Research map / Architectures and efficiency / Attention and transformers
Staircase Attention for Recurrent Processing of Sequences
Da JU, Stephen Roller, Sainbayar Sukhbaatar, Jason E Weston
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 |
|---|---|---|---|
| increase recurrent steps from 1 to 8 (ablation) | Staircase · Pushshift.io Reddit valid | perplexity: 26.4 -> 22.9 | Improved |
| use Staircase with recurrence instead of Transformer-XL | Staircase · Pushshift.io Reddit test | perplexity: 26.2 -> 22.6 | Improved |
| use Staircase instead of Transformer-XL | Staircase · BASE Data valid | perplexity: 28.0 -> 23.0 | Improved |
| use Ladder instead of Transformer-XL | Ladder · Enwik8 test | bpc: 1.15 -> 1.11 | Improved |
| use Staircase instead of Transformer-XL on state tracking | Staircase · Random Walk | error %: 84.1 -> 0.1 | Improved |
| use Staircase instead of Transformer-XL on Algorithm task | Staircase · Algorithm | error %: 48.7 -> 0.2 | Improved |
| use Ladder (no time recurrence) instead of Staircase on Algorithm task (ablation) | Ladder · Algorithm | error %: 0.2 -> 45.1 | Hurt |
| use Cached Staircase instead of Staircase on Algorithm task (ablation) | Cached Staircase · Algorithm | error %: 0.2 -> 1.2 | Hurt |
| double depth of Transformer-XL (218M params) vs Staircase 8 steps at 117M (ablation) | Transformer-XL (2x) · Pushshift.io Reddit valid | perplexity: 23.7 vs 22.9 | Hurt |
| replace Staircase with Universal Transformer (shared all layers, 29M) (ablation) | Universal Transformer · Pushshift.io Reddit valid | perplexity: 25.0 -> 39.9 | Hurt |
| reduce forward size from 128 to 64 at 2 recurrent steps (step size 256 to 128) (ablation) | Staircase · Pushshift.io Reddit valid | perplexity: 25.0 -> 25.2 | Hurt |
| reduce forward size from 64 to 32 at 4 recurrent steps (ablation) | Staircase · Pushshift.io Reddit valid | perplexity: 23.7 -> 23.8 | Hurt |
Closest papers by meaning
- Forgetting Transformer: Softmax Attention with a Forget Gate (ICLR 2025 · similarity 0.88)
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.86)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.86)
- Scaling Stick-Breaking Attention: An Efficient Implementation and In-depth Study (ICLR 2025 · similarity 0.86)
- Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers (Student Abstract) (AAAI 2024 · similarity 0.86)
- Curse of High Dimensionality Issue in Transformer for Long Context Modeling (ICML 2025 · 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.
