Abstract

This paper argues for a pluralist approach to representation in large language models. There are two parts to this pluralism, the first is that we should recognise more than one vehicle of representation in transformer models. Call this vehicle pluralism. Rather than identifying the vehicles of representation with a single component of a system, e.g. individual neurons, patterns of activation, regions in the activation space, we should acknowledge multiple systems of representation within a network operating with different vehicles. The second claim is that we should recognise that there are different formats of representation in transformer models. Transformer models do not operate with a purely analogue, structural, or symbolic architecture but are a hybrid system of representation. Finally, I will discuss how this relates to several working hypotheses about representation that have become adopted in the field of mechanistic interpretability.