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
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 |
|---|---|---|---|
| remove all LayerNorm blocks by sequential FakeLN replacement and fine-tuning | GPT-2 XL · The Pile-filtered | cross-entropy loss: 2.3821 -> 2.3992 (+0.0253 vs best) | Hurt |
| remove all LayerNorm blocks | GPT-2 Small · The Pile-filtered | cross-entropy loss: 2.7899 -> 2.8757 | Hurt |
| remove all LayerNorm blocks | GPT-2 XL · The Pile | cross-entropy loss: 2.4436 -> 130.2197 | Hurt |
| extend fine-tuning of the LN-free model (ablation) | GPT-2 LN-free · OpenWebText | cross-entropy loss gap to vanilla | No clear effect |
| add auxiliary std-consistency loss at the final LN position (ablation) | LN-removal fine-tuning · GPT-2 fine-tuning | loss spikes / stability | Improved |
| remove auxiliary loss from LN removal protocol (ablation) | LN-removal protocol · larger GPT-2 models | scaling to larger models | Hurt |
| remove LayerNorm | GPT-2 Small · direct logit attribution | NMAE between DLA and direct effect: 49.07% -> 0.00% | Improved |
| remove LayerNorm | GPT-2 Small · attribution patching on IOI prompts | attribution error improvement score: mean -0.026 | No clear effect |
| remove LayerNorm | GPT-2 · attention sink rate | sink rate: 55.3% -> 45.3% | Hurt |
| remove LayerNorm | GPT-2 Medium · ablation of top-3 confidence neurons | change in cross-entropy | Hurt |
| remove LayerNorm | GPT-2 Medium · calibration | ECE: 0.0019 -> 0.034 | Hurt |
Closest papers by meaning
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · similarity 0.91)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.91)
- Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA (ICLR 2025 · similarity 0.89)
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.88)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.88)
- Mix-LN: Unleashing the Power of Deeper Layers by Combining Pre-LN and Post-LN (ICLR 2025 · similarity 0.88)
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.
