Abstract: In mixed-model assembly job shops,significant differences exist between production batches in the processing and assembly stages,which can readily lead to work-in-process accumulation,assembly errors,missing parts and even product quality defects. This study addressed the lot streaming scheduling problem in mixed-model assembly job shops,formulated a mixed-integer linear programming model,and proposed a learning-assisted multi-objective evolutionary algorithm based on decomposition strategy(LMOEA/D)to solve the model. First,two neural network surrogate models were constructed,and a cooperative initialization strategy was established to improve the quality of initial solutions. Second,a crossover object selection mechanism was designed using reinforcement learning to adaptively adjust the global search space. In addition,four objective-oriented neighborhood search operators were constructed to enhance the convergence ability of the algorithm in the specified objective space. Experimental results show that,compared with the classic multi-objective optimization algorithms,the proposed algorithm has remarkable advantages in convergence and diversity metrics,which is conducive to the precise control of work-in-process inventory and the stabilization of the production process.