Research map / Architectures and efficiency / Attention and transformers
Combiner: Full Attention Transformer with Sparse Computation Cost
Hongyu Ren, Hanjun Dai, Zihang Dai, Mengjiao Yang, Jure Leskovec, Dale Schuurmans, Bo Dai
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 sparse attention with Combiner full attention using the same axial pattern (ablation) | 6-layer image density models · CIFAR-10 | Bits/Dim: 3.666 -> 3.050 | Improved |
| replace sparse attention with Combiner full attention using the same fixed pattern (ablation) | 6-layer image density models · CIFAR-10 | Bits/Dim: 3.408 -> 3.321 | Improved |
| replace sparse attention with Combiner full attention using the same Logsparse pattern (ablation) | 6-layer image density models · CIFAR-10 | Bits/Dim: 4.253 -> 3.366 | Improved |
| add a second partition plan through a mixture (ablation) | Combiner-Axial · CIFAR-10, 6-layer | Bits/Dim: 3.050 -> 3.040 | Improved |
| replace sparse attention with Combiner full attention using the same axial pattern (ablation) | 6-layer image density models · ImageNet-64 | Bits/Dim: 4.032 -> 3.585 | Improved |
| replace sparse attention with Combiner full attention using the same fixed pattern (ablation) | 6-layer image density models · ImageNet-64 | Bits/Dim: 3.696 -> 3.654 | Improved |
| replace sparse attention with Combiner full attention using the same Logsparse pattern (ablation) | 6-layer image density models · ImageNet-64 | Bits/Dim: 4.351 -> 3.795 | Improved |
| add a second partition plan through a mixture (ablation) | Combiner-Axial · ImageNet-64, 6-layer | Bits/Dim: 3.585 -> 3.585 | No clear effect |
| use DeepSets-Max instead of max pooling for abstraction summaries (ablation) | Combiner · Wiki-40B, 8k | perplexity: 16.60 -> 16.29 | Improved |
| use DeepSets-Mean instead of mean pooling for abstraction summaries (ablation) | Combiner · Wiki-40B, 8k | perplexity: 16.54 -> 16.48 | Improved |
| use mean pooling instead of max pooling for abstraction summaries (ablation) | Combiner · Wiki-40B, 8k | perplexity: 16.60 -> 16.54 | No clear effect |
| increase context length from 2k to 8k | Combiner-Fixed · Wiki-40B | perplexity: 18.04 -> 16.60 | Improved |
Closest papers by meaning
- SEA: Sparse Linear Attention with Estimated Attention Mask (ICLR 2024 · similarity 0.89)
- Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction (ICLR 2025 · similarity 0.88)
- Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights (NeurIPS 2026 · similarity 0.88)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.88)
- ZeroS: Zero‑Sum Linear Attention for Efficient Transformers (NeurIPS 2025 · similarity 0.88)
- Pool Me Wisely: On the Effect of Pooling in Transformer-Based Models (NeurIPS 2025 · similarity 0.88)
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.
