The UML shares with reflective architectures the idea that self-definition of languages and systems is a key principle for building and maintaining complex systems. The UML is now defined by a four-layer metalevel structure, enabling a flexible and extensible definition of models by metamodels, and even a self-description of the meta-metamodel (the MOF). This metalevel dimension of UMLis currently restricted to structural reflection. But recently a new extension to the UML, called the Action Semantics (AS), has been proposed for standardization to the OMG. This paper explores how this proposed extension brings a behavioural reflection dimension to the UML. Indeed, we show that it is not only possible but quite effective to use the AS for manipulating UMLmo dels (including the AS metamodel). Besides elegant conceptual achievements, such as a metacircular definition of the AS, reflective modeling with the AS leverages on the UML metalevel architecture to provide the benefits of a reflective approach, in terms of separation of concerns, within a mainstream industrial context. A complete model can now be built as an ideal model representing the core concepts in the application, to which non-functional requirements are integrated as fully traceable transformations over this ideal model. For example, this approach paves the way for powerful UML-defined semantics-based model transformations such as refactoring, aspect weaving, application of design patterns or round-trip engineering.