AI

Metonymic Circuits for Abstract Concept Grounding in Vision Transformers

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 02:18:24 UTC.

Briefings

This study investigates how Vision Transformers ground abstract concepts like 'angry' when direct visual evidence is scarce. The central finding is the discovery of metonymic circuits where concrete anchor concepts (e.g., 'fire') bridge visual signals to abstract semantics. Experiments on a curated icon dataset reveal structured circuits where perceptual primitives dominate early layers, and object-like anchors precede abstract targets. Causal interventions confirm these intermediates are functionally involved in grounding abstract concepts.

Key points

• Vision Transformers use metonymic circuits to ground abstract concepts like 'angry' when direct visual evidence is limited.
• Perceptual primitives dominate early layers while object-like anchors precede abstract targets in the circuit structure.

Context

The research addresses the challenge of abstract concept recognition in Vision Transformers when training data lacks direct referential evidence. This work provides a mechanism where concrete anchors bridge visual signals to abstract semantics.

Limitations

Evidence is limited to a carefully curated icon dataset. Causal interventions validate findings but do not provide full generalization across all visual domains. Preprint status (arXiv:2610.06928v1) indicates uncertainty regarding final validation and reproducibility.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 03:55:10 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifyzuhqso34s2icfh3252g547y3d6htn65g4ey6e2rxmldxq43lry