During requirements elicitation and preliminary design, a system’s behavior is typically partially specified: some behavior is defined as either forbidden or required, while other behavior is not yet categorized as either of those. The goal is then to gradually refine the specification and finally arrive at a complete behavioral description. Partial-behavior modeling formalisms, such as Modal Transition Systems, can support such gradual refinement. However, several challenges still remain, particularly in the context of hierarchical architectural specifications where (sub)system models are composed of smaller subsystems and components, and, in turn, abstracted to be made more compact and analyzable. Refinement of a behavior specification can be performed using models of varying scopes (e.g., different subsystems) and levels of abstraction, depending on the stakeholder needs. The primary challenge then becomes ensuring that a refinement of a model is correct when that model is a composition (or abstraction) of other models; this problem has not been addressed in the existing literature. In this paper, we propose a framework that supports reasoning about behavior refinement of composite and abstract models. Our framework assures that (1) a refinement of a composition is realized with refinements of the individual composed models and (2) a refinement of an abstract model is interpreted in terms of a refinement of the detailed model.