Metonymic Circuits for Abstract Concept Grounding in Vision Transformers

Vision transformers (ViTs) have achieved remarkable generalization across visual domains, yet little is known about how they internally represent the structure of the world. To address this gap, we use Cross-Layer Transcoders (CLTs) to read concept circuits from ViTs: directed graphs whose nodes correspond to sparse, interpretable concepts and edges capture concept interactions across layers. Our method yields two complementary views of model behavior. The global concept circuit is input-invariant and can be recovered directly from learned cross-layer weights, exposing the reusable "world knowledge" encoded in the model. The instance concept circuit is input-dependent and identifies the concepts and pathways actually used for a specific prediction, enabling faithful example-level explanations. We demonstrate the utility of concept circuits in three ways: (1) Automatic spurious correlation discovery: leveraging the statistics of our global concept circuits to identify shortcut dependencies within the model. (2) Spurious correlation removal: intervening on the instance concept circuit to steer the model towards correct predictions. Empirical results show that our method outperforms existing counterparts by 11.0% on the Waterbird dataset. (3) Model comparison: contrasting the global concept circuits of different foundation models (e.g., CLIP vs. DINO) to reveal how supervision paradigms shape representational structure. Our code is available at https://github.com/deep-real/VisionCLT

Recent work reports that vision--language models (VLMs) struggle to establish and maintain stable reference in repeated reference games. Rather than ask which VLM does best, we ask a more basic question: do you need a large pretrained VLM for this at all? On grounding a single director utterance to one of twelve tangram silhouettes, we compare six off-the-shelf VLMs against a transparent baseline that uses no learned visual representation: classical SIFT keypoint matching and a signal-quality index over retrieved images. On identical trials, the transparent baseline matches the strongest VLM (SigLIP-large) and significantly outperforms the other five, including every CLIP and OpenCLIP variant. The baseline additionally retrieves external images, so this is not a matched-information comparison; what it shows is that a learned visual representation is not the bottleneck for this task: given retrieved images, a shape-appropriate classical similarity suffices. Along the way we find that abstract-grounding ability varies widely across VLMs (15--39% top-1; chance 8.33%, humans $\approx$77--80%), so the weakness is model-specific rather than intrinsic to contrastive pretraining; on the 1{,}013-shape KiloGram benchmark the pattern generalizes for CLIP, with per-shape difficulty tracking human shape-nameability. The pipeline is a classical, inspectable alternative rather than a learned one. We close by sketching how an explicit, inspectable representation of listener-side pact state could carry this approach into interactive multi-turn reference, which we leave to future work. Code available in supplementary material.

Concept-based explanations help users understand vision models through recognizable visual patterns. However, existing methods often rely on correlational signals without directly validating which image cues support prediction-relevant internal features. To this end, we introduce VisionLogic, a post-hoc framework that grounds these features in visual concepts through intervention-based validation. VisionLogic first identifies compact sets of features whose contributions reproduce the model's original prediction. It then represents their activation states as predicates using class-specific thresholds. An iterative refinement procedure grounds these predicates in visual regions through ablation tests. A region is accepted when its removal deactivates the corresponding predicate, linking the feature's numerical role to visual evidence in the input. The same predicates allow us to examine how features are activated, selected, and reused across images and classes. Across CNNs and vision transformers on ImageNet-1k, we find that only a few features are selected to explain each prediction, and frequently active features are not always selected. In a large-scale human evaluation with 465 participants, VisionLogic significantly improves participants' understanding of model behavior over established methods ACE and CRAFT. Code is available at https://github.com/allengeng123/VisionLogic.