Research map / Architectures and efficiency / Attention and transformers
Sparse attention for long-context LLMs
Selects or compresses tokens and keys to cut attention cost in long-context LLM inference.
147 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Attention and transformers, keyword cluster “global · local · spatial”
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.
13 papers in 2023–24 and 133 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.2% to 11.4%.
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
- 2021 · EL-Attention: Memory Efficient Lossless Attention for Generation (ICML 2021 · 2 citations)
- 2023 · Personalized Dialogue Generation with Persona-Adaptive Attention (AAAI 2023 · 25 citations)
- 2024 · MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · 29 citations)
- 2025 · MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 2025 · 8 citations)
- 2026 · SlimInfer: Accelerating Long-Context LLM Inference via Dynamic Token Pruning (AAAI 2026 · 1 citation)
Earliest papers here
- EL-Attention: Memory Efficient Lossless Attention for Generation (ICML 2021 · 2 citations)
- Personalized Dialogue Generation with Persona-Adaptive Attention (AAAI 2023 · 25 citations)
- Focused Transformer: Contrastive Training for Context Scaling (NeurIPS 2023 · 13 citations)
What has been tried
47 papers here report 536 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 9%, 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 |
|---|---|---|---|---|
| Replace dynamically built sparse indices with static indices for Vertical-Slash and Block-Sparse heads | InfiniteBench, LLaMA-3-8B-262K · Avg. score | 38.8 -> 31.9 | Hurt | MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · 29 citations) |
| Insert LayerExpert in every layer instead of every other layer | Ego4D Narration Stream Validation · Fluency | 45.2% -> 42.5% | Hurt | VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation (NeurIPS 2024 · 9 citations) |
| Use finer-grained block segmentation at consistent 75% attention sparsity | Validation set · LM loss | 1e-2 | Helped | MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 2025 · 8 citations) |
| add CLS-MS multi-task training to CAR | GlobalVoice Arabic-to-English · ROUGE-2 | 6.41 -> 5.00 | Hurt | CAR-Transformer: Cross-Attention Reinforcement Transformer for Cross-Lingual Summarization (AAAI 2024 · 7 citations) |
| Replace the Taylor feature map with CosFormer | Pile language modeling · quality | Mixed | Simple linear attention language models balance the recall-throughput tradeoff (ICML 2024 · 4 citations) | |
| Add SMR and LMR | MSC test · BLEU-avg | 17.32 -> 19.33 | Helped | StreamingDialogue: Prolonged Dialogue Learning via Long Context Compression with Minimal Losses (NeurIPS 2024 · 3 citations) |
| Use token-choice routing instead of expert-choice routing | Few-shot average accuracy, 16.5e18 FLOPs · accuracy | 42.6 -> 40.0 | Hurt | Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation (NeurIPS 2025 · 2 citations) |
| Use fine-grained interleaved beacon tokens instead of appending all beacons at chunk end | Single-Doc QA, LongBench, Qwen-2-7B, 32K, x4 · score | 35.2 -> 40.5 | Helped | Long Context Compression with Activation Beacon (ICLR 2025 · 2 citations) |
| Fuse attention and SSM heads in parallel rather than sequentially | Commonsense reasoning · accuracy | +1.08% | Helped | Hymba: A Hybrid-head Architecture for Small Language Models (ICLR 2025 · 2 citations) |
| Add centering before LSH sampling | RULER retrieval and FWE tasks · Accuracy | Without centering, the accuracy drops to almost zero in retrieval (NIAH) and degrades to 65% in FWE. | Helped | MagicPIG: LSH Sampling for Efficient LLM Generation (ICLR 2025 · 2 citations) |
| Add adaptive selection of SAGEAttn-vB for suitable layers, using SAGEAttn-B otherwise | CogvideoX attention · TOPS | 292.17 -> 327.57 | Helped | SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · 2 citations) |
| Use pre-rotary rather than post-rotary PCA transformation | Llama3-8B and Mistral-7B · perplexity | about 5 for the full attention to over 10 | Helped | Loki: Low-rank Keys for Efficient Sparse Attention (NeurIPS 2024 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 75 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- RULER · 16 of 75 papers
- LongBench · 15 of 75 papers
- GSM8K · 11 of 75 papers
- PG-19 · 8 of 75 papers
- SlimPajama · 7 of 75 papers
- C4 · 6 of 75 papers
- Needle-in-a-Haystack · 6 of 75 papers
- TriviaQA · 6 of 75 papers
Baselines
- StreamingLLM · 14 of 75 papers
- H2O · 8 of 75 papers
- Quest · 8 of 75 papers
- MInference · 7 of 75 papers
- SnapKV · 6 of 75 papers
- Mamba2 · 5 of 75 papers
- InfLLM · 4 of 75 papers
- FlashAttention2 · 4 of 75 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 146 papers; from OpenAlex and the papers' affiliations).
- Amazon · 7 papers
- Google · 6 papers
- University of California, Berkeley · 5 papers
- NVIDIA · 5 papers
- Meta · 5 papers
- Tsinghua University · 5 papers
- Korea Advanced Institute of Science and Technology · 4 papers
- Massachusetts Institute of Technology · 4 papers
- Google DeepMind · 4 papers
- National University of Singapore · 4 papers
Use it
Download BibTeX of 23 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
- Block-sparse attention kernels (60 papers)
- KV cache eviction and compression (92 papers)
Elsewhere in the topic
- Transformer expressivity and learnability (217 papers)
- State space and recurrent sequence models (172 papers)
- Sparse and structured attention (163 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).