Research map / Architectures and efficiency / Attention and transformers
Efficient language transformers
Searches and streamlines transformer architectures for language modeling, translation, and stable training.
105 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Attention and transformers, keyword cluster “transformers · language · icl”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
23 papers in 2023–24 and 36 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 3.9% to 3.1%.
Open questions the data points to
- Papers report Adaptive Inputs on WIKITEXT-103 test (perplexity) anywhere from 18.65 to 24: comparisons against it depend on whose number you take.
No longer reported by papers here: WMT14 En-De (BLEU, last in 2023).
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 · Layer-Wise Coordination between Encoder and Decoder for Neural Machine Translation (NeurIPS 2018 · 115 citations)
- 2019 · Insertion Transformer: Flexible Sequence Generation via Insertion Operations (ICML 2019 · 154 citations)
- 2020 · Meshed-Memory Transformer for Image Captioning (CVPR 2020 · 1,212 citations)
- 2021 · Share or Not? Learning to Schedule Language-Specific Capacity for Multilingual Translation (ICLR 2021 · 49 citations)
- 2022 · Improving Language Models by Retrieving from Trillions of Tokens (ICML 2022 · 298 citations)
- 2023 · Mass-Editing Memory in a Transformer (ICLR 2023 · 54 citations)
- 2024 · MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · 171 citations)
- 2025 · TAMER: Tree-Aware Transformer for Handwritten Mathematical Expression Recognition (AAAI 2025 · 16 citations)
- 2026 · XSpecMesh: Quality-Preserving Auto-Regressive Mesh Generation Acceleration via Multi-Head Speculative Decoding (ICML 2026 · 0 citations)
Earliest papers here
- Layer-Wise Coordination between Encoder and Decoder for Neural Machine Translation (NeurIPS 2018 · 115 citations)
- Tree-to-tree Neural Networks for Program Translation (NeurIPS 2018 · 82 citations)
- Multi-Channel Encoder for Neural Machine Translation (AAAI 2018 · 28 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.
WMT14 En-De · BLEU (higher is better)
7 papers, 2018–2023 · not reported since 2023
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 29.01 | Layer-Wise Coordination between Encoder and Decoder for Neural Machine Translation (NeurIPS 2018 · 115 citations) |
| 2019 | 29.89 | Tied Transformers: Neural Machine Translation with Shared Encoder and Decoder (AAAI 2019 · 65 citations) |
What has been tried
78 papers here report 922 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 8%, hurt 23%, 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 residual quantization | HumanML3D · FID | 0.093±.004 -> 0.045±.002 | Helped | MoMask: Generative Masked Modeling of 3D Human Motions (CVPR 2024 · 171 citations) |
| remove temporal downsampling | LibriSpeech test-clean and test-other · WER (%) | 2.21 -> 2.23; 4.79 -> 5.09 | Hurt | Zipformer: A faster and better encoder for automatic speech recognition (ICLR 2024 · 28 citations) |
| add the Tree-Aware Module | CROHME 2014/2016/2019 test · ExpRate | 58.38 -> 60.39; 56.98 -> 59.08; 59.12 -> 61.32 | Helped | TAMER: Tree-Aware Transformer for Handwritten Mathematical Expression Recognition (AAAI 2025 · 16 citations) |
| Use ETP instead of parameterized softmax for semantic relations | 20NG · TD | 0.391 -> 0.983 | Helped | FASTopic: Pretrained Transformer is a Fast, Adaptive, Stable, and Transferable Topic Model (NeurIPS 2024 · 13 citations) |
| Stack high-resolution visual tokens instead of repeated original tokens | Phi-3 evaluation suite · AVG | 49.1 -> 51.1 | Helped | DeepStack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs (NeurIPS 2024 · 9 citations) |
| increase compression to (5, 10, 20)× | XNLI · accuracy | 68.74 -> 67.50 | Hurt | MAGNET: Improving the Multilingual Fairness of Language Models with Adaptive Gradient-Based Tokenization (NeurIPS 2024 · 4 citations) |
| Use the proposed probabilistic RVQ instead of conventional RVQ | Speech reconstruction · PESQ | 2.54 -> 2.63 | Helped | CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech (ICLR 2024 · 2 citations) |
| Add KD after Step 3 fusion | Accuracy · score | 81.69 -> 85.28 | Helped | FFN Fusion: Rethinking Sequential Computation in Large Language Models (NeurIPS 2025 · 1 citation) |
| Use an independent non-autoregressive decoder instead of an autoregressive decoder | Voice Search test sets · WER (%) | 4.3 -> 4.5; 14.3 -> 15.3; 17.8 -> 27.3; 23.8 -> 23.5 | Mixed | Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers (NeurIPS 2024 · 1 citation) |
| Use the Jacobian transformation Ws instead of the Bias approximator (s+b) | BATS morphological tasks · faithfulness (top-1 match rate) | Helped | Linear Decoding of Morphology Relations in Language Models (Student Abstract) (AAAI 2025 · 0 citations) | |
| Replace posterior alignment inference with direct prior alignment | MuST-C En-Es and En-De, chunk 320 ms · BLEU | average 1.03 BLEU lower than Transducer baseline | Hurt | Overcoming Non-monotonicity in Transducer-based Streaming Generation (ICML 2025 · 0 citations) |
| Ablate the bigram subnetwork | OSCAR held-out language modeling · evaluation loss | Hurt | Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · 0 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 |
|---|---|---|---|---|
| Adaptive Inputs | WIKITEXT-103 test | perplexity | 18.65 – 24 | 2 papers, e.g. ICML 2022 and ICLR 2020 |
What reviewers will expect
Datasets and baselines named in the 18 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- FineWeb · 2 of 18 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 59 papers; from OpenAlex and the papers' affiliations).
- Google · 4 papers
- Tencent · 3 papers
- Chinese Academy of Sciences · 3 papers
- Microsoft Research Asia · 2 papers
- Technion · 2 papers
- Massachusetts Institute of Technology · 2 papers
- Alibaba Group · 2 papers
- Tel Aviv 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
- Transformer block and depth design (89 papers)
- Language model pretraining and induction (99 papers)
- Transformer expressivity and circuits (56 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)
- Sparse attention for long-context LLMs (147 papers)
- Attention head and block redesign (127 papers)
- Efficient and robust vision transformers (124 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).