Research map / Architectures and efficiency / Attention and transformers
Magneto: A Foundation Transformer
Hongyu Wang, Shuming Ma, Shaohan Huang, Li Dong, Wenhui Wang, Zhiliang Peng, Yu Wu, Payal Bajaj, Saksham Singhal, Alon Benhaim, Barun Patra, Zhun Liu and 3 more
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 |
|---|---|---|---|
| Sub-LN plus DeepNet-style initialization instead of Pre-LN | GPT-style causal LM (72L) · Zero-shot, one-shot, four-shot in-context learning | Avg. accuracy: +1.7 zero-shot vs Pre-LN (65.8 vs 64.1); +0.4 vs Normformer zero-shot | Improved |
| Sub-LN plus DeepNet-style initialization | BERT-base masked LM · GLUE dev | Avg. score: 86.3 vs 85.7 Post-LN, 84.6 Pre-LN | Improved |
| Sub-LN plus DeepNet-style initialization | 18L-18L encoder-decoder NMT · OPUS-100 | BLEU: +0.5 over Pre-LN (31.0 vs 30.5) | Improved |
| Post-LN instead of Sub-LN in 18L-18L NMT | encoder-decoder NMT · OPUS-100 | BLEU: did not converge | Hurt |
| Sub-LN plus DeepNet-style initialization for BEiT-style ViT | ViT-Base and ViT-Large (masked image modeling) · ImageNet-1k val top-1 and ADE20k mIoU | top-1 accuracy / mIoU: +0.4 top-1 at 12L, +0.6 at 24L; +0.8 mIoU at 12L, +0.4 at 24L | Improved |
| Sub-LN plus DeepNet-style initialization | Transformer Transducer on LibriSpeech · LibriSpeech 960h without LM, 18L and 36L | WER (lower is better): test-other 6.16 vs 6.62 (18L); 5.56 vs 6.04 (36L) | Improved |
| Sub-LN plus DeepNet-style initialization | BEiT-3-style vision-language pretraining (24L base) · VQA test-dev and NLVR2 test-P | VQA score / accuracy: +0.63 VQA test-dev; +0.54 NLVR2 test-P | Improved |
| Higher learning rate enabled by Sub-LN+DeepNet init (1e-3 vs diverged Pre-LN and Normformer) at 24L (ablation) | GPT-style causal LM 24L · Zero-shot avg accuracy | Avg. accuracy: MAGNETO 61.4 at lr 1e-3; Pre-LN and Normformer diverged | Improved |
| Sub-LN plus DeepNet-style initialization with larger learning rate (36L LibriSpeech) | Transformer Transducer 36L · LibriSpeech training stability | peak learning rate: MAGNETO up to 3e-3 vs Transformer 1.5e-3 | Improved |
Closest papers by meaning
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.87)
- A Multi-Level Framework for Accelerating Training Transformer Models (ICLR 2024 · similarity 0.86)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.86)
- GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling (NeurIPS 2025 · similarity 0.86)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.86)
- HaploVL: A Single-Transformer Baseline for Multi-Modal Understanding (ICML 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.
