Learning How to Localize Change Points: In-Context Adaptation across Change Mechanisms

Modern sequential data often exhibit heterogeneous structural changes, where the appropriate localization strategy may depend on the underlying change mechanism. Change-point localization has traditionally relied on carefully designed statistics tailored to specific mechanisms, meaning that different types of changes typically require different localization procedures. We propose In-Context Change-Point Localization (ICCP), a task-level framework that enables a pretrained Transformer to learn adaptive localization strategies from previously solved change-point tasks. By using solved tasks as demonstrations, ICCP localizes the change point of a new sequence through the prompt alone, without parameter updates or retraining. We provide a theoretical analysis of this in-context adaptation process, establishing finite-sample localization error bounds that characterize the roles of pretraining data, contextual information, and sequence information. Under suitable regularity and scaling conditions, we further prove relative localization consistency. Experiments across diverse change mechanisms demonstrate that ICCP exploits contextual demonstrations and improves with increasing pretraining and contextual information.