One of the most important goals and tasks of many E-learning systems is adaptivity, i.e., to automatically create individual learning environments taking into account a learner's special needs, interests, goals, and preferences, be it in a corporate, government, healthcare, or higher education environment. An adaptive E-learning system must be able to capture information about the learner, to interpret this information and thus to determine which content will be presented. On the one hand, this requires a lot of skills from the field of user modeling. On the other hand, an appropriate learning objects repository is required. What kind of learning objects is required to enable a system's adaptivity? Which types of learning objects can be distinguished, and why? This will be discussed in this work, and a new architecture enabling different levels of adaptivity by the use of multidimensional, atomic learning objects is presented and discussed.