How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Training

TL;DR

This paper investigates how large language models internalize new knowledge during continual pre-training by analyzing the evolution of knowledge circuits, revealing key phases and patterns that influence knowledge acquisition.

Contribution

It introduces a knowledge circuit perspective to understand LLMs' knowledge acquisition, providing new insights into circuit evolution during continual pre-training.

Findings

Knowledge relevance affects new knowledge acquisition.

Circuit evolution shifts from formation to optimization phase.

Knowledge circuits follow a deep-to-shallow evolution pattern.

Abstract

Despite exceptional capabilities in knowledge-intensive tasks, Large Language Models (LLMs) face a critical gap in understanding how they internalize new knowledge, particularly how to structurally embed acquired knowledge in their neural computations. We address this issue through the lens of knowledge circuit evolution, identifying computational subgraphs that facilitate knowledge storage and processing. Our systematic analysis of circuit evolution throughout continual pre-training reveals several key findings: (1) the acquisition of new knowledge is influenced by its relevance to pre-existing knowledge; (2) the evolution of knowledge circuits exhibits a distinct phase shift from formation to optimization; (3) the evolution of knowledge circuits follows a deep-to-shallow pattern. These insights not only advance our theoretical understanding of the mechanisms of new knowledge acquisition in LLMs, but also provide potential implications for improving continual pre-training strategies to enhance model performance. Code and data will be available at https://github.com/zjunlp/DynamicKnowledgeCircuits.