Research map / Architectures and efficiency / Attention and transformers
Linear attention approximations
Approximates softmax attention with kernel, linear, or gated attention for linear-time complexity.
76 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Attention and transformers, keyword cluster “sequence · long · modeling”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
14 papers in 2023–24 and 47 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.4% to 4.0%.
Open questions the data points to
- Papers report Performer on WikiText-103 (perplexity) anywhere from 26.8 to 66: comparisons against it depend on whose number you take.
- Papers report Performer on LRA Pathfinder (accuracy (%)) anywhere from 68.6 to 77.05: comparisons against it depend on whose number you take.
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
- 2018 · Monotonic Chunkwise Attention (ICLR 2018 · 32 citations)
- 2020 · Real-Time Emotion Recognition via Attention Gated Hierarchical Memory Network (AAAI 2020 · 140 citations)
- 2021 · Random Feature Attention (ICLR 2021 · 116 citations)
- 2022 · FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022 · 498 citations)
- 2023 · Mega: Moving Average Equipped Gated Attention (ICLR 2023 · 37 citations)
- 2024 · Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · 14 citations)
- 2025 · Differential Transformer (ICLR 2025 · 13 citations)
- 2026 · SEMA: a Scalable and Efficient Mamba like Attention via Token Localization and Averaging (ICML 2026 · 0 citations)
Earliest papers here
- Monotonic Chunkwise Attention (ICLR 2018 · 32 citations)
- Real-Time Emotion Recognition via Attention Gated Hierarchical Memory Network (AAAI 2020 · 140 citations)
- Sparse Sinkhorn Attention (ICML 2020 · 114 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ImageNet-1K · Top-1 accuracy (higher is better)
5 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 80.9 | Stable, Fast and Accurate: Kernelized Attention with Relative Positional Encoding (NeurIPS 2021 · 7 citations) |
| 2024 | 84.5 | Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · 14 citations) |
LRA Image · accuracy (higher is better)
4 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 47.86 | Luna: Linear Unified Nested Attention (NeurIPS 2021 · 48 citations) |
| 2023 | 90.43 | Mega: Moving Average Equipped Gated Attention (ICLR 2023 · 37 citations) |
What has been tried
45 papers here report 534 measured changes (a component added, removed or swapped, compared with a baseline): helped 63%, no clear effect 11%, hurt 24%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add injective InLine attention instead of normalized linear attention | ImageNet-1K · accuracy | 77.3 -> 79.8 | Helped | Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · 14 citations) |
| Add GroupNorm to differential attention | 1.4B language model validation set · validation loss | 3.122 -> 3.062 | Helped | Differential Transformer (ICLR 2025 · 13 citations) |
| add global bidirectional scanning to causal DiG | ImageNet 256×256 · FID-50K | 175.84 -> 69.28 | Helped | DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention (CVPR 2025 · 10 citations) |
| add data-dependent gating rather than use linear attention without a gate | 340M model, 7B tokens · training perplexity | 23.21 -> 14.77 | Helped | Gated Linear Attention Transformers with Hardware-Efficient Training (ICML 2024 · 5 citations) |
| Replace sequential cross-attention with bi-directional cross-attention at 12 layers | ImageNet1K, 120 epochs · Top-1 accuracy | 73.79{\scriptstyle~±~0.32} -> 73.86{\scriptstyle~±~0.39} | No clear effect | Perceiving Longer Sequences With Bi-Directional Cross-Attention Transformers (NeurIPS 2024 · 5 citations) |
| Replace the large-kernel patch embedding with the improved convolutional patch embedding | ImageNet-1K val, 224² · top-1 accuracy | 66.1 -> 73.8 | Helped | ViG: Linear-complexity Visual Sequence Learning with Gated Linear Attention (AAAI 2025 · 4 citations) |
| Add input-dependent decay gating to ABC | Slimpajama, 340M, 10B tokens · PPL | 16.94 -> 13.51 | Helped | Gated Slot Attention for Efficient Linear-Time Sequence Modeling (NeurIPS 2024 · 4 citations) |
| replace the Hedgehog feature map with the T2R feature map while retaining attention distillation | GLUE · performance recovery (%) | 99.3 -> 93.5 | Hurt | The Hedgehog & the Porcupine: Expressive Linear Attentions with Softmax Mimicry (ICLR 2024 · 3 citations) |
| Add Sparse Attention Layer (SAL) | Taobao and Short Video · GAUC | Helped | Iterative Sparse Attention for Long-sequence Recommendation (AAAI 2025 · 2 citations) | |
| remove pretrained SD query, key, and value initialization | COCO validation 1024×1024 · FID | 30.86 -> 36.89 | Hurt | Parallel Sequence Modeling via Generalized Spatial Propagation Network (CVPR 2025 · 2 citations) |
| Increase the codebook size from S = 256 to S = 1024 | Enwik8 validation · BPB | 1.010 -> 1.000 | Helped | Transformer-VQ: Linear-Time Transformers via Vector Quantization (ICLR 2024 · 2 citations) |
| add local exact polynomial attention | PG19 test, 32k context · perplexity | 12.19 -> 11.34 | Helped | PolySketchFormer: Fast Transformers via Sketching Polynomial Kernels (ICML 2024 · 2 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Performer | WikiText-103 | perplexity | 26.8 – 66 | 2 papers, e.g. ICLR 2024 and NeurIPS 2021 |
| Performer | LRA Pathfinder | accuracy (%) | 68.6 – 77.05 | 2 papers, e.g. NeurIPS 2021 and ICLR 2025 |
| Reformer | WikiText-103 | perplexity | 25.6 – 27.68 | 2 papers, e.g. ICLR 2024 and NeurIPS 2021 |
What reviewers will expect
Datasets and baselines named in the 25 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet-1K · 6 of 25 papers
- ImageNet · 3 of 25 papers
- WikiText-103 · 3 of 25 papers
- PiQA · 2 of 25 papers
- ARC-E · 2 of 25 papers
- Winogrande · 2 of 25 papers
- MQAR · 2 of 25 papers
- SQuAD · 2 of 25 papers
Baselines
- Mamba-2 · 5 of 25 papers
- Mamba · 4 of 25 papers
- Performer · 4 of 25 papers
- Transformer · 4 of 25 papers
- DeltaNet · 3 of 25 papers
- cosFormer · 3 of 25 papers
- FlashAttention-2 · 3 of 25 papers
- DIN · 2 of 25 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 61 papers; from OpenAlex and the papers' affiliations).
- Google · 4 papers
- Tencent · 3 papers
- Carnegie Mellon University · 3 papers
- Tsinghua University · 2 papers
- Huazhong University of Science & Technology · 2 papers
- NVIDIA · 2 papers
- Meta · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Stanford University · 2 papers
- Shanghai AI Lab · 2 papers
Use it
Download BibTeX of 33 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
- State space and recurrent sequence models (172 papers)
- Sparse efficient transformers for long context (76 papers)
- Mamba architectures (40 papers)
Elsewhere in the topic
- Transformer expressivity and learnability (217 papers)
- Sparse and structured attention (163 papers)
- Sparse attention for long-context LLMs (147 papers)
- Attention head and block redesign (127 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)
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).