Abstract: To address the challenges in passive underwater acoustic source localization, such as performance degradation and limited generalization caused by low signal-to-noise ratio (SNR) and environmental mismatch in complex marine waveguides (charaterized by multipath propagation, severe noise interference, and environmental time-variance), this paper systematically reviews the technological evolution and frontier advancements in this field, providing a reference for enhancing the reliability and engineering applicability of localization technologies in ocean exploration and underwater warfare scenarios. First, the developmental stages from geometric measurement to physical field matching are outlined. The key characteristics of traditional model-driven methods, including time-delay estimation-based geometric localization, acoustic field interference parameter inversion, and matched-field processing (MFP) are comparatively analyzed in dimensions such as engineering applicability and environmental sensitivity. Second, focusing on recent research progress of deep learning in passive localization, the evolution of task modeling paradigms, feature input representations, and network architectures (progressing from CNNs to hybrid architectures of U-Net and Transformers) are systematically summarized. Finally, addressing the core bottlenecks of "data scarcity and domain shift," coping strategies including simulation-based data augmentation, transfer learning, self-supervised learning, and unsupervised domain adaptation are discussed in depth. The synthesized findings indicate that while traditional model-driven methods are generally constrained by a high reliance on prior knowledge of the ocean environment, deep learning methods demonstrate significant potential for robustness against environmental mismatch through data-driven approaches. Currently, the deep learning localization paradigm is evolving from singular black-box mapping toward multi-task joint inference. Utilizing multi-task joint constraints and attention mechanisms to process spatial statistical features (e.g., covariance matrices and spectrograms) has become the mainstream solution to effectively mitigate multipath ambiguities and feature space confusion. To facilitate future engineering implementation, deep learning-based passive localization requires synergistic breakthroughs along four pathways: constructing multi-environment test benchmarks on the data front; transitioning the learning paradigm toward "self-supervised representation + few-shot supervision"; deepening multi-task and attention constraints on the model front; and deeply coupling physics-informed neural networks (PINNs) with traditional MFP to ensure physical consistency. These efforts will propel the data-physiccal dual driven underwater acoustic localization models toward cross-domain applicability, interpretability, and rigorous evaluability.