arXiv:2607.14938v2 Announce Type: replace Abstract: Wireless observations capture radio signal responses formed through interactions with propagation environments and spatial geometry. In integrated sensing and communication, such observations have become an important basis for high-accuracy localization beyond conventional channel estimation. Learning-driven methods learn implicit relations between channel propagation and spatial position, enabling location inference under complex channel conditions. However, the useful information is tightly coupled with spatial layout, environmental dynamics, and system configurations. This coupling obscures the inference process and weakens performance consistency across scenarios. In this paper, we review learning-driven high-accuracy localization techniques with channel representations as the organizing view. The survey covers typical channel observation forms and analyzes their physical meanings. We also review channel feature extraction and representation learning methods, and summarize methods according to the acquisition, organization, adaptation, and reuse of channel representations. Typical methods are compared in terms of accuracy, applicable conditions, data requirements, and generalization. We model the localization process as a unified ``wireless observation--channel representation--location inference'' framework and use information bottleneck theory to highlight that the quality and usability of channel representations are critical to exploiting propagation information and play a decisive role in localization performance. Finally, we summarize the key challenges in moving from experimental studies to real deployment and present our perspectives on these issues.