mlsanityData to 14 Oct 2026

Decomposing Prediction Mechanisms for In-Context Recall

arXiv 2507.01414 ↗First posted 2 Jul 2025cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

We introduce a new family of toy problems that combine features of linear-regression-style continuous in-context learning (ICL) with discrete associative recall. We pretrain transformer models on sample traces from this toy, specifically symbolically-labeled interleaved state observations from randomly drawn linear deterministic dynamical systems. We study if the transformer models can recall the state of a sequence previously seen in its context when prompted to do so with the corresponding in-context label. Taking a closer look at this task, it becomes clear that the model must perform two functions: (1) identify which system's state should be recalled and apply that system to its last seen state, and (2) continuing to apply the correct system to predict the subsequent states. Training dynamics reveal that the first capability emerges well into a model's training. Surprisingly, the second capability, of continuing the prediction of a resumed sequence, develops much earlier. Via out-of-distribution experiments, and a mechanistic analysis on model weights via edge pruning, we find that next-token prediction for this toy problem involves at least two separate mechanisms. One mechanism uses the discrete symbolic labels to do the associative recall required to predict the start of a resumption of a previously seen sequence. The second mechanism, which is largely agnostic to the discrete symbolic labels, performs a "Bayesian-style" prediction based on the previous token and the context. These two mechanisms have different learning dynamics. To confirm that this multi-mechanism (manifesting as separate phase transitions) phenomenon is not just an artifact of our toy setting, we used OLMo training checkpoints on an ICL translation task to see a similar phenomenon: a decisive gap in the emergence of first-task-token performance vs second-task-token performance.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 58 of its 85 references that are in the map. First posted on arXiv 2 Jul 2025.

Builds on

58 of 85 references.

References · 58

  1. 2020Language Models are Few-Shot Learnerscited by 32,419
  2. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  3. 2022Emergent Abilities of Large Language Modelscited by 2,512
  4. 2022Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language modelscited by 1,998
  5. 2022Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?cited by 1,434
  6. 2022In-context Learning and Induction Headscited by 964
  7. 2021An Explanation of In-context Learning as Implicit Bayesian Inferencecited by 938
  8. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  9. 2022Transformers learn in-context by gradient descentcited by 732
  10. 2022What Can Transformers Learn In-Context? A Case Study of Simple Function Classescited by 715
  11. 2017The Implicit Bias of Gradient Descent on Separable Datacited by 655
  12. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 629
38 more
  1. 2024OLMo: Accelerating the Science of Language Modelscited by 613
  2. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 523
  3. 2020The Pitfalls of Simplicity Bias in Neural Networkscited by 351
  4. 2017Stochastic Gradient Descent as Approximate Bayesian Inferencecited by 300
  5. 2020Gradient Starvation: A Learning Proclivity in Neural Networkscited by 267
  6. 2023Transformers as Algorithms: Generalization and Stability in In-context Learningcited by 222
  7. 2022Predictability and Surprise in Large Generative Modelscited by 210
  8. 2023Zoology: Measuring and Improving Recall in Efficient Language Modelscited by 204
  9. 2023The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networkscited by 156
  10. 2022Omnigrok: Grokking Beyond Algorithmic Datacited by 146
  11. 2023Pretraining task diversity and the emergence of non-Bayesian in-context learning for regressioncited by 146
  12. 2023The Quantization Model of Neural Scalingcited by 138
  13. 2023Sudden Drops in the Loss: Syntax Acquisition, Phase Transitions, and Simplicity Bias in MLMscited by 124
  14. 2023The mechanistic basis of data dependence and abrupt learning in an in-context classification taskcited by 118
  15. 2024The Evolution of Statistical Induction Heads: In-Context Learning Markov Chainscited by 118
  16. 2023Are Emergent Abilities in Large Language Models just In-Context Learning?cited by 109
  17. 2023Grokking as the Transition from Lazy to Rich Training Dynamicscited by 108
  18. 2023How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression?cited by 96
  19. 2024Understanding Emergent Abilities of Language Models from the Loss Perspectivecited by 94
  20. 2024In-Context Language Learning: Architectures and Algorithmscited by 89
  21. 2024What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formationcited by 89
  22. 2023Grokking modular arithmeticcited by 83
  23. 2023Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokkingcited by 80
  24. 2023Symbol tuning improves in-context learning in language modelscited by 75
  25. 2025Which Attention Heads Matter for In-Context Learning?cited by 61
  26. 2024Finding Transformer Circuits with Edge Pruningcited by 60
  27. 2023The Local Learning Coefficient: A Singularity-Aware Complexity Measurecited by 51
  28. 2024Deep Networks Always Grok and Here is Whycited by 51
  29. 2020Is SGD a Bayesian sampler? Well, almostcited by 50
  30. 2023Unifying Grokking and Double Descentcited by 50
  31. 2025Grokking at the Edge of Numerical Stabilitycited by 46
  32. 2025Training Dynamics of In-Context Learning in Linear Attentioncited by 46
  33. 2024Dual Operating Modes of In-Context Learningcited by 43
  34. 2023Dynamical versus Bayesian Phase Transitions in a Toy Model of Superpositioncited by 29
  35. 2024Emergence in non-neural models: grokking modular arithmetic via average gradient outer productcited by 28
  36. 2024Transformers on Markov Data: Constant Depth Sufficescited by 27
  37. 2024A Percolation Model of Emergence: Analyzing Transformers Trained on a Formal Languagecited by 26
  38. 2021Multi-scale Feature Learning Dynamics: Insights for Double Descentcited by 24

50 most cited of 58 All 58 (JSON)

Built on by

Not cited yet.