Component model synthesis (CMS) has been widely used for model order reduction in dynamic analysis of jointed structures with localized non-linearities. The main drawback of these CMS methods is that their computational efficiency largely depends on the size of contact friction interfaces. This work proposes an adaptive reduction approach to improve these CMS based reduction methods in the application to the assembled structure with frictional interfaces. The main idea of this method is that, instead of retaining the whole frictional interface DOFs in the reduced model, only those DOFs in a slipping or separating condition are retained. This would significantly reduce the size of classical CMS based reduced models for dynamical analysis of jointed structure with micro-slip motion, leading to an impressive computational saving. This novel approach is based on a reformulated dynamic system that consists of a underlying linearised system and an adaptive internal variable to account the effects of non-linear contact friction force on the interface. The paper also describes the detailed implementation of the proposed approach with harmonic balance method for non-linear spectral analysis, where a new updating algorithm is put forward to enable the size of the reduced model can be automatically updated according to the contact condition of interface nodes. Two distinct FE joint models are used to validate the proposed method. It is demonstrated that the new approach can achieve a considerable computing speed-up comparing to the classic CMS approach while maintain the same accuracy.