Mechanics of Next Token Prediction with Self-Attention

arXiv:2403.08081v1 Announce Type: new
Abstract: Transformer-based language models are trained on large datasets to predict the next token given an input sequence. Despite this simple training objective, they have led to revolutionary advances in natural language processing. Underlying this success is the self-attention mechanism. In this work, we ask: $\textit{What}$ $\textit{does}$ $\textit{a}$ $\textit{single}$ $\textit{self-attention}$ $\textit{layer}$ $\textit{learn}$ $\textit{from}$ $\textit{next-token}$ $\textit{prediction?}$ We show that training self-attention with gradient descent learns an automaton which generates the next token in two distinct steps: $\textbf{(1)}$ …

abstract advances arxiv attention cs.ai cs.cl cs.lg datasets language language models language processing large datasets math.oc natural natural language natural language processing next prediction processing self-attention simple success token training transformer type work