Decomposing Prediction Mechanisms for In-Context Recall

• URL: http://arxiv.org/abs/2507.01414v1
• Date: Wed, 02 Jul 2025 07:09:09 GMT
• Title: Decomposing Prediction Mechanisms for In-Context Recall
• Authors: Sultan Daniels, Dylan Davis, Dhruv Gautam, Wentinn Liao, Gireeja Ranade, Anant Sahai,
• Abstract summary: We introduce a new family of toy problems that combine features of linear-regression-style continuous in-context learning (ICL) with discrete associative recall.<n>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.
• Score: 4.148170164455114
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors [61.92704516732144]
  We show that the most robust features for correctness prediction are those that play a distinctive causal role in the model's behavior.<n>We propose two methods that leverage causal mechanisms to predict the correctness of model outputs.
  arXiv  Detail & Related papers  (2025-05-17T00:31:39Z)
• (How) Do Language Models Track State? [50.516691979518164]
  Transformer language models (LMs) exhibit behaviors that appear to require tracking the unobserved state of an evolving world.<n>We study state tracking in LMs trained or fine-tuned to compose permutations.<n>We show that LMs consistently learn one of two state tracking mechanisms for this task.
  arXiv  Detail & Related papers  (2025-03-04T18:31:02Z)
• Looking Beyond The Top-1: Transformers Determine Top Tokens In Order [13.032106683136394]
  We analyze the computation performed by Transformers in the layers after the top-1 prediction has become fixed. We find that these saturation events happen in order of the corresponding tokens' ranking. We propose an underlying mechanism of task transition for this sequential saturation.
  arXiv  Detail & Related papers  (2024-10-26T16:00:38Z)
• In-Context Convergence of Transformers [63.04956160537308]
  We study the learning dynamics of a one-layer transformer with softmax attention trained via gradient descent. For data with imbalanced features, we show that the learning dynamics take a stage-wise convergence process.
  arXiv  Detail & Related papers  (2023-10-08T17:55:33Z)
• From system models to class models: An in-context learning paradigm [0.0]
  We introduce a novel paradigm for system identification, addressing two primary tasks: one-step-ahead prediction and multi-step simulation. We learn a meta model that represents a class of dynamical systems. For one-step prediction, a GPT-like decoder-only architecture is utilized, whereas the simulation problem employs an encoder-decoder structure.
  arXiv  Detail & Related papers  (2023-08-25T13:50:17Z)
• Mutual Exclusivity Training and Primitive Augmentation to Induce Compositionality [84.94877848357896]
  Recent datasets expose the lack of the systematic generalization ability in standard sequence-to-sequence models. We analyze this behavior of seq2seq models and identify two contributing factors: a lack of mutual exclusivity bias and the tendency to memorize whole examples. We show substantial empirical improvements using standard sequence-to-sequence models on two widely-used compositionality datasets.
  arXiv  Detail & Related papers  (2022-11-28T17:36:41Z)
• Self-Regulated Learning for Egocentric Video Activity Anticipation [147.9783215348252]
  Self-Regulated Learning (SRL) aims to regulate the intermediate representation consecutively to produce representation that emphasizes the novel information in the frame of the current time-stamp. SRL sharply outperforms existing state-of-the-art in most cases on two egocentric video datasets and two third-person video datasets.
  arXiv  Detail & Related papers  (2021-11-23T03:29:18Z)
• Anticipating synchronization with machine learning [1.0958014189747356]
  In applications of dynamical systems, it is desired to predict the onset of synchronization. We develop a prediction framework that is model free and fully data driven. We demonstrate the machine-learning based framework using representative chaotic models and small network systems.
  arXiv  Detail & Related papers  (2021-03-13T03:51:48Z)
• A Novel Anomaly Detection Algorithm for Hybrid Production Systems based on Deep Learning and Timed Automata [73.38551379469533]
  DAD:DeepAnomalyDetection is a new approach for automatic model learning and anomaly detection in hybrid production systems. It combines deep learning and timed automata for creating behavioral model from observations. The algorithm has been applied to few data sets including two from real systems and has shown promising results.
  arXiv  Detail & Related papers  (2020-10-29T08:27:43Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.