Towards transparent representations: on internal structure and external world modeling in LLMs

Abstract

Large language models (LLMs) generalize far beyond their training distribution, enabling impressive downstream performance in domains vastly different from their pretraining distribution. In this thesis, we develop a data-centric view on machine learning. We suggest that the deep generalization of LLMs is best understood through studying the relationships between the four fundamental components of this data generalization: pretraining data, test-time inputs, model outputs, and internal structure. Of these, we present two full research studies characterizing test-time inputs and internal structure. Chapter 1 develops the data-centric view of machine learning, and outline the thesis. Chapter 2 presents Breakpoint, a method of generating difficult coding tasks for models at a large scale that attempts to disambiguate the factors that make problems at test-time difficult. Chapter 3 analyzes the structure of gradient-based jailbreaks in LLMs. We argue that even though GBJs are more out of distribution than even random text, they induce a low-rank, structured change in models. Finally, Chapter 4 discusses the recent rise of reasoning models and proposing some lines of future work in the data-centric view towards developing more robust understanding of LLMs.