Research map / Architectures and efficiency / Attention and transformers
Compositional De-Attention Networks
Yi Tay, Anh Tuan Luu, Aston Zhang, Shuohang Wang, Siu Cheung Hui
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 context-query attention with CoDA | DecaProp · Quasar-T | EM: 38.6 -> 41.3 | Improved |
| replace attention with CoDA | DecaProp · SearchQA | EM: 56.8 -> 57.2 | No clear effect |
| replace attention with CoDA | DecompAtt · WikiQA | MAP: 66.4 -> 70.5 | Improved |
| replace attention with CoDA | DecompAtt · TrecQA | MAP: 80.6 -> 80.0 | Hurt |
| replace attention with CoDA | DecompAtt · SNLI | accuracy: 84.61 -> 85.71 | Improved |
| replace attention with CoDA | Transformer · IWSLT'15 En-Vi | BLEU: 28.43 -> 29.84 | Improved |
| replace attention with CoDA | Transformer · MLU arithmetic | sequence accuracy: 76.1 -> 84.3 | Improved |
| replace attention with CoDA | Transformer · AlgoLisp | sequence accuracy: 96.8 -> 97.7 | Improved |
| replace attention with CoDA | Transformer · IMDb and SST | accuracy: 82.6 -> 83.3 (IMDb), 78.9 -> 80.6 (SST) | Improved |
| use tanh(N) instead of sigmoid(N) as gate (ablation) | CoDA · IWSLT'15 En-Vi dev | BLEU: 27.8 -> 23.1 | Hurt |
| use arctan(N) instead of sigmoid(N) as gate (ablation) | CoDA · IWSLT'15 En-Vi dev | BLEU: 27.8 -> 21.9 | Hurt |
| use sigmoid(E) instead of tanh(E) (ablation) | CoDA · IWSLT'15 En-Vi dev | BLEU: 27.8 -> 27.5 | Hurt |
Closest papers by meaning
- Compositional Attention: Disentangling Search and Retrieval (ICLR 2022 · similarity 0.88)
- Attention as a Hypernetwork (ICLR 2025 · similarity 0.85)
- In-Context Compositional Learning vis Sparse Coding Transformer (NeurIPS 2025 · similarity 0.85)
- Order-Level Attention Similarity Across Language Models: A Latent Commonality (NeurIPS 2025 · similarity 0.84)
- Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · similarity 0.84)
- Scaling can lead to compositional generalization (NeurIPS 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.
