Research map / Architectures and efficiency / Attention and transformers
Sparse and structured attention
Redesigns self-attention with sparsity, structure, and rank analysis for efficiency and interpretability.
163 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Attention and transformers, keyword cluster “kv · cache · long context”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
26 papers in 2023–24 and 65 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 4.5% to 5.6%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · A Convolutional Attention Network for Extreme Summarization of Source Code (ICML 2016 · 135 citations)
- 2017 · Bidirectional Attention Flow for Machine Comprehension (ICLR 2017 · 1,592 citations)
- 2018 · Hierarchical Recurrent Attention Network for Response Generation (AAAI 2018 · 151 citations)
- 2019 · End-To-End Multi-Task Learning With Attention (CVPR 2019 · 1,203 citations)
- 2020 · Exploring Self-Attention for Image Recognition (CVPR 2020 · 958 citations)
- 2021 · Scaling Local Self-Attention for Parameter Efficient Visual Backbones (CVPR 2021 · 461 citations)
- 2022 · On the Integration of Self-Attention and Convolution (CVPR 2022 · 551 citations)
- 2023 · Neighborhood Attention Transformer (CVPR 2023 · 474 citations)
- 2024 · An Attentive Inductive Bias for Sequential Recommendation beyond the Self-Attention (AAAI 2024 · 124 citations)
- 2025 · Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · 3 citations)
- 2026 · Ordinary Least Squares as an Attention Mechanism (NeurIPS 2026 · 2 citations)
Earliest papers here
- A Convolutional Attention Network for Extreme Summarization of Source Code (ICML 2016 · 135 citations)
- Can Active Memory Replace Attention? (NeurIPS 2016 · 16 citations)
- Learning to Generate with Memory (ICML 2016 · 14 citations)
What has been tried
104 papers here report 1,145 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 8%, hurt 19%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove attentive inductive bias and use only self-attention | Beauty, Sports, Toys, Yelp, LastFM, ML-1M · HR@20 | Beauty: 0.1373 -> 0.1265; Sports: 0.0858 -> 0.0779 | Hurt | An Attentive Inductive Bias for Sequential Recommendation beyond the Self-Attention (AAAI 2024 · 124 citations) |
| Add residual self-attention | YFCC100M and ScanNet · AUC@20° | 82.00 -> 82.15; 57.16 -> 57.97 | Helped | ResMatch: Residual Attention Learning for Feature Matching (AAAI 2024 · 13 citations) |
| Restrict attention to positional information rather than semantic information | Histogram task test set · accuracy | 0.99885 ± 0.0004; 0.99895 ± 0.0012 | No clear effect | A Phase Transition between Positional and Semantic Learning in a Solvable Model of Dot-Product Attention (NeurIPS 2024 · 4 citations) |
| Add λ₂ weighting to attention-head averaging | ImageNet Segmentation · Acc | 84.00 -> 84.12 | Helped | Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · 3 citations) |
| use continuous linear attention instead of the discrete mixer | Radial kernel learning OOD1 · operator test error | 25.81% -> 8.10% | Helped | Nonlocal Attention Operator: Materializing Hidden Knowledge Towards Interpretable Physics Discovery (NeurIPS 2024 · 3 citations) |
| replace Rotary with another positional embedding | Pile-CC validation · Sink₁ᵋ (%) | Rotary 18.18; NoPE 20.35; Absolute PE 32.73; Learnable PE 33.13; Relative PE 35.53; ALiBi 20.78 | No clear effect | When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · 2 citations) |
| Reduce the prediction block size b_Q | Image completion and meta-regression · log-likelihood | Helped | Memory Efficient Neural Processes via Constant Memory Attention Block (ICML 2024 · 1 citation) | |
| Increase PAP iterations from 4 to 6 at the first layer | ImageNet-1K val · Top-1 accuracy | 70.94 -> 71.49 | Helped | Unveiling the Hidden Structure of Self-Attention via Kernel Principal Component Analysis (NeurIPS 2024 · 1 citation) |
| Apply full LayerNorm in every layer | Darcy flow, 441 tokens, 2 layers · test error | 3.09% -> 3.45% | Hurt | Neural Interpretable PDEs: Harmonizing Fourier Insights with Attention for Scalable and Interpretable Physics Discovery (ICML 2025 · 0 citations) |
| remove the CBSA contraction step | ImageNet-1K · top-1 accuracy | 63.2 -> 63.8; 71.4 -> 71.8 | Helped | Towards Interpretable and Efficient Attention: Compressing All by Contracting a Few (NeurIPS 2025 · 0 citations) |
| Remove the attention bias | ImageNet-1K · Acc@1 | 87.144% -> 9.376% | Hurt | FlashBias: Fast Computation of Attention with Bias (NeurIPS 2025 · 0 citations) |
| add feature simulation | Books3, training length 1024 · perplexity | 22.43 -> 22.14 | Helped | SAS: Simulated Attention Score (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 30 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 2 of 30 papers
- ImageNet-1K · 2 of 30 papers
- Wikitext · 2 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 91 papers; from OpenAlex and the papers' affiliations).
- The University of Hong Kong · 4 papers
- National University of Singapore · 4 papers
- Meta · 3 papers
- Shanghai Jiao Tong University · 2 papers
- Lehigh University · 2 papers
- Wuhan University · 2 papers
- Georgia Institute of Technology · 2 papers
- University of California, Riverside · 2 papers
- University of Illinois, Urbana Champaign · 2 papers
- New York University · 2 papers
Use it
Download BibTeX of 39 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Attention head and block redesign (127 papers)
- Efficient image transformers (84 papers)
- Long-context KV and memory attention (88 papers)
- Transformer neural operators for PDEs (69 papers)
Elsewhere in the topic
- Transformer expressivity and learnability (217 papers)
- State space and recurrent sequence models (172 papers)
- Sparse attention for long-context LLMs (147 papers)
- Efficient and robust vision transformers (124 papers)
- Efficient language transformers (105 papers)
- Language model pretraining and induction (99 papers)
- Hybrid and interpretable ViT designs (94 papers)
- KV cache eviction and compression (92 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).