Research map / Architectures and efficiency / Attention and transformers
Stronger Normalization-Free Transformers
Mingzhi Chen, Taiming Lu, Jiachen Zhu, Mingjie Sun, Zhuang Liu
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 Derf | ViT-B · ImageNet-1K | top-1 accuracy: 82.3% -> 82.8% | Improved |
| replace LayerNorm with Derf | DiT-XL/2 · ImageNet-1K generation | FID: 19.94 -> 18.92 | Improved |
| replace LayerNorm with DyT | DiT-XL/2 · ImageNet-1K generation | FID: 19.94 -> 20.83 | Hurt |
| replace LayerNorm with Derf | HyenaDNA · GenomicBenchmarks | accuracy: 85.2% -> 85.7% | Improved |
| replace LayerNorm with Derf | wav2vec 2.0 Base · LibriSpeech | validation loss: 1.95 -> 1.93 | Improved |
| replace LayerNorm with Derf | GPT-2 124M · OpenWebText | validation loss: 2.94 -> 2.94 | No clear effect |
| replace LayerNorm with Derf | ViT-B · ImageNet-1K train set (eval mode) | training loss: 0.2623 -> 0.2681 | Hurt |
| add learnable shift s to erf (ablation) | Derf · ImageNet-1K ViT-B | top-1 accuracy: 82.6% -> 82.8% | Improved |
| use per-channel vector s instead of scalar (ablation) | Derf · ImageNet-1K ViT-B | top-1 accuracy: 82.8% vs 82.8% | No clear effect |
| scale tanh by 1.205 to approximate erf (ablation) | DyT · ImageNet-1K ViT-B | top-1 accuracy: 82.6% -> 82.7% | No clear effect |
| shift function off zero center (|shift| >= 2 fails) (ablation) | point-wise functions · ImageNet-1K ViT-B | top-1 accuracy: 82.5% -> 81.6% (erf vertical +1) | Hurt |
| make function unbounded (mix in linear term) (ablation) | erf, tanh · ImageNet-1K ViT-B | top-1 accuracy: 82.6% -> 82.4% (erf, 0.01 linear) | Hurt |
Closest papers by meaning
- Transformers without Normalization (CVPR 2025 · similarity 0.90)
- Learning in Compact Spaces with Approximately Normalized Transformer (NeurIPS 2025 · similarity 0.85)
- Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.85)
- Dataset Distillation with Neural Characteristic Function: A Minmax Perspective (CVPR 2025 · similarity 0.85)
- Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion (ICLR 2024 · similarity 0.85)
- On the Importance of Gaussianizing Representations (ICML 2025 · similarity 0.84)
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.
