Research map / Architectures and efficiency / Attention and transformers
Searching for Efficient Transformers for Language Modeling
David So, Wojciech Mańke, Hanxiao Liu, Zihang Dai, Noam Shazeer, Quoc V Le
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 |
|---|---|---|---|
| Add squared ReLU activations and depthwise convolution after Q, K, and V projections | Transformer (Primer-EZ) · LM1B, Tensor2Tensor TPUv2 | Speedup: 2.34 +/- 0.04 | Improved |
| Add squared ReLU activations and depthwise convolution after Q, K, and V projections | Transformer (Primer-EZ) · C4, 110M parameters | PPLX: 20.61 -> 19.83 | Improved |
| Add Primer-EZ modifications to Switch Transformer | Switch Transformer · C4, 550M parameters | PPLX: 17.16 -> 16.56 | Improved |
| Add squared ReLU | Synthesizer · C4, 145M parameters | PPLX: 20.35 -> 19.55 | Improved |
| Replace ReLU with squared ReLU (ablation) | Transformer feed-forward block · Autoregressive language modeling | perplexity | Improved |
| Add 3x1 depthwise convolution after each Q, K, and V projection (ablation) | Transformer self-attention · Autoregressive language modeling | perplexity | Improved |
| Replace the Q/K projections with a shared Q and K depthwise representation (ablation) | Primer · Autoregressive language modeling | performance | Hurt |
| Apply the second normalization after the feed-forward transformation (ablation) | Primer · Autoregressive language modeling | performance | Mixed |
| Use custom normalization with x(x - μ) (ablation) | Primer · Autoregressive language modeling | performance | Mixed |
| Increase the feed-forward upward projection to 12X while controlling for parameters (ablation) | Transformer · Small models (~35M parameters) | perplexity | Improved |
| Increase the feed-forward upward projection to 12X while controlling for parameters (ablation) | Transformer · Larger models | performance | No clear effect |
| Add per-channel learnable scalars after attention softmax (ablation) | Primer · Fixed-length language modeling | perplexity | Improved |
Closest papers by meaning
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.90)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.89)
- Progressive Residual Warmup for Language Model Pretraining (NeurIPS 2026 · similarity 0.89)
- Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers (NeurIPS 2024 · similarity 0.88)
- Dynamic Convolutions Improve Transformers (NeurIPS 2026 · similarity 0.88)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · similarity 0.88)
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.
