Research map / Architectures and efficiency / Attention and transformers
Attention head mechanisms
Studies attention heads, induction circuits, and positional encodings behind length generalization.
75 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Attention and transformers, keyword cluster “vision · vit · transformers”
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.
15 papers in 2023–24 and 57 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.6% to 4.9%.
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
- 2019 · A Task in a Suit and a Tie: Paraphrase Generation with Semantic Augmentation (AAAI 2019)
- 2021 · On Scalar Embedding of Relative Positions in Attention Models (AAAI 2021 · 4 citations)
- 2022 · Attention-based Neural Cellular Automata (NeurIPS 2022 · 4 citations)
- 2023 · Representational Strengths and Limitations of Transformers (NeurIPS 2023 · 13 citations)
- 2024 · AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · 9 citations)
- 2025 · On the Power of Convolution-Augmented Transformer (AAAI 2025 · 1 citation)
- 2026 · Induction Heads Interpolate N-Grams (ICML 2026 · 0 citations)
Earliest papers here
- A Task in a Suit and a Tie: Paraphrase Generation with Semantic Augmentation (AAAI 2019)
- On Scalar Embedding of Relative Positions in Attention Models (AAAI 2021 · 4 citations)
- Attention-based Neural Cellular Automata (NeurIPS 2022 · 4 citations)
What has been tried
43 papers here report 414 measured changes (a component added, removed or swapped, compared with a baseline): helped 51%, no clear effect 12%, hurt 31%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace the ε-rule with the γ-rule in ViT linear layers | ImageNet · faithfulness | 2.53 -> 6.06 | Helped | AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · 9 citations) |
| Ignore bottom 1% of dimensions with the lowest LN γ when calculating FF contextualization | Wikipedia · contextualization change (1 - ρ), layer average | 0.21 to 0.09 | Hurt | Analyzing Feed-Forward Blocks in Transformers through the Lens of Attention Maps (ICLR 2024 · 7 citations) |
| Patch corrupted-state activations at the third-move target square in layer 10 | Filtered Lichess puzzles · mean log-odds reduction | 1.88 ± 0.04 | Hurt | Evidence of Learned Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2024 · 3 citations) |
| add dynamic projection and gating instead of static projection | language modeling · perplexity | 11.17 -> 10.83 | Helped | Improving Transformers with Dynamically Composable Multi-Head Attention (ICML 2024 · 2 citations) |
| add RoPE to MH-CAT | Wikitext · perplexity | 27.417 -> 25.858 | Helped | On the Power of Convolution-Augmented Transformer (AAAI 2025 · 1 citation) |
| Add Fourier Series (FS) to FoPE | Gutenberg Books; length generalization · perplexity and Passkey Retrieval accuracy | Helped | Fourier Position Embedding: Enhancing Attention’s Periodic Extension for Length Generalization (ICML 2025 · 1 citation) | |
| Add LayerNorm to the attention block | Increasing context length · stable rank | No clear effect | Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers (ICML 2025 · 1 citation) | |
| Remove long-term memory after training | Persistent-knowledge benchmarks (13-task avg) · average score | 56.8 vs 56.6 | No clear effect | Memory Mosaics at scale (NeurIPS 2025 · 1 citation) |
| Repeat the relevant token B times within the context | Single-location linear regression · plateau length | T_\mathrm{plateau} = 1.51 \, T^{\,0.99} \, B^{-0.99} \, d^{\,0.49} | Helped | The emergence of sparse attention: impact of data distribution and benefits of repetition (NeurIPS 2025 · 1 citation) |
| Zero the highest-composition component in an inhibition head | IOI · inhibition score | reduces the Inhibition score by 7-14% | Hurt | Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · 1 citation) |
| add the dual-run design (run twice, using saved head outputs) | Parametric datasets average, Type 3 (coherent conflict), Gemma · accuracy (%) | 52.7 -> 73.4 | Helped | Taming Knowledge Conflicts in Language Models (ICML 2025 · 0 citations) |
| Increase context length | Linear manifolds · MSE | Helped | In-Context Denoising with One-Layer Transformers: Connections between Attention and Associative Memory Retrieval (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 38 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- C4 · 4 of 38 papers
- SlimPajama · 2 of 38 papers
- PG19 · 2 of 38 papers
- RULER · 2 of 38 papers
- Fineweb-Edu · 2 of 38 papers
- The Pile · 2 of 38 papers
Baselines
- linear attention · 5 of 38 papers
- ALiBi · 4 of 38 papers
- NoPE · 3 of 38 papers
- RoPE · 3 of 38 papers
- Mamba · 2 of 38 papers
- GLA · 2 of 38 papers
- FIRE · 2 of 38 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 72 papers; from OpenAlex and the papers' affiliations).
- Microsoft · 4 papers
- University of California, Berkeley · 3 papers
- Meta · 3 papers
- New York University · 3 papers
- Yale University · 2 papers
- Beijing University of Posts and Telecommunications · 2 papers
- Google DeepMind · 2 papers
- University of Illinois, Urbana Champaign · 2 papers
- Flatiron Institute · 2 papers
- Harvard University · 2 papers
Use it
Download BibTeX of 24 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
- Transformer expressivity and learnability (217 papers)
- Training dynamics of in-context learning (62 papers)
- Vision transformer design (30 papers)
Elsewhere in the topic
- State space and recurrent sequence models (172 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).