AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models

Yearim Kim, Sangyu Han, Nojun Kwak

TPAMI 2026 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining · 0 citations (OpenAlex)

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
add class-centering plus positive clamping to sharing ratios (vs Raw) (ablation)SRD-input (VGG16) · ImageNet-S50Attribution Localization: .490 -> .561Improved
add class-centering plus positive clamping to sharing ratios (vs Raw) (ablation)SRD-input (VGG16) · ImageNet-S50Sparseness: .746 -> .788Improved
add class-centering plus positive clamping to sharing ratios (vs Raw) (ablation)SRD-input (ResNet50) · ImageNet-S50Sparseness: .596 -> .724Improved
add positive clamping after class-centering (vs Mean) (ablation)SRD-input (VGG16) · ImageNet-S50Attribution Localization: .404 -> .561Improved
add positive clamping after class-centering (vs Mean) (ablation)SRD-input (ResNet50) · ImageNet-S50Pointing Game: .958 -> .953Hurt
use SRD-high instead of LayerCAM-highattribution method (VGG16) · ImageNet-S50Attribution Localization: .435 -> .566Improved
use SRD-input instead of GradCAMattribution method (ResNet50) · ImageNet-S50Stability: .757 -> .104Improved
use SRD instead of Fullgrad on ReLU activationResNet50 CIFAR-100 · CIFAR-100Fidelity: 0.038 -> 0.078Improved
use SRD with Tanh activation instead of FullgradResNet50 CIFAR-100 · CIFAR-100Fidelity: 0.107 -> 0.115Improved
use input-scale SRD instead of high-layer SRDSRD (VGG16) · ImageNet-S50Fidelity: .206 -> .069Hurt
use bisecting k-means instead of SAE for concept extraction at Layer1.1CAFE concept extraction · ResNet50 PFVsRel-l2 reconstruction error: 1.8952 -> 0.3651Improved
use SAE instead of bisecting k-means for concept extraction at Layer4.2CAFE concept extraction · ResNet50 PFVsRel-l2 reconstruction error: 0.5095 -> 0.4816Improved

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