A closer alignment of mutual expectations between technical systems and their users regarding functionality and interactions is supposed to improve their overall performance. In general, such an alignment is realized by automatically adapting the appearance and the behavior of a system. Adaptation may be based on parameters regarding the task to be fulfilled, the surrounding context, or the user himself. Among the latter, current emphasis of research is shifting from a user’s trails in the system (for instance, to derive his level of expertise) towards transient aspects (like his current mental or emotional state). For educational technology, in particular, adapting the presented information and the tasks to be solved to the current personal needs of a learner promises a higher motivation and thus a better learning outcome. Tasks which are equally challenging and motivating can keep the users in a state of flow and thus foster enduring engagement. This is of certain importance for difficult topics and/or learners with disabilities. The chapter explains the complex cause-and-effect models behind adaptive training systems, the mechanisms that can be facilitated to implement them, and empirical results from a clinical study. We exemplify this for the training of emotion recognition by people with autism, but not limited to this user group. For this purpose, we present two approaches. One is to extent the Elo algorithm regarding dimensions of difficulty in social cognition. This allows not only to judge the difficulty of tasks and the skills of users, but also to freely generate well-suited tasks. The second approach is to make use of socio-emotional signals of the learners in order to further adapt the training system. We discuss current possibilities and remaining challenges for these approaches.