Research map / Architectures and efficiency / Attention and transformers
World Knowledge in the Weights: Reading Concept Circuits of Vision Transformers
Yanlin Chen, Tang Li, Xi Peng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| steer with CLT instance concept circuit (ablate background-specific concepts) | CLIP-ViT-B/32 linear probe · Waterbirds (waterbirds on land, worst group) | accuracy: 0.48 -> 0.71 | Improved |
| CLT steering costs accuracy on majority group | CLIP-ViT-B/32 linear probe · Waterbirds (landbirds on land) | accuracy: 0.98 -> 0.95 | Hurt |
| ablate randomly chosen concepts instead of circuit-identified ones (ablation) | CLIP-ViT-B/32 linear probe · Waterbirds (waterbirds on land, worst group) | accuracy: 0.48 -> 0.22 | Hurt |
| add co-activation reweighting to global weights (ablation) | global concept circuit · ImageNet val (50 classes), top-10 concept mean-ablation | accuracy after ablation: 0.48 -> 0.22 (CLIP) | Improved |
| use global weights rather than co-activation statistics alone (ablation) | global concept circuit · ImageNet val (50 classes), top-10 concept mean-ablation | accuracy after ablation: 0.72 -> 0.62 (DINO) | Improved |
| CLT instead of SAE for instance concept circuits | CLIP instance circuits · ImageNet val (50 classes) | faithfulness and completeness: +24.4% faithfulness, +22.1% completeness (CLIP, average) | Improved |
| increase CLT L0 sparsity and dictionary size (ablation) | CLIP CLT · ImageNet train (hyperparameter sweep) | explained variance vs monosemanticity | Mixed |
Closest papers by meaning
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · similarity 0.90)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.89)
- CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification (CVPR 2025 · similarity 0.89)
- NeuronEye: Query-Guided Visual Concept Activation for Vision-Language Reasoning (NeurIPS 2026 · similarity 0.88)
- Interpreting vision transformers via residual replacement model (NeurIPS 2025 · similarity 0.88)
- Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objects (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
