Contemporary engineering and scientific applications frequently rely on black-box optimization, where objective functions are expensive to evaluate and analytically opaque. While evolutionary algorithms are widely employed for such problems, their performance degrades severely in high-dimensional, irregular search landscapes. This study introduces an evolutionary framework that addresses this challenge through adaptive, autoencoder-driven dimensionality reduction. The core of the method lies in simultaneous exploration of the original decision space and a learned latent representation, which is continuously updated using population data during the search. This dual-space strategy allows the algorithm to exploit problem-specific structures, and its efficacy is rigorously assessed on established numerical benchmarks. The experimental results have demonstrated that the proposed approach enhances the core search algorithm’s search space exploration.