Next-token pretraining implies in-context learning

TL;DR

This paper demonstrates that in-context learning naturally results from standard next-token pretraining, with a theoretical framework predicting its dynamics and experimental validation showing phase transitions and loss scaling.

Contribution

It establishes a foundational information-theoretic explanation for in-context learning as an emergent property of standard pretraining, not an exotic phenomenon.

Findings

Predicts in-context learning dynamics using information theory

Reproduces phase transitions in training loss for induction heads

Shows power-law scaling of in-context loss

Abstract

We argue that in-context learning (ICL) predictably arises from standard self-supervised next-token pretraining, rather than being an exotic emergent property. This work establishes the foundational principles of this emergence by focusing on in-distribution ICL, demonstrating how models necessarily adapt to context when trained on token sequences, especially from non-ergodic sources. Our information-theoretic framework precisely predicts these in-distribution ICL dynamics (i.e., context-dependent loss reduction). We verify this with experiments using synthetic datasets of differing types of correlational structure, reproducing characteristic phenomena like phase transitions in training loss for induction head formation and power-law scaling of in-context loss. We further show that a model's in-context performance on any task is mathematically coupled to the ensemble of tasks seen in pretraining, offering a fundamental explanation, grounded in architecture- and modality-independent principles, for such inference-time learning.