Identifying and Analyzing Task-Encoding Tokens in Large Language Models

In-context learning (ICL) has become an effective solution for few-shot learning in natural language processing. However, our understanding of ICL's working mechanisms is limited, specifically regarding how models learn to perform tasks from ICL demonstrations. For example, unexpectedly large changes in performance can arise from small changes in the prompt, leaving prompt design a largely empirical endeavour. In this paper, we investigate this problem by identifying and analyzing task-encoding tokens on whose representations the task performance depends. Using experiments that ablate the representations of different token types, we find that template and stopword tokens are the most prone to be task-encoding. In addition, we demonstrate experimentally that lexical meaning, repetition, and text formatting are the main distinguishing characteristics of these tokens. Our work sheds light on how large language models (LLMs) learn to perform a task from demonstrations, deepens our understanding of the varied roles different types of tokens play in LLMs, and provides insights for avoiding instability from improperly utilizing task-encoding tokens.