Providing feedback, including hints, is one of the key steps in the tutoring process. However, a persistent challenge in the development of intelligent tutoring systems (ITSs) is how to determine accurately when a student needs help, and then determine what the best help is for that individual student. This study investigated the feasibility of predicting students' need for help in an ITS using Item Response Theory (IRT). It involved analysis of data from the PACT Geometry Tutor and a randomized study that compared three versions of a tutoring system, including one that used IRT. The analysis of extant data showed that the use of hints was related to the students' beginning ability and to the size of the gap between that initial ability and the difficulty of the item. In the comparative study, I worked with a staff from the Principled Assessment Design for Inquiry project to develop three versions of a self-assessment system used with the Full Option Science System curriculum on Force and Motion. The system helped middleschool students to learn to select and use equations when solving physics problems about speed. The full version of the tutor used IRT to give students hints appropriate to the size of their learning gap. The second version of the tutor provided feedback on errors made, but gave no hints on how to repair those errors. The third version of the tutor gave neither error feedback nor hints. I found statistically significant differences between the mean learning gains of both the Full Tutor and the Feedback Only groups when compared to the No Help group. There was a moderate effect size when the mean learning gains of the Full Tutor and Feedback Only groups were compared to the gains of the group with No Help. The study also showed that some assumptions made in the design of the intelligent hints system about the stability of item difficulty across occasions and across item isomorphs generated from item shells in the system were incorrect. Future improvements and possible research are discussed.