隨著 Vision Transformer（ViT）在電腦視覺領域的廣泛應用，如何有效解釋其內部決策機制已成為可解釋人工智慧（XAI）的重要課題。然而，現有可解釋性方法多針對具備 [CLS] token 的標準 ViT 設計，難以直接套用於採用 Patch Merging 機制的階層式 ViT 架構；且現有方法普遍僅提供空間定位資訊，缺乏語意層次的解釋，因而形成了空間定位與語意理解之間的分析缺口。 本研究以自行設計之階層式視覺轉換器 HierarchicalExplainableViT為分析對象，提出一套結合位置感知 K-means 分群、Grad-CAM 梯度定位與視覺語言模型（VLM）語意標注的可解釋性分析框架。針對 HierarchicalExplainableViT 各 Stage 輸出的 Position-wise 特徵向量，本研究採用餘弦等價 K-means 進行無監督分群，擷取各群聚之代表性 Patch 進行視覺化；並透過對 Patch Merging 結構的回溯映射，將 token 層級的 Grad-CAM 梯度正確還原至原始輸入座標，實現空間資訊的精準對齊。 於語意標注階段，本研究引入 Qwen2-VL 視覺語言模型對高階層之代表 Patch 進行自動描述，並輔以人工觀察加以驗證，以橋接空間定位與人類可理解語意之間的落差。 本研究於 Colored MNIST、Colored Fashion MNIST、BloodMNIST 及 Caltech-101 等四個資料集上進行實驗。結果顯示，HierarchicalExplainableViT 在參數規模顯著較小的前提下，達到與基準模型相當具競爭力的分類準確率；位置感知 K-means 分群能有效呈現模型由低階色彩與紋理、逐步過渡至高階語意結構的多尺度視覺概念。此外，實驗亦發現不同合併比例組合雖對準確率影響有限，但會顯著改變各階層感受野形狀，進而影響視覺化的空間特性。最後，本研究也指出 VLM 在高解析度情境下能提供具參考價值的語意描述，但在低解析度資料集上標注品質受限之限制。本框架不依賴額外監督資訊，即可揭示無 CLS Token 階層式 架構之內部表徵演化，為其設計原理提供具價值的分析視角。;With the widespread application of Vision Transformers (ViTs) in computer vision， effectively explaining their internal decision-making mechanisms has become a critical issue in Explainable Artificial Intelligence (XAI). However， existing explainability methods are predominantly designed for standard ViTs containing a [CLS] token， making them difficult to directly apply to hierarchical ViT architectures that utilize Patch Merging mechanisms. Furthermore， current approaches generally provide only spatial localization information and lack corresponding semantic-level explanations， leading to an analytical gap between spatial localization and semantic understanding. This study introduces an explainability analysis framework tailored for a self-designed hierarchical Vision Transformer， HierarchicalExplainableViT， by combining position-aware K-means clustering， Grad-CAM gradient ocalization， and Vision-Language Model (VLM) semantic annotation. Addressing the position-wise feature vectors output at each stage of HierarchicalExplainableViT， we employ cosine-equivalence K-means for unsupervised clustering and visualize the representative patches extracted from each cluster. By backmapping through the Patch Merging structure， token-level Grad-CAM gradients are accurately restored to the original input coordinates， achieving precise alignment of spatial information. In the semantic annotation phase， the Qwen2-VL vii model is introduced to automatically describe representative patches at higher hierarchical layers， supplemented by human observation to bridge the gap between spatial localization and human-comprehensible semantics. Experiments are conducted on four datasets: Colored MNIST， Colored Fashion MNIST， BloodMNIST， and Caltech-101. The results demonstrate that HierarchicalExplainableViT achieves competitive classification accuracy while using significantly fewer parameters than the baseline models. The positionaware K-means clustering effectively reveals multi-scale visual concepts learned across different stages of the model， progressively transitioning from low-level color and texture features to high-level semantic structures. Additionally， we find that while different merging ratio combinations have limited impact on classification accuracy， they significantly alter the receptive field shapes at various stages， thereby affecting the spatial characteristics of Grad-CAM localization and cluster visualization. Finally， this study highlights that while the VLM provides valuable semantic descriptions in high-resolution scenarios， its annotation quality is constrained in low-resolution datasets. Without relying on extra supervisory information， the proposed framework successfully uncovers the internal representation evolution of non-[CLS]-token hierarchical architectures， providing valuable analytical insights into their underlying design principles.
