Research suggests LLM in-context learning stems from specific pre-training data

JoshPurtell · x · 2026-08-26

The author replies that some research papers suggest LLMs derive their in-context learning (ICL) ability from specific kinds of data in web pre-training. If you ablate this data, the model loses ICL. This implies ICL is not just an emergent property of generic next-token prediction but relies on specific patterns in the training data.

Related event: Study: LLM In-Context Learning Stems from Parallel Structures in Pretraining Data(2 posts)→

Original post →

More from Research

Research channel →