Investigating the Pre-Training Dynamics of In-Context Learning: Task Recognition vs. Task Learning

TL;DR

This paper investigates the pre-training dynamics of in-context learning, revealing a competitive relationship between task recognition and task learning abilities, and proposes an adaptive ensemble method to enhance ICL performance.

Contribution

It is the first to analyze the competitive dynamics between TR and TL during pre-training and introduces an adaptive ensemble approach to improve ICL performance.

Findings

TR and TL are competitive during pre-training.

Negative correlation between competition and ICL performance.

Adaptive ensemble learning boosts ICL, outperforming larger models.

Abstract

The emergence of in-context learning (ICL) is potentially attributed to two major abilities: task recognition (TR) for recognizing the task from demonstrations and utilizing pre-trained priors, and task learning (TL) for learning from demonstrations. However, relationships between the two abilities and how such relationships affect the emergence of ICL is unclear. In this paper, we take the first step by examining the pre-training dynamics of the emergence of ICL. With carefully designed metrics, we find that these two abilities are, in fact, competitive during pre-training. Moreover, we observe a strong negative correlation between the competition and ICL performance. Further analysis of common pre-training factors (i.e., model size, dataset size, and data curriculum) demonstrates possible ways to manage the competition. Based on these insights, we propose a simple yet effective method to better integrate these two abilities for ICL at inference time. Through adaptive ensemble learning, the performance of ICL can be significantly boosted, enabling two small models to outperform a larger one with more than twice the parameters. The code is available at https://github.com/RUCAIBox/Competitive-ICL.