Evaluating and Understanding Declarative and Procedural Knowledge in Language Models

Abstract

Advancements in language model development have led to models with a broad range of useful capabilities, including those related to knowledge. These knowledge capabilities can be conceptualized as improvements in declarative knowledge ("knowing that", being able to recall and use factual information) and procedural knowledge ("knowing how", the ability to perform tasks or skills). While language models have improved substantially along both axes, many open questions remain regarding the extent of models' declarative knowledge in various domains, the predictability of their procedural knowledge capabilities, and how each of these arises from the underlying datasets that models are trained on. In this thesis, we explore work in each of these directions, expanding our understanding of both declarative and procedural knowledge in language models. We begin with a study of declarative knowledge in language models via the task of scientific knowledge base completion, finding that language models effectively complement classical text-free methods on this task and are particularly valuable when applied to unseen scientific entities and concepts. Our work proceeds to investigate language models' procedural knowledge capabilities by attempting to predict task performance from instructions, discovering that this remains a challenging task across various data- and model-related experimental settings. We then analyze how knowledge is related to language models' training data by proposing an experimental recipe for performing interventional analyses on training, demonstrating the utility of this recipe by applying it to study various model behaviors. Finally, we conclude with thoughts on potential future directions on applying language models' declarative knowledge for scientific discovery, modernizing the experimental study of predicting models' procedural knowledge, and adapting our proposed experimental methodology to better understand how different stages of training each contribute to models' knowledge capabilities.