When Context Sticks: Studying Interference in In-Context Learning

TL;DR

This study explores how earlier examples in in-context learning prompts interfere with model adaptation, revealing persistent bias effects and the impact of training curricula on model resilience.

Contribution

It provides empirical evidence of context interference in ICL and analyzes how training strategies influence model robustness to task switches.

Findings

Linear examples bias predictions negatively, especially with more preceding examples.

Quadratic examples help reduce error but with diminishing returns.

Sequential training on target functions enhances model recovery speed.

Abstract

This paper investigates context stickiness in in-context learning (ICL), a phenomenon where earlier examples in a prompt interfere with a transformer's ability to adapt to later tasks. Using synthetic regression tasks over linear and quadratic functions, we examine how models trained under sequential, mixed, and random curricula handle abrupt task switches during inference. By sweeping over structured combinations of misleading linear examples followed by recovery quadratic examples, we quantify how prior context biases prediction error and how quickly models realign. Our results show strong evidence of persistent interference: more preceding linear examples reliably degrade quadratic predictions, while additional quadratic examples reduce error but with diminishing returns. We further find that training curricula significantly modulate resilience, with sequential training on the target function class yielding the fastest recovery, and surprisingly, random training producing the least robust behavior.