We propose a novel dual-path evolutionary framework for open-world test-time learning, designed to effectively address the stability dilemma. The framework introduces three key innovations: (1) an adaptive thresholding mechanism that formulates OOD detection as a constrained dual-objective optimization problem, enabling dynamic adjustment of decision boundaries; (2) a dual-path masked autoencoder (MAE) architecture that constructs the solution space through collaborative optimization of local and global feature representations; and (3) a distribution alignment module guided by gradients, which introduces a domain adaptation strategy to avoid inefficient search in high-dimensional feature spaces. Experiments on the CIFAR-10 benchmark demonstrate that the proposed method performs significantly better than existing approaches in challenging scenarios such as noise contamination and cross-domain interference. It achieves a 12.3% improvement in known-class retention and an 8.7% increase in novel-category discovery. Additional ablation studies further confirm the effectiveness of each module in enhancing model robustness. Overall, this method offers a new technical perspective for online adaptation in open environments.