Research map / Architectures and efficiency / Attention and transformers
RAT+: Train Dense, Infer Sparse - Recurrence Augmented Attention for Dilated Inference
Xiuying Wei, Caglar Gulcehre
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 |
|---|---|---|---|
| add recurrence to dense attention (ablation) | attention 1B from scratch · Val PPL T=4096 | perplexity: 7.44 -> 7.38 | Improved |
| add active recurrence learning (joint sparse+dense training) (ablation) | RAT+ L=T · Val PPL at D=64 | perplexity: 9.96 -> 7.66 | Improved |
| use fixed overlapped chunk L=64 instead of L=D at inference (ablation) | RAT+ · Val PPL at D=1 | perplexity: 8.39 -> 7.4 | Improved |
| per-batch joint training instead of 0.5/0.5 loss mixing (ablation) | RAT+ · Val PPL at D=64 | perplexity: 7.82 -> 7.66 | Improved |
| add 1B-token resolution adaptation (ablation) | RAT+ L=T L*=64 · Val PPL at D=16 | perplexity: 8.03 -> 7.52 | Improved |
| add recurrence on attention keys (ablation) | RAT+ 200M · PG19 PPL at D=1 | perplexity: 11.68 -> 10.71 | Improved |
| apply recurrence on qkv inputs instead of k,v (ablation) | RAT+ 200M · PG19 PPL at D=128 | perplexity: 11.39 -> 12.09 | Hurt |
| replace RoPE with NoPE (ablation) | RAT+ · Val PPL at D=64 | perplexity: 7.66 -> 7.63 | No clear effect |
| add recurrence to top-k block attention (RAT+ vs attention) (ablation) | Quest D=64 K=16 · NIAH-MK-2 | accuracy: 63.2 -> 93.8 | Improved |
| combine dilation with top-k block (ablation) | Quest D=64 K=8 on RAT+ · NIAH-MK-2 | accuracy: 57.4 -> 97.4 | Improved |
| combine dilation with top-k block (ablation) | Quest D=64 K=8 on RAT+ · NIAH-MK-3 | accuracy: 11.0 -> 79.6 | Improved |
| increase dilation from 1 to 64 | RAT+ 1.5B · commonsense avg | accuracy: 59.21 -> 57.46 | Hurt |
Closest papers by meaning
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.92)
- Depth-Recurrent Attention Mixtures: Giving Latent Reasoning the Attention it Deserves (NeurIPS 2026 · similarity 0.89)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.86)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.86)
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.86)
- Recursive Inference Scaling: A Winning Path to Scalable Inference in Language and Multimodal Systems (NeurIPS 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.
