Research map / Architectures and efficiency / Attention and transformers
Focused Hierarchical RNNs for Conditional Sequence Processing
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 a question-conditioned focused hierarchical encoder | FHE · SearchQA | F1: 55.3 -> 56.7 validation; 51.9 -> 53.4 test; AMANDA test 56.6 | Mixed |
| use a question-conditioned focused hierarchical encoder | FHE · MS MARCO | Bleu-1: 24.8 -> 27.3 validation; 28 -> 30 test | Helped |
| use a question-conditioned focused hierarchical encoder | FHE · MS MARCO | Rouge-L: 26.5 -> 26.7 validation; 28 -> 30 test | Mixed |
| use FHE instead of LSTM2 | Pixel-by-Pixel MNIST QA · Pixel-by-Pixel MNIST QA validation | accuracy: 98.4 -> 99.1 | Helped |
| add question-context elementwise product to the gate features (ablation) | FHE · MS MARCO validation | Bleu-1: 18.5 -> 27.3 | Helped |
| add question-context elementwise product to the gate features (ablation) | FHE · MS MARCO validation | Rouge-L: 19.3 -> 26.7 | Helped |
| add pointer softmax (ablation) | FHE · MS MARCO validation | Bleu-1: 20.5 -> 27.3 | Helped |
| add pointer softmax (ablation) | FHE · MS MARCO validation | Rouge-L: 18.7 -> 26.7 | Helped |
| use learned boundaries rather than non-learned boundaries (ablation) | FHE · MS MARCO validation | Bleu-1: 23.5 -> 27.3 | Helped |
| use learned boundaries rather than non-learned boundaries (ablation) | FHE · MS MARCO validation | Rouge-L: 24 -> 26.7 | Helped |
| train on sequences of length 200, then use FHE on longer sequences | FHE-fixed · Picking task, length 400 | accuracy: 61.4 -> 97.6 | Helped |
| train on sequences of length 200, then use FHE on longer sequences | FHE-fixed · Picking task, length 800 | accuracy: 39.7 -> 95.6 | Helped |
Closest papers by meaning
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.89)
- MesaNet: Sequence Modeling by Locally Optimal Test-Time Training (ICLR 2026 · similarity 0.89)
- Hierarchically Gated Recurrent Neural Network for Sequence Modeling (NeurIPS 2023 · similarity 0.89)
- RNN-Based Sequence-Preserved Attention for Dependency Parsing (AAAI 2018 · similarity 0.88)
- Word Attention for Sequence to Sequence Text Understanding (AAAI 2018 · similarity 0.88)
- Explaining Modern Gated-Linear RNNs via a Unified Implicit Attention Formulation (ICLR 2025 · similarity 0.87)
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.