Research map / Architectures and efficiency / Attention and transformers
DeepNet: Scaling Transformers to 1,000 Layers
Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Dongdong Zhang, Furu Wei
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 |
|---|---|---|---|
| scale depth from 12 to 48 layers (ablation) | baseline multilingual NMT · OPUS-100 | average BLEU: 24.5 -> 27.7 | Improved |
| scale DeepNet depth from 200 to 1000 layers (ablation) | DeepNet · OPUS-100 | average BLEU: 31.1 -> 32.1 | Improved |
| use DeepNorm at 1000 layers instead of 48-layer baseline | multilingual NMT · OPUS-100 | average BLEU: 27.7 -> 32.1 | Improved |
| use deep narrow DeepNet 200L 3.2B instead of M2M-100 48L 12B | multilingual NMT · Flores | BLEU: 13.6 -> 18.6 | Improved |
| use DeepNet instead of M2M-100 | multilingual NMT · WMT | BLEU: 31.9 -> 33.9 | Improved |
| use DeepNet instead of M2M-100 | multilingual NMT · OPUS | BLEU: 18.4 -> 23.0 | Improved |
| use DeepNet instead of M2M-100 | multilingual NMT · TED | BLEU: 18.7 -> 20.1 | Improved |
| use DeepNorm instead of Post-LN at 100L-100L | Transformer · WMT-17 En-De | BLEU: diverged -> 28.9 | Improved |
| use DeepNorm instead of Pre-LN at 50L-50L | Transformer · WMT-17 En-De | BLEU: 28.0 -> 29.0 | Improved |
| use DeepNorm instead of Post-LN at 6L-6L | Transformer · WMT-17 En-De | BLEU: 28.1 -> 27.8 | No clear effect |
| increase DeepNet depth from 18L-18L to 50L-50L (ablation) | DeepNet · WMT-17 En-De | BLEU: 28.8 -> 29.0 | Improved |
| increase R-Fixup depth from 18L-18L to 50L-50L (ablation) | R-Fixup (no LN) · WMT-17 En-De | BLEU: 28.4 -> 27.7 | Hurt |
Closest papers by meaning
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.89)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.88)
- Global Convergence in Training Large-Scale Transformers (NeurIPS 2024 · similarity 0.87)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.87)
- Peri-LN: Revisiting Normalization Layer in the Transformer Architecture (ICML 2025 · similarity 0.87)
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.86)
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.
