Research map / Architectures and efficiency / Attention and transformers
Transformers without Normalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dynamic Tanh (DyT), an element-wise tanh(ωx) with learnable scale and affine parameters, as a drop-in replacement for LayerNorm/RMSNorm so Transformers can train without normalization statistics.
Problem
Normalization layers are considered indispensable for training modern deep networks, especially Transformers, but they add computation and their mechanism is not well understood.
Approach
LayerNorm in trained Transformers behaves like a scaled tanh (linear near zero, squashing extreme values); DyT replaces the normalization layer with a per-element tanh(ωx) plus per-channel affine parameters, requiring no activation statistics.
Key results
- DyT matches or slightly exceeds LN on ImageNet-1K supervised classification (e.g., ViT-B 82.5% vs 82.3%, ViT-L 83.6% vs 83.1%).
- On LLaMA pretraining (7B to 70B, 200B tokens), DyT matches RMSNorm zero-shot scores (e.g., 0.513 vs 0.513 for 7B) with training loss within 0.01.
- Removing tanh (identity) causes training to diverge (ViT-B fails at 61.0% accuracy), while tanh gives 82.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Layer Norm with DyT | ViT-B · ImageNet-1K | top-1 accuracy: 82.3% -> 82.5% | Helped |
| Replace Layer Norm with DyT | ViT-L · ImageNet-1K | top-1 accuracy: 83.1% -> 83.6% | Helped |
| Replace Layer Norm with DyT | ConvNeXt-B · ImageNet-1K | top-1 accuracy: 83.7% -> 83.7% | No clear effect |
| Replace Layer Norm with DyT | MAE ViT-L · ImageNet-1K fine-tuning | accuracy: 85.5% -> 85.4% | No clear effect |
| Replace Layer Norm with DyT | DINO ViT-B patch size 8 · ImageNet-1K | accuracy: 84.1% -> 84.5% | Helped |
| Replace Layer Norm with DyT | DiT-B · ImageNet-1K | FID: 64.9 -> 63.9 | Helped |
| Replace Layer Norm with DyT | DiT-XL · ImageNet-1K | FID: 19.9 -> 20.8 | Hurt |
| Replace RMSNorm with DyT | LLaMA 7B · 15 zero-shot lm-eval tasks | average score: 0.513 -> 0.513 | No clear effect |
| Replace tanh with identity, retaining ω (ablation) | ViT-B DyT · ImageNet-1K | accuracy: 82.5% -> 61.0% → failed | Hurt |
| Replace tanh with hardtanh (ablation) | ViT-B DyT · ImageNet-1K | accuracy: 82.5% -> 82.2% | Hurt |
| Replace tanh with sigmoid (ablation) | ViT-B DyT · ImageNet-1K | accuracy: 82.5% -> 81.6% | Hurt |
| Remove learnable ω (ablation) | ViT-B DyT with tanh · ImageNet-1K | accuracy: 82.5% -> 81.1% | Hurt |
Surprising findings
- Without any squashing function, ViT training diverges, showing the bounded nonlinearity is essential rather than just the learnable scale.
- DyT struggles to directly replace batch normalization in ResNets according to preliminary experiments.
Limitations
- Experiments only cover networks using LayerNorm or RMSNorm; replacing batch normalization in ResNets is hard.
- LLM training requires careful tuning of the initial scale parameter ω0 for good performance.
- Training of larger ViT models can become unstable without lowering ω0 or the learning rate.
Experimental setup
Datasets: ImageNet-1K LibriSpeech The Pile GenomicBenchmarks human reference genome
Benchmarks: ImageNet-1K GenomicBenchmarks lm-eval 15 zero-shot tasks LibriSpeech
Metrics: top-1 accuracy FID validation loss training loss average zero-shot score
Baselines: LayerNorm RMSNorm hardtanh sigmoid identity (no squashing)
Builds on: layer normalization RMSNorm Vision Transformer Diffusion Transformer LLaMA
Closest papers by meaning
- Peri-LN: Revisiting Normalization Layer in the Transformer Architecture (ICML 2025 · similarity 0.87)
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.87)
- Subcritical Signal Propagation at Initialization in Normalization-Free Transformers (NeurIPS 2026 · similarity 0.86)
- SLAB: Efficient Transformers with Simplified Linear Attention and Progressive Re-parameterized Batch Normalization (ICML 2024 · similarity 0.86)
- Learning in Compact Spaces with Approximately Normalized Transformer (NeurIPS 2025 · similarity 0.86)
- Impact of Layer Norm on Memorization and Generalization in Transformers (NeurIPS 2025 · similarity 0.85)
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.