Research map / Architectures and efficiency / RNNs and normalization
Root Mean Square Layer Normalization
Biao Zhang, Rico Sennrich
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 |
|---|---|---|---|
| Replace LayerNorm with RMSNorm | RNNSearch · WMT14 English-German, TensorFlow newstest2014 | SacreBLEU: 22.6 -> 22.4 | No clear effect |
| Replace LayerNorm with RMSNorm | RNNSearch · WMT14 English-German, TensorFlow newstest2017 | SacreBLEU: 23.6 -> 23.7 | No clear effect |
| Replace LayerNorm with RMSNorm | RNNSearch · WMT14 English-German, TensorFlow | seconds per 1k training steps: 665±32.5s -> 501±11.8s ({24.7%}) | Improved |
| Replace RMS normalization with L2 normalization (ablation) | RNNSearch · WMT14 English-German, TensorFlow newstest2014 | SacreBLEU: 22.4 -> 20.7 | Hurt |
| Estimate RMS from 6.25% of summed inputs (ablation) | RNNSearch · WMT14 English-German, TensorFlow newstest2014 | SacreBLEU: 22.4 -> 22.6 | No clear effect |
| Estimate RMS from 6.25% of summed inputs (ablation) | RNNSearch · WMT14 English-German, TensorFlow newstest2017 | SacreBLEU: 23.7 -> 23.1 | No clear effect |
| Remove normalization (ablation) | Transformer · WMT14 English-German | SacreBLEU | Hurt |
| Replace LayerNorm with RMSNorm | Transformer · WMT14 English-German, newstest2014 | SacreBLEU: 26.6 -> 26.8 | No clear effect |
| Replace LayerNorm with RMSNorm | Transformer · WMT14 English-German, newstest2017 | SacreBLEU: 27.7 -> 27.7 | No clear effect |
| Replace LayerNorm with RMSNorm | Transformer · WMT14 English-German | seconds per 1k training steps: 248±1.31s -> 231±0.04s ({6.9%}) | Improved |
| Replace LayerNorm with RMSNorm | Order-embedding model · Microsoft COCO image-caption retrieval | Caption Retrieval R@1: 47.9 -> 48.7 | Improved |
| Estimate RMS from a subset of summed inputs (ablation) | Order-embedding model · Microsoft COCO image-caption retrieval | Caption Retrieval R@1: 48.7 -> 46.8 | Mixed |
Closest papers by meaning
- Understanding and Improving Layer Normalization (NeurIPS 2019 · similarity 0.90)
- Pre-RMSNorm and Pre-CRMSNorm Transformers: Equivalent and Efficient Pre-LN Transformers (NeurIPS 2023 · similarity 0.89)
- Impact of Layer Norm on Memorization and Generalization in Transformers (NeurIPS 2025 · similarity 0.86)
- On the Nonlinearity of Layer Normalization (ICML 2024 · similarity 0.86)
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.84)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.83)
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.
