In-Context Learning Strategies Emerge Rationally

• URL: http://arxiv.org/abs/2506.17859v2
• Date: Thu, 26 Jun 2025 16:54:57 GMT
• Title: In-Context Learning Strategies Emerge Rationally
• Authors: Daniel Wurgaft, Ekdeep Singh Lubana, Core Francisco Park, Hidenori Tanaka, Gautam Reddy, Noah D. Goodman,
• Abstract summary: We show that when trained to learn a mixture of tasks, strategies learned by a model for performing ICL can be captured by a family of Bayesian predictors.<n>We develop a hierarchical Bayesian framework that almost perfectly predicts Transformer next-token predictions throughout training.<n>Our work advances an explanatory and predictive account of ICL grounded in tradeoffs between strategy loss and complexity.
• Score: 36.21854069812912
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent work analyzing in-context learning (ICL) has identified a broad set of strategies that describe model behavior in different experimental conditions. We aim to unify these findings by asking why a model learns these disparate strategies in the first place. Specifically, we start with the observation that when trained to learn a mixture of tasks, as is popular in the literature, the strategies learned by a model for performing ICL can be captured by a family of Bayesian predictors: a memorizing predictor, which assumes a discrete prior on the set of seen tasks, and a generalizing predictor, where the prior matches the underlying task distribution. Adopting the normative lens of rational analysis, where a learner's behavior is explained as an optimal adaptation to data given computational constraints, we develop a hierarchical Bayesian framework that almost perfectly predicts Transformer next-token predictions throughout training -- without assuming access to its weights. Under this framework, pretraining is viewed as a process of updating the posterior probability of different strategies, and inference-time behavior as a posterior-weighted average over these strategies' predictions. Our framework draws on common assumptions about neural network learning dynamics, which make explicit a tradeoff between loss and complexity among candidate strategies: beyond how well it explains the data, a model's preference towards implementing a strategy is dictated by its complexity. This helps explain well-known ICL phenomena, while offering novel predictions: e.g., we show a superlinear trend in the timescale for transitioning from generalization to memorization as task diversity increases. Overall, our work advances an explanatory and predictive account of ICL grounded in tradeoffs between strategy loss and complexity.

Related papers

• A Theory of Inference Compute Scaling: Reasoning through Directed Stochastic Skill Search [15.387256204743407]
  Large language models (LLMs) demand considerable computational, energy, and financial resources during both training and deployment.<n>Inference costs now represent a significant and growing component of the overall resource burden.<n>We introduce directed skill search (DS3), a general framework that represents inference as expressive over a learned skill graph.
  arXiv  Detail & Related papers  (2025-06-10T14:47:48Z)
• In-Context Linear Regression Demystified: Training Dynamics and Mechanistic Interpretability of Multi-Head Softmax Attention [52.159541540613915]
  We study how multi-head softmax attention models are trained to perform in-context learning on linear data.<n>Our results reveal that in-context learning ability emerges from the trained transformer as an aggregated effect of its architecture and the underlying data distribution.
  arXiv  Detail & Related papers  (2025-03-17T02:00:49Z)
• Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining [55.262510814326035]
  Existing reweighting strategies primarily focus on group-level data importance.<n>We introduce novel algorithms for dynamic, instance-level data reweighting.<n>Our framework allows us to devise reweighting strategies deprioritizing redundant or uninformative data.
  arXiv  Detail & Related papers  (2025-02-10T17:57:15Z)
• Practical Performative Policy Learning with Strategic Agents [8.361090623217246]
  We study the performative policy learning problem, where agents adjust their features in response to a released policy to improve their potential outcomes.<n>We propose a gradient-based policy optimization algorithm with a differentiable classifier as a substitute for the high-dimensional distribution map.
  arXiv  Detail & Related papers  (2024-12-02T10:09:44Z)
• Restoring balance: principled under/oversampling of data for optimal classification [0.0]
  Class imbalance in real-world data poses a common bottleneck for machine learning tasks.<n> Mitigation strategies, such as under or oversampling the data depending on their abundances, are routinely proposed and tested empirically.<n>We provide a sharp prediction of the effects of under/oversampling strategies depending on class imbalance, first and second moments of the data, and the metrics of performance considered.
  arXiv  Detail & Related papers  (2024-05-15T17:45:34Z)
• Data Assimilation in Chaotic Systems Using Deep Reinforcement Learning [0.5999777817331317]
  Data assimilation plays a pivotal role in diverse applications, ranging from climate predictions and weather forecasts to trajectory planning for autonomous vehicles. Recent advancements have seen the emergence of deep learning approaches in this domain, primarily within a supervised learning framework. In this study, we introduce a novel DA strategy that utilizes reinforcement learning (RL) to apply state corrections using full or partial observations of the state variables.
  arXiv  Detail & Related papers  (2024-01-01T06:53:36Z)
• Hierarchical Decomposition of Prompt-Based Continual Learning: Rethinking Obscured Sub-optimality [55.88910947643436]
  Self-supervised pre-training is essential for handling vast quantities of unlabeled data in practice. HiDe-Prompt is an innovative approach that explicitly optimize the hierarchical components with an ensemble of task-specific prompts and statistics. Our experiments demonstrate the superior performance of HiDe-Prompt and its robustness to pre-training paradigms in continual learning.
  arXiv  Detail & Related papers  (2023-10-11T06:51:46Z)
• An Analysis of Initial Training Strategies for Exemplar-Free Class-Incremental Learning [36.619804184427245]
  Class-Incremental Learning (CIL) aims to build classification models from data streams. Due to catastrophic forgetting, CIL is particularly challenging when examples from past classes cannot be stored. Use of models pre-trained in a self-supervised way on large amounts of data has recently gained momentum.
  arXiv  Detail & Related papers  (2023-08-22T14:06:40Z)
• Towards Out-of-Distribution Sequential Event Prediction: A Causal Treatment [72.50906475214457]
  The goal of sequential event prediction is to estimate the next event based on a sequence of historical events. In practice, the next-event prediction models are trained with sequential data collected at one time. We propose a framework with hierarchical branching structures for learning context-specific representations.
  arXiv  Detail & Related papers  (2022-10-24T07:54:13Z)
• Learning from Heterogeneous Data Based on Social Interactions over Graphs [58.34060409467834]
  This work proposes a decentralized architecture, where individual agents aim at solving a classification problem while observing streaming features of different dimensions. We show that the. strategy enables the agents to learn consistently under this highly-heterogeneous setting. We show that the. strategy enables the agents to learn consistently under this highly-heterogeneous setting.
  arXiv  Detail & Related papers  (2021-12-17T12:47:18Z)
• Predicting Deep Neural Network Generalization with Perturbation Response Curves [58.8755389068888]
  We propose a new framework for evaluating the generalization capabilities of trained networks. Specifically, we introduce two new measures for accurately predicting generalization gaps. We attain better predictive scores than the current state-of-the-art measures on a majority of tasks in the Predicting Generalization in Deep Learning (PGDL) NeurIPS 2020 competition.
  arXiv  Detail & Related papers  (2021-06-09T01:37:36Z)

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.