Research map / Architectures and efficiency / Attention and transformers
Tree-Structured Attention with Hierarchical Accumulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use tree-based hierarchical attention instead of the strongest listed baseline | Machine translation models · IWSLT'14 En-De | BLEU: 28.43 -> 29.47 | Helped |
| use tree-based hierarchical attention instead of the strongest listed baseline | Machine translation models · IWSLT'13 En-Fr | BLEU: 43.75 -> 45.53 | Helped |
| use tree-based hierarchical attention instead of the strongest listed baseline | Machine translation models · IWSLT'14 De-En and IWSLT'13 Fr-En | BLEU: 34.72 -> 35.96; 43.08 -> 44.34 | Helped |
| use tree-based hierarchical attention instead of the strongest listed baseline | Machine translation models · WMT'14 En-De Base and Big | BLEU: 27.48 -> 28.40; 29.70 -> 29.95 | Helped |
| use tree-based attention instead of the strongest listed baseline | Text classification models · SST-5, SST-2, IMDB, and SVA | accuracy: 43.9 -> 47.4; 82.0 -> 84.3; 86.5 -> 90.1; 96.2 -> 98.0 | Helped |
| add hierarchical embeddings (ablation) | Tree Transformer · IWSLT'14 En-De | BLEU: 29.20 -> 29.47 | Helped |
| add hierarchical embeddings (ablation) | Tree Transformer · IWSLT'13 En-Fr | BLEU: 44.80 -> 45.53 | Helped |
| add hierarchical embeddings (ablation) | Tree Transformer · SST-5 | accuracy: 46.1 -> 47.4 | Helped |
| add subtree masking (ablation) | Tree Transformer · IWSLT'14 En-De | BLEU: 29.05 -> 29.47 | Helped |
| add subtree masking (ablation) | Tree Transformer · IWSLT'13 En-Fr | BLEU: 45.07 -> 45.53 | Helped |
| add subtree masking (ablation) | Tree Transformer · SST-5 | accuracy: 45.7 -> 47.4 | Helped |
| add hierarchical embeddings and subtree masking together (ablation) | Tree Transformer · IWSLT'14 En-De, IWSLT'13 En-Fr, and SST-5 | BLEU / accuracy: 28.98 -> 29.47; 44.50 -> 45.53; 45.0 -> 47.4 | Helped |
Closest papers by meaning
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.86)
- From Self-Attention to Markov Models: Unveiling the Dynamics of Generative Transformers (ICML 2024 · similarity 0.86)
- Stack Attention: Improving the Ability of Transformers to Model Hierarchical Patterns (ICLR 2024 · similarity 0.86)
- From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs (ICLR 2025 · similarity 0.86)
- AdmTree: Compressing Lengthy Context with Adaptive Semantic Trees (NeurIPS 2025 · similarity 0.86)
- Low-Rank Approximation for Sparse Attention in Multi-Modal LLMs (CVPR 2024 · similarity 0.86)
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.