Llama See, Llama Do: Contextual Entrainment and Distraction in LLMs

What it says

Examines a mechanism through which irrelevant tokens already present in context pull later model generations towards them.

Why it matters here

Provides a more precise research term for one mechanism often described informally as context bleed.

Editorial caution

Peer review increases confidence in the reported method and result; it does not make every adjacent claim universal.