Research map / Architectures and efficiency / Attention and transformers
Seeing Through Circuits: Faithful Mechanistic Interpretability for Vision Transformers
Nina Żukowska, Wolfgang Stammer, Bernt Schiele, Jonas Fischer
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 |
|---|---|---|---|
| use sequential activation patching on corrupted inpainted data (Vi-CD) instead of EAP / EAP-IG | ViT-B and OpenCLIP class circuits · ImageNet classes | faithfulness vs sparsity: about 10x sparser circuits at comparable accuracy | Improved |
| use EAP-IG instead of EAP (ablation) | edge attribution baselines on ViT · ImageNet classes | faithfulness | Improved |
| steer typographic-attack circuit (big text) | OpenCLIP ViT-B/32 · Corrupted ImageNet | ASR Top-1 (%): 39.1 -> 2.8 | Improved |
| steer typographic-attack circuit (big text) | OpenCLIP ViT-B/32 · Clean ImageNet | Top-1 accuracy (%): 57.0 -> 55.8 | Hurt |
| steer bezel-attack circuit | OpenCLIP ViT-B/32 · Clean ImageNet | Top-1 accuracy (%): 57.1 -> 49.9 | Hurt |
| steer typographic circuit (bezel) | OpenCLIP ViT-B/32 · Corrupted ImageNet | Top-1 accuracy (%): 34.4 -> 45.0 | Improved |
| steer weapon-related circuits (Assault Rifle, Revolver, Rifle) | OpenCLIP ViT-B/32 · RoCOCO danger benchmark | RSMS (%): 11.68 -> 5.15 / 5.69 / 4.86 | Improved |
| steer Rifle circuit on RoCOCO | OpenCLIP ViT-B/32 · RoCOCO retrieval | R_mean (%): 62.74 -> 64.37 | Improved |
| steer Assault Rifle and Revolver circuits on RoCOCO | OpenCLIP ViT-B/32 · RoCOCO retrieval | R@5 (%): 69.48 -> 68.59 / 68.39 | No clear effect |
| steer random circuits of same size instead of discovered circuits (ablation) | OpenCLIP typographic defense · Corrupted ImageNet | ASR and accuracy retention | Hurt |
Closest papers by meaning
- Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.90)
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.90)
- Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.89)
- InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques (NeurIPS 2024 · similarity 0.87)
- Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits (NeurIPS 2025 · similarity 0.87)
- Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? (ICLR 2025 · similarity 0.87)
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.
