mlsanityData to 14 Oct 2026

Metonymic Circuits for Abstract Concept Grounding in Vision Transformers

arXiv 2610.06928 ↗First posted 3 Oct 2026cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 25 of its 72 references that are in the map. First posted on arXiv 3 Oct 2026.

Builds on

25 of 72 references.

References · 25

  1. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  2. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  3. 2023Shikra: Unleashing Multimodal LLM's Referential Dialogue Magiccited by 906
  4. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  5. 2023Ferret: Refer and Ground Anything Anywhere at Any Granularitycited by 509
  6. 2023GLaMM: Pixel Grounding Large Multimodal Modelcited by 455
  7. 2019The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervisioncited by 372
  8. 2024Transcoders Find Interpretable LLM Feature Circuitscited by 200
  9. 2023Interpreting CLIP's Image Representation via Text-Based Decompositioncited by 196
  10. 2020Grounded Situation Recognitioncited by 78
  11. 2025Interpreting CLIP with Hierarchical Sparse Autoencoderscited by 60
  12. 2025Steering CLIP's vision transformer with sparse autoencoderscited by 38
13 more
  1. 2022Exploring Visual Interpretability for Contrastive Language-Image Pre-trainingcited by 30
  2. 2024Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIPcited by 20
  3. 2025DECEPTICON: How Dark Patterns Manipulate Web Agentscited by 15
  4. 2025It's a (Blind) Match! Towards Vision-Language Correspondence without Parallel Datacited by 14
  5. 2024Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objectscited by 13
  6. 2022Grammar-Based Grounded Lexicon Learningcited by 11
  7. 2025Probing the Representational Power of Sparse Autoencoders in Vision Modelscited by 10
  8. 2026Circuit Tracing in Vision-Language Models: Understanding the Internal Mechanisms of Multimodal Thinkingcited by 8
  9. 2025VisualTrap: A Stealthy Backdoor Attack on GUI Agents via Visual Grounding Manipulationcited by 6
  10. 2025The Mechanistic Emergence of Symbol Grounding in Language Modelscited by 6
  11. 2025Causal Interpretation of Sparse Autoencoder Features in Visioncited by 5
  12. 2014Context-based Word Acquisition for Situated Dialogue in a Virtual Worldcited by 4
  13. 2026A Computational Approach to Visual Metonymycited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Symbolic Grounding Reveals Representational Bottlenecks in Abstract Visual Reasoning
  2. 2023Emergence and Function of Abstract Representations in Self-Supervised Transformers
  3. 2026"World Knowledge" in the Weights: Reading Concept Circuits of Vision Transformers
  4. 2023Describe me an Aucklet: Generating Grounded Perceptual Category Descriptions
  5. 2026Transcoders Trace Visual Grounding and Hallucinations in Vision-Language Models
  6. 2025Concept-RuleNet: Grounded Multi-Agent Neurosymbolic Reasoning in Vision Language Models
  7. 2024What Makes a Maze Look Like a Maze?
  8. 2025Uncovering Grounding IDs: How External Cues Shape Multimodal Binding