Speech dereverberation is challenging for various speech processing systems. Recently, phase recovery is proved to be significant for improving speech quality and intelligibility, and numerous supervised speech dereverberation algorithms begin focusing on complex spectrum estimation. However, these methods recover clean speech phase at the expense of severe magnitude distortion due to the magnitude-phase compensation effect. To address this problem, we propose a novel loss-guided two-stage framework to progressively guide the process of complex spectrum recovery. In the first stage, a bifurcated network is proposed to separately optimize the magnitude and phase of the complex spectrum coarsely by two distinct loss functions. After that, a reunited network is devised to exploit the complementary characteristics of previous estimations and further refine the complex spectrum. A mathematical derivation is presented to reveal the magnitude-phase compromise phenomenon and validate the rationality of the proposed objective optimization strategy. Experimental results demonstrate that the proposed method improves both speech quality and intelligibility in the dereverberation task, and outperforms other baseline methods.