mitrovic et al., 2009; Mitrovic & Ohlsson, 1999; Suraweera, Mitrovic, & Martin, 2007; VanLehn et al., 2005). However, most ITSs have restricted user interfaces, which confine reasoning strategies of students during problem solving, thus ignoring the fact that students could use dissimilar strategies to solve a given question. Furthermore, student learning problems could be diagnosed from the derivation of their answers. In order to interpret students' mathematical problem-solving behaviors, this study developed a Model-tracing Intelligent Tutor (MIT) to diagnose students' learning problems and provide learning feedback for individual students. A quasi-experiment was conducted in an elementary school to evaluate the effectiveness of the proposed approach, in which 124 fifth graders participated. The experimental results show that the model-tracing-based learning diagnosis system is significantly more helpful to the students in learning mathematics than the conventional web-based test in terms of learning achievements.