中文线性探测的盲区：面向权重空间学习的多视角探测

开源模型仓库“模型丛林”中检查点常缺文档。权重空间学习虽可直接分析参数，但计算成本高。现有探测方法通过可学习向量提取置换等变表示，提供轻量级替代，但受限于单视图，需改进以提升模型识别能力。

arXiv:2605.23410v1 Announce Type: cross Abstract: The explosive growth of open-source model repositories has created a Model Jungle, where checkpoints are frequently shared without adequate documentation or metadata. While weight-space learning offers a pathway to identify and analyze these models directly from their parameters, processing full-scale weights is computationally prohibitive. Probing-based methods have emerged as a lightweight alternative, extracting permutation-equivariant representations via learnable probe vectors. However, existing probing methods are limited by a single-view