Exploring Concept Depth: How Large Language Models Acquire Knowledge at Different Layers?

• DOI: 10.48550/arxiv.2404.07066
• Publication Date: 2024-04-10
• Authors: Mingyu Jin; Qinkai Yu; Jingyuan Huang; Qingcheng Zeng; Zhenting Wang; Wenyue Hua; Haiyan Zhao; Kai Mei; Yanda Meng; Kaize Ding; Fan Yang; Mengnan Du; Yongfeng Zhang
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Science - Computation and Language; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computation and Language (cs.CL); Machine Learning (cs.LG)

This paper studies the phenomenon that different concepts are learned in layers of large language models, i.e. more difficult fully acquired with deeper layers. We define difficulty by level abstraction, and here it is crudely categorized factual, emotional, inferential. Each category contains a spectrum tasks, arranged from simple to complex. For example, within factual dimension, tasks range lie detection categorizing mathematical problems. employ probing technique extract representations model apply these classification tasks. Our findings reveal models tend efficiently classify simpler indicating shallower Conversely, complex may only be discernible at layers, if all. explores implications for our understanding learning processes internal representations. implementation available \url{https://github.com/Luckfort/CD}.