Research map / Architectures and efficiency / Attention and transformers
Training dynamics of in-context learning
Proves how transformers acquire in-context learning and induction heads during training.
62 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · 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.
21 papers in 2023–24 and 40 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 3.6% to 3.4%.
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
- 2022 · What Can Transformers Learn In-Context? A Case Study of Simple Function Classes (NeurIPS 2022 · 58 citations)
- 2023 · The Transient Nature of Emergent In-Context Learning in Transformers (NeurIPS 2023 · 4 citations)
- 2024 · Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks (ICML 2024 · 10 citations)
- 2025 · TabICL: A Tabular Foundation Model for In-Context Learning on Large Data (ICML 2025 · 9 citations)
- 2026 · Optimal Attention Temperature Improves the Robustness of In-Context Learning under Distribution Shift in High Dimensions (ICML 2026 · 0 citations)
Earliest papers here
- What Can Transformers Learn In-Context? A Case Study of Simple Function Classes (NeurIPS 2022 · 58 citations)
- The Transient Nature of Emergent In-Context Learning in Transformers (NeurIPS 2023 · 4 citations)
- Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks (ICML 2024 · 10 citations)
What has been tried
42 papers here report 390 measured changes (a component added, removed or swapped, compared with a baseline): helped 52%, no clear effect 13%, hurt 26%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add an attention layer within the Mamba blocks | Vector MQAR · test loss (MSE) | Helped | Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks (ICML 2024 · 10 citations) | |
| add tree-based SCMs to synthetic pretraining | 200 classification datasets · relative accuracy | Helped | TabICL: A Tabular Foundation Model for In-Context Learning on Large Data (ICML 2025 · 9 citations) | |
| replace head-specific shift magnitudes with one magnitude shared across heads | VQAv2 · accuracy | 57.89 -> 59.64 | Hurt | Mimic In-Context Learning for Multimodal Tasks (CVPR 2025 · 4 citations) |
| Add NGH¹, NGH², and NGH³ to RetNet | RegBench, N=2500 · TVD | 0.392 -> 0.217 | Helped | In-Context Language Learning: Architectures and Algorithms (ICML 2024 · 3 citations) |
| Add two fully-connected layers after the relational bottleneck | Line oddball · loss | Helped | MLPs Learn In-Context on Regression and Classification Tasks (ICLR 2025 · 2 citations) | |
| Set β₂ to 0, preventing extra target attention to contextual labels | In-context classification task · ICL accuracy | No clear effect | The mechanistic basis of data dependence and abrupt learning in an in-context classification task (ICLR 2024 · 2 citations) | |
| Use a one-layer transformer embedding before the upper module | Linear ICL pasting, D-dimensional, noise 0.1 · test risk | Helped | How Do Transformers Learn In-Context Beyond Simple Functions? A Case Study on Learning with Representations (ICLR 2024 · 2 citations) | |
| Use multiple Layer 2 induction heads instead of restricting training to one head | Few-shot Omniglot task · training dynamics | Helped | What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation (ICML 2024 · 2 citations) | |
| concatenate an irrelevant GSM-IC sentence to an SST-2 input-label pair | SST-2 attention analysis, 100 evaluation prompts · attention correlation magnitude | Mixed | Why Larger Language Models Do In-context Learning Differently? (ICML 2024 · 2 citations) | |
| Increase task Lipschitzness L | Lipschitzness ablation · attention matrix spectral norm ||M|| | Helped | In-Context Learning with Transformers: Softmax Attention Adapts to Function Lipschitzness (NeurIPS 2024 · 2 citations) | |
| Add extra context loss instead of extra weights loss | Shape3D D_rnd · accuracy | λ = 0.1 | No clear effect | Breaking through the learning plateaus of in-context learning in Transformer (ICML 2024 · 1 citation) |
| sample pre-training task vectors randomly instead of using the rectangular rule | modular arithmetic out-of-distribution generalization · accuracy | Hurt | Learning to grok: Emergence of in-context learning and skill composition in modular arithmetic tasks (NeurIPS 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 28 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Baselines
- LoRA · 2 of 28 papers
- Zero-shot · 2 of 28 papers
- dMMSE estimator · 2 of 28 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 · 5 papers
- University of Michigan - Ann Arbor · 3 papers
- Massachusetts Institute of Technology · 3 papers
- National University of Singapore · 3 papers
- Google DeepMind · 3 papers
- Nanyang Technological University · 2 papers
- Microsoft · 2 papers
- University of Pennsylvania · 2 papers
- The Ohio State University · 2 papers
- University of Wisconsin - Madison · 2 papers
Use it
Download BibTeX of 22 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)
- Attention head mechanisms (75 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).