Abstract: To address the challenges of low efficiency and difficulty in optimizing manually designed neural network architectures, this paper reviews the research progress in neural architecture search (NAS) in recent years. Different from traditional reviews that often analyze NAS components individually, this paper builds a three-dimensional review framework based on NAS literature published since 2017, which includes core methods, multi-objective optimization, and cross-domain collaboration. Through analysis, five core NAS methods are identified: Reinforcement learning, evolutionary algorithms, gradient-based optimization, zero-cost proxies, and one-shot training. The corresponding multi-objective optimization strategies for these methods are analyzed, and cross-domain collaboration mechanisms are discussed from two aspects: hardware architecture integration and domain-specific knowledge fusion. Using this three-dimensional analysis framework, the paper describes the evolutionary patterns of NAS technologies, identifies technical challenges and future research directions in core NAS methods, multi-objective optimization, and cross-domain collaboration, and provides references for solving the problems of existing approaches. This contributes to the further development of NAS technologies.