AI Paper Atlas

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

CVPR 2026 · Architectures and efficiency › Attention and transformers · approach: Theory of self-attention dynamics

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 changedWhereMeasuredEffect
ablate the top-16 spurious attention heads found by CIGMA (training-free)InternVL2-26B · CIFAR-10 (zero-shot)accuracy: 92.0 -> 99.6Improved
ablate spurious headsInternVL2-26B · CIFAR-100 (zero-shot)BIR: 0.43 -> 0.051Improved
ablate spurious heads on top of fine-tuningInternVL2-26B (ft) · CIFAR-10 (fine-tuned)accuracy: 95.1 -> 99.8Improved
apply token pruning (TopV)InternVL2-26B · CIFAR-10 (zero-shot)accuracy: 92.0 -> 91.1Hurt
apply DivPrune token pruningInternVL2-26B · CIFAR-10 (zero-shot)accuracy: 92.0 -> 92.0No clear effect
apply MDP pruningInternVL2-26B · CIFAR-10 (zero-shot)accuracy: 92.0 -> 92.3No clear effect
fine-tune with CoBalT spurious mitigationInternVL2-26B · CIFAR-10 (fine-tuned)accuracy: 95.1 -> 93.1Hurt
fine-tune with RAVL spurious mitigationLLaVA-1.5-13B · CIFAR-10 (fine-tuned)accuracy: 93.1 -> 91.5Hurt
increase number of pruned heads K from 16 to 32 (ablation)CIGMA · Tiny-ImageNet, InternVL2-26Baccuracy: 82.1 -> 80.2Hurt
prune K=16 heads versus none (ablation)CIGMA · Tiny-ImageNet, InternVL2-26Baccuracy: 79.6 -> 82.1Improved
use 10% instead of 40% of true-positive images to score heads (ablation)CIGMA · Tiny-ImageNet, InternVL2-26Baccuracy: 82.1 -> 81.1Hurt
use 100% instead of 40% of true-positive images to score heads (ablation)CIGMA · Tiny-ImageNet, InternVL2-26Baccuracy: 82.1 -> 81.6No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy