In this paper, we propose and prove the theorem regarding the stability of attributes in a decision system. Based on the theorem, we propose the LRA framework for accelerating rough set algorithms. It is a general-purpose framework which can be applied to almost all rough set methods significantly . Theoretical analysis guarantees high efficiency. Note that the enhancement of efficiency will not lead to any decrease of the classification accuracy. Besides, we provide a simpler prove for the positive approximation acceleration framework.