发表机构
University of Michigan; Vector Institute; University of Pennsylvania; Amazon; University of Waterloo（密歇根大学; 向量研究所; 宾夕法尼亚大学; 亚马逊; 滑铁卢大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究提出并验证了视觉 Transformer 中抽象概念接地依赖具体锚定概念的转喻机制，通过 Transcoders 分析 CLIP 和 DINO 编码器，在图标数据集上发现结构化电路，并通过因果干预证实其功能作用。
AI 中文摘要
我们研究了当训练数据提供的直接指称证据有限时，视觉 Transformer（Vision Transformers）如何对抽象概念（例如“愤怒”）进行接地（grounding）。我们假设存在一种转喻接地机制，其中抽象预测由具体的、可解释的锚定概念（例如“火”）驱动，这些锚定概念将视觉信号与抽象语义连接起来。通过在 CLIP 和 DINO 视觉编码器上应用 Transcoders，我们恢复了可与更具体概念的语义标签相关联的中间特征，并追踪它们在抽象概念识别背后的电路中的贡献。在一个精心策划的图标数据集上的实验揭示了结构化的转喻电路，其中感知原语主导早期层，而类似物体的锚定概念先于抽象目标出现。包含渲染文本的图像则转而采用一种独特的感知到文本的路径。因果干预进一步验证了转喻中间体在功能上参与了抽象概念的接地过程。
英文摘要
We study how Vision Transformers ground abstract concepts (e.g., angry) when training data provide limited direct referential evidence. We hypothesize a metonymic grounding mechanism in which abstract predictions are driven by concrete, interpretable anchor concepts (e.g., fire) that bridge visual signals to abstract semantics. By applying Transcoders on CLIP and DINO vision encoders, we recover intermediate features that can be associated with semantic labels for more concrete concepts, and trace their contributions in circuits underlying abstract concept recognition. Experiments on a carefully curated icon dataset reveal structured metonymic circuits, in which perceptual primitives dominate early layers and object-like anchors precede abstract targets. Images containing rendered text instead recruit a distinct perceptual-to-textual route. Causal interventions further validate that metonymic intermediates are functionally involved in grounding abstract concepts.
CommentsEMNLP 2026 Main. Project Website: https://github.com/jingjingjing-ding/metonymic-circuits
