In this Letter, we propose a novel, to the best of our knowledge, adaptive feedback-driven probabilistic constellation-shaping (FBD-PCS) method based on the robustness evaluation criteria and employ variational autoencoder (VAE)-based equalizers to implement polarization demultiplexing and nonlinear equalization for the recovery of high-order PCS-QAM signals. We experimentally demonstrate the fiber-THz 2 times 2 MIMO system with a net rate of 366.4 Gbit/s using dual-polarization 40 Gbaud PCS-64QAM signal over a 20 km SSMF and 6 m wireless link. Specifically, the feedback mechanism drives the fiber-THz system to solve optimization problems, adaptively matching the optimized distribution of transmitted symbols that maximizes normalized generalized mutual information (NGMI). We also examine six scenarios to explore nonlinear resistances of FBD-PCS symbols and the robustness of VAE-based equalizers. The results demonstrate the superiority of FBD-PCS over the Maxwell-Boltzmann (M-B) distributions in practical nonlinear-dominant systems. Additionally, the FBD-PCS signals can break limitations for ultrahigh rate transmission with the help of advanced equalizers.