Research map / Architectures and efficiency / Attention and transformers
DeLighT: Deep and Light-weight Transformer
Sachin Mehta, Marjan Ghazvininejad, Srinivasan Iyer, Luke Zettlemoyer, Hannaneh Hajishirzi
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 the Transformer with DeLighT | Machine translation models · WMT'14 En-Fr test | BLEU: 39.2 -> 40.5 (+1.3) | Improved |
| Replace the Transformer with DeLighT | Machine translation models · WMT'16 En-Ro test | BLEU: 34.3 -> 34.7 (+0.4) | Improved |
| Replace the Transformer with DeLighT | Machine translation models · WMT'14 En-De test | BLEU: 27.7 -> 27.6 (-0.1) | No clear effect |
| Replace Transformer-XL with DeLighT | Language modeling · WikiText-103 test | perplexity: 24.03 -> 24.14 | No clear effect |
| Use block-wise depth scaling instead of uniform scaling (ablation) | DeLighT · WikiText-103 validation | perplexity: 48.58 -> 46.30 | Improved |
| Increase width multiplier w_m from 2 to 4 with uniform depth (ablation) | DeLighT · WikiText-103 validation | perplexity: 48.58 -> 45.10 | Improved |
| Increase block-wise depth range from N_min=4, N_max=8 to N_min=4, N_max=12 (ablation) | DeLighT · WikiText-103 validation | perplexity: 46.30 -> 41.18 | Improved |
| Increase width multiplier w_m from 2 to 3 (ablation) | DeLighT · WikiText-103 validation | perplexity: 41.18 -> 39.92 | Improved |
| Increase model dimension d_m from 256 to 384 (ablation) | DeLighT · WikiText-103 validation | perplexity: 41.18 -> 35.14 | Improved |
| Reverse block scaling, placing deeper and wider blocks near input (ablation) | DeLighT · WikiText-103 validation | perplexity: 41.18 -> 43.10 | Hurt |
| Move DeLighT transformation after attention (ablation) | DeLighT block · WikiText-103 validation | perplexity: 34.20 -> 39.02 | Hurt |
| Move DeLighT transformation after attention and retain light-weight FFN (ablation) | DeLighT block · WikiText-103 validation | perplexity: 34.20 -> 39.43 | Hurt |
Later papers that compared with this method
2 later accepted papers report Delight as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Learning Multiscale Transformer Models for Sequence Generation (ICML 2022 · better on 1 of 1 shared result)
- Shapeshifter: a Parameter-efficient Transformer using Factorized Reshaped Matrices (NeurIPS 2021 · better on 3 of 4 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention (NeurIPS 2024 · similarity 0.86)
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.86)
- DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging (NeurIPS 2024 · similarity 0.86)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.85)
- What Does It Mean to Be a Transformer? Insights from a Theoretical Hessian Analysis (ICLR 2025 · similarity 0.85)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (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.
