Vision transformers reveal reusable world knowledge inside their layers
"World Knowledge" in the Weights: Reading Concept Circuits of Vision Transformers
Computer Vision and Pattern Recognition
Summary
Vision transformers are a type of AI model that can understand images well, but how they internally represent concepts from the world is not clear. The authors introduce a way to map out these internal concepts and how they connect across the model’s layers, creating what they call concept circuits. These circuits can show which ideas the model uses all the time and which are involved in specific decisions. This understanding helps identify when the model relies on misleading cues and allows fixing these mistakes. The method also lets people compare different models to see how their training shapes what they learn about the world.
vision transformer (ViT)concept circuitscross-layer transcoders (CLTs)model interpretabilityspurious correlationsfoundation modelsmodel explanationmachine learningrepresentation learningshortcut dependencies
Authors
Yanlin Chen, Tang Li, Xi Peng
Abstract
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