Large-scale multiobjective optimization problems (LSMOPs), characterized by a large number of decision variables, frequently arise in real-world scenarios. To address these problems, we propose a hierarchical layered mutual learning (HLML) strategy in this paper which aims to accelerate population convergence, while preserving diversity by stratifying the population into three levels via non-dominated sorting and employing two multilayer perceptrons (MLPs) for targeted learning. Specifically, two MLPs with one hidden layer are constructed. To obtain suitable training data for these models, the current population is divided into three levels based on non-dominated sorting. Individuals from the high-quality layer are paired with those from the intermediate and low-quality layers according to the minimum angle criterion, generating two datasets that represent the most promising directions for rapid convergence of inferior individuals. Subsequently, the two MLPs are updated via using these datasets. Finally, a Multi-Level Mutual Learning-Based Evolutionary Search for Large-Scale Multi-Objective Optimization (MLLMOEA) is designed by utilizing the offspring population generated by HLML. The effectiveness of this algorithm is validated through comparisons with four state-of-the-art evolutionary algorithms, showing consistently better performance and computational efficiency across different LSMOPs.