Referring to the Item Response Theory (IRT), there are actually few researches of the relation between the performance and the latent traits or the abilities of mentally retarded students. The prime goal of this study is to introduce IRT and Computerized Adaptive Test (CAT) to special education, hoping to develop a feasible way to evaluate mentally retarded students’ ability with a scientific method. This research will include the analysis of course outlines and assessment constructed out of three-parameter logistic model in the subject of practical mathematics for students in special arranged classes established at vocational schools. By using course outlines, we may accurately predict students’ ability in practical math. With the better understanding of students’ ability in the subject, we can then use it as a basis to construct a useful CAT system on the network. By the powerful calculation of the computer, the CAT system can immediately assigns question to target students by their item response pattern. The pool of questions consider offering the largest amount of information to subject. The results of research are: 1. The established index test model is in accordance with the mathematics ability evaluation for mentally retarded student. This index paper accorded with IRT of three-parameter logistic model. 2. Three-parameter logistic model can accurately evaluate mild mentally retarded student's mathematics ability. It also solves the problem of the classical test which often fails to effectively present mentally retarded students’ ability because it has the defect of giving out small number of questions and having too many people resulted in the same test score. 3. The CAT meets the ‘Individualized’ and ‘Adaptive’ needs of what is requited in special education. By the system, each student who takes the online tests will face different series of questions. In line with student’s individual ability, the amount of questions on the test will be given accordingly. The result of examining can save the examination questions and answering time up to 23.56% in average. Moreover, we also can exactly evaluate specific mentally retarded student's mathematics ability accurately.
