mlsanityData to 14 Oct 2026

The emergence of sparse attention: impact of data distribution and benefits of repetition

arXiv 2505.17863 ↗First posted 23 May 2025cs.LG cs.NEcited by 25

Not on the map. Why?

Abstract (arXiv)

Emergence is a fascinating property of large language models and neural networks more broadly: as models scale and train for longer, they sometimes develop new abilities in sudden ways. Despite initial studies, we still lack a comprehensive understanding of how and when these abilities emerge. To address this gap, we study the emergence over training of sparse attention, a critical and frequently observed attention pattern in Transformers. By combining theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant, we uncover the mechanics driving sparse attention emergence and reveal that emergence timing follows power laws based on task structure, architecture, and optimizer choice. We additionally find that repetition can greatly speed up emergence. Finally, we confirm these results on a well-studied in-context associative recall task. Our findings provide a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence.

Key facts. Cited by 25 papers in the map (data through 14 Oct 2026). Builds on 50 of its 73 references that are in the map. First posted on arXiv 23 May 2025.

Builds on

50 of 73 references.

References · 50

  1. 2017Attention Is All You Needcited by 60,361
  2. 2020Language Models are Few-Shot Learnerscited by 32,419
  3. 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,305
  4. 2023GPT-4 Technical Reportcited by 18,396
  5. 2020Scaling Laws for Neural Language Modelscited by 6,586
  6. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,718
  7. 2022Training Compute-Optimal Large Language Modelscited by 3,019
  8. 2022Emergent Abilities of Large Language Modelscited by 2,512
  9. 2022Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language modelscited by 1,998
  10. 2022In-context Learning and Induction Headscited by 964
  11. 2023Vision Transformers Need Registerscited by 871
  12. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
38 more
  1. 2013Exact solutions to the nonlinear dynamics of learning in deep linear neural networkscited by 763
  2. 2021Deduplicating Training Data Makes Language Models Bettercited by 714
  3. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 629
  4. 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 528
  5. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 523
  6. 2021Gradient Descent on Neural Networks Typically Occurs at the Edge of Stabilitycited by 427
  7. 2024Parallelizing Linear Transformers with the Delta Rule over Sequence Lengthcited by 396
  8. 2023In-Context Learning Creates Task Vectorscited by 312
  9. 2022Data Distributional Properties Drive Emergent In-Context Learning in Transformerscited by 309
  10. 2023Function Vectors in Large Language Modelscited by 308
  11. 2023Physics of Language Models: Part 3.1, Knowledge Storage and Extractioncited by 270
  12. 2024Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Modelscited by 242
  13. 2022Predictability and Surprise in Large Generative Modelscited by 210
  14. 2023Zoology: Measuring and Improving Recall in Efficient Language Modelscited by 204
  15. 2024Foundational Challenges in Assuring Alignment and Safety of Large Language Modelscited by 199
  16. 2018A mathematical theory of semantic development in deep neural networkscited by 186
  17. 2023Birth of a Transformer: A Memory Viewpointcited by 161
  18. 2022Scaling Laws and Interpretability of Learning from Repeated Datacited by 152
  19. 2023SGD learning on neural networks: leap complexity and saddle-to-saddle dynamicscited by 143
  20. 2022The merged-staircase property: a necessary and nearly sufficient condition for SGD learning of sparse functions on two-layer neural networkscited by 142
  21. 2024Why Transformers Need Adam: A Hessian Perspectivecited by 124
  22. 2023The mechanistic basis of data dependence and abrupt learning in an in-context classification taskcited by 118
  23. 2024The Evolution of Statistical Induction Heads: In-Context Learning Markov Chainscited by 118
  24. 2024How Transformers Learn Causal Structure with Gradient Descentcited by 118
  25. 2023A Theory for Emergence of Complex Skills in Language Modelscited by 113
  26. 2024Understanding Emergent Abilities of Language Models from the Loss Perspectivecited by 94
  27. 2024What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formationcited by 89
  28. 2025Findings of the BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corporacited by 88
  29. 2021Saddle-to-Saddle Dynamics in Deep Linear Networks: Small Initialization Training, Symmetry, and Sparsitycited by 85
  30. 2023The Transient Nature of Emergent In-Context Learning in Transformerscited by 76
  31. 2024Training Dynamics of Multi-Head Softmax Attention for In-Context Learning: Emergence, Convergence, and Optimalitycited by 74
  32. 2023Saddle-to-Saddle Dynamics in Diagonal Linear Networkscited by 57
  33. 2024Competition Dynamics Shape Algorithmic Phases of In-Context Learningcited by 54
  34. 2025Training Dynamics of In-Context Learning in Linear Attentioncited by 46
  35. 2025How do language models learn facts? Dynamics, curricula and hallucinationscited by 37
  36. 2024Attention layers provably solve single-location regressioncited by 17
  37. 2024Emergent properties with repeated examplescited by 11
  38. 2024Predicting Emergent Capabilities by Finetuningcited by 10

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 8; 2026: 17.

25 papers.

Sort:

By area:

  1. 2025What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformerscited by 7
  2. 2025Predicting the Emergence of Induction Heads in Language Model Pretrainingcited by 7
  3. 2025In-Context Algebracited by 5
  4. 2026Dissecting Multimodal In-Context Learning: Modality Asymmetries and Circuit Dynamics in modern Transformerscited by 5
  5. 2026Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Stepscited by 4
  6. 2026What Makes Position Zero Special? A Mechanistic Study of Position Zero Attention Sinks in LLMscited by 3
  7. 2026Kuramoto Oscillatory Phase Encoding: Neuro-inspired Synchronization for Improved Learning Efficiencycited by 3
  8. 2025Block Sparse Flash Attentioncited by 2
  9. 2026Incremental Learning of Sparse Attention Patterns in Transformerscited by 2
  10. 2026Gradient Flow Polarizes Softmax Outputs towards Low-Entropy Solutionscited by 2
  11. 2026Invariant Learning Dynamics of Transformers in Inductive Reasoning Taskscited by 2
  12. 2025Understanding the Staged Dynamics of Transformers in Learning Latent Structurecited by 1
13 more
  1. 2026Transformers Learn Latent Mixture Models In-Context via Mirror Descentcited by 1
  2. 2026Grokking as Structural Inference: Transformers Need Bayesian Lottery Ticketscited by 1
  3. 2026SIFT: Selective-Index For Fast Compute of RAG Prefill by Exploiting Attention Invariancecited by 1
  4. 2026Emergent Capabilities Arise Randomly from Learning Sparse Attention Patternscited by 1
  5. 2026Induction Heads Interpolate N-Gramscited by 1
  6. 2026Train Smarter, Not Longer: Memorization-Guided Data Reuse for Efficient LLM Trainingcited by 1
  7. 2025A solution to generalized learning from small training sets found in infants repeated visual experiences of individual objectscited by 0
  8. 2025In-Context Learning Without Copyingcited by 0
  9. 2025On the Universality of Transformer Architectures; How Much Attention Is Enough?cited by 0
  10. 2026Contrast Matters: Understanding Robustness of In-Context Fine-Tuning to Target-Context Relatednesscited by 0
  11. 2026Less Data, Faster Training: repeating smaller datasets speeds up learning via sampling biasescited by 0
  12. 2026Meta-Attention: Bayesian Per-Token Routing for Efficient Transformer Inferencecited by 0
  13. 2026Emergent One-Third Scaling Law as Attention Tries to Concentratecited by 0