Vision Transformers trained mostly on concrete words can still recognize abstract concepts such as angry. This work hypothesizes that they do so through metonymic grounding: a small set of concrete visual cues acts as a bridge. To test this, the researchers collected an icon dataset of abstract concepts with fine-grained labels for the concrete depicted elements, then analyzed CLIP and DINO encoders with Transcoders, sparse feature-extraction modules applied to MLP layers.
Tracing these features revealed structured metonymic circuits. Early layers encode perceptual primitives like colors and simple shapes, then object-like anchors such as fire precede the emergence of abstract targets. Images containing rendered text recruit a separate perceptual-to-textual route. Causal interventions confirmed that the concrete intermediate features are functionally involved, not merely correlated.
Why it matters: abstract concepts are notoriously hard to ground because they lack direct referents. Showing that current vision models rely on concrete, human-interpretable anchors makes their behavior more auditable, and opens a path toward detecting spurious cues in icon and multimodal systems.