AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior

Florian Eichin, Yupei Du, Philipp Mondorf, Maria Matveev, Barbara Plank, Michael A. Hedderich

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 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
use 100 instead of 10 integration steps in the test feature map (ablation)EPK representation · MOD-113 TransformerEPK accuracy: 0.748 -> 1.0Improved
use 100 instead of 10 integration steps (ablation)EPK representation · MOD-113 TransformerKL divergence: 0.885 -> 0.0Improved
use 100 instead of 10 integration steps (ablation)EPK representation · CIFAR-2 ResNet9EPK accuracy: 0.997 -> 1.0Improved
use 100 instead of 10 integration steps (ablation)EPK representation · MNIST CNNEPK accuracy: 0.998 -> 1.0Improved
prune CNN weights by ExPLAIND kernel importance (training-free) vs retrain-and-prune baselineCNN, MNIST · MNIST CNN pruning 70% to 99% sparsitytest accuracyNo clear effect
prune CNN weights by ExPLAIND kernel importance vs magnitude-pruning baselineCNN, MNIST · MNIST CNN pruningKL divergence to original modelImproved
keep only top-ranked update steps (about 60% sparsity) in decomposition (ablation)ExPLAIND · MNIST CNNreconstructed prediction accuracyNo clear effect
initialize Attention and Linear-1 layers from grokked model, rest random (ablation)Grokking Transformer · MOD-113generalization speedImproved
start from earlier-checkpoint representation pipeline instead of final (ablation)Grokking Transformer · MOD-113final accuracyHurt
swap grokked outer layers into intermediate checkpoints (ablation)Grokking Transformer · MOD-113predictions of checkpointsImproved
use data influence scores derived from ExPLAIND for causal data attributionExPLAIND · CIFAR and MNISTdata attributionNo 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