Research map / Architectures and efficiency / Attention and transformers
CIGMA: Causal Information-Gain Mechanistic Attribution of Attention Heads in Vision Transformers
Maisha Maliha, Dean F. Hougen
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 |
|---|---|---|---|
| ablate the top-16 spurious attention heads found by CIGMA (training-free) | InternVL2-26B · CIFAR-10 (zero-shot) | accuracy: 92.0 -> 99.6 | Improved |
| ablate spurious heads | InternVL2-26B · CIFAR-100 (zero-shot) | BIR: 0.43 -> 0.051 | Improved |
| ablate spurious heads on top of fine-tuning | InternVL2-26B (ft) · CIFAR-10 (fine-tuned) | accuracy: 95.1 -> 99.8 | Improved |
| apply token pruning (TopV) | InternVL2-26B · CIFAR-10 (zero-shot) | accuracy: 92.0 -> 91.1 | Hurt |
| apply DivPrune token pruning | InternVL2-26B · CIFAR-10 (zero-shot) | accuracy: 92.0 -> 92.0 | No clear effect |
| apply MDP pruning | InternVL2-26B · CIFAR-10 (zero-shot) | accuracy: 92.0 -> 92.3 | No clear effect |
| fine-tune with CoBalT spurious mitigation | InternVL2-26B · CIFAR-10 (fine-tuned) | accuracy: 95.1 -> 93.1 | Hurt |
| fine-tune with RAVL spurious mitigation | LLaVA-1.5-13B · CIFAR-10 (fine-tuned) | accuracy: 93.1 -> 91.5 | Hurt |
| increase number of pruned heads K from 16 to 32 (ablation) | CIGMA · Tiny-ImageNet, InternVL2-26B | accuracy: 82.1 -> 80.2 | Hurt |
| prune K=16 heads versus none (ablation) | CIGMA · Tiny-ImageNet, InternVL2-26B | accuracy: 79.6 -> 82.1 | Improved |
| use 10% instead of 40% of true-positive images to score heads (ablation) | CIGMA · Tiny-ImageNet, InternVL2-26B | accuracy: 82.1 -> 81.1 | Hurt |
| use 100% instead of 40% of true-positive images to score heads (ablation) | CIGMA · Tiny-ImageNet, InternVL2-26B | accuracy: 82.1 -> 81.6 | No clear effect |
Closest papers by meaning
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · similarity 0.89)
- Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers (CVPR 2025 · similarity 0.88)
- Token Transformation Matters: Towards Faithful Post-hoc Explanation for Vision Transformer (CVPR 2024 · similarity 0.88)
- On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · similarity 0.87)
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 2024 · similarity 0.87)
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · 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.
