AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Transformers Don’t Need LayerNorm at Inference Time: Scaling LayerNorm Removal to GPT-2 XL and Implications for Mechanistic Interpretability

Luca Baroni, Galvin Khara, Joachim Schaeffer, Marat Subkhankulov, Stefan Heimersheim

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws

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
remove all LayerNorm blocks by sequential FakeLN replacement and fine-tuningGPT-2 XL · The Pile-filteredcross-entropy loss: 2.3821 -> 2.3992 (+0.0253 vs best)Hurt
remove all LayerNorm blocksGPT-2 Small · The Pile-filteredcross-entropy loss: 2.7899 -> 2.8757Hurt
remove all LayerNorm blocksGPT-2 XL · The Pilecross-entropy loss: 2.4436 -> 130.2197Hurt
extend fine-tuning of the LN-free model (ablation)GPT-2 LN-free · OpenWebTextcross-entropy loss gap to vanillaNo clear effect
add auxiliary std-consistency loss at the final LN position (ablation)LN-removal fine-tuning · GPT-2 fine-tuningloss spikes / stabilityImproved
remove auxiliary loss from LN removal protocol (ablation)LN-removal protocol · larger GPT-2 modelsscaling to larger modelsHurt
remove LayerNormGPT-2 Small · direct logit attributionNMAE between DLA and direct effect: 49.07% -> 0.00%Improved
remove LayerNormGPT-2 Small · attribution patching on IOI promptsattribution error improvement score: mean -0.026No clear effect
remove LayerNormGPT-2 · attention sink ratesink rate: 55.3% -> 45.3%Hurt
remove LayerNormGPT-2 Medium · ablation of top-3 confidence neuronschange in cross-entropyHurt
remove LayerNormGPT-2 Medium · calibrationECE: 0.0019 -> 0.034Hurt

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