Patches of Nonlinearity: Instruction Vectors in Large Language Models

TL;DR

This paper investigates how instruction-specific representations, called Instruction Vectors, are formed and used in large language models, revealing their localized nature and complex non-linear interactions across model layers.

Contribution

It introduces a novel method to localize information processing in language models and uncovers the dual linear and non-linear properties of instruction representations.

Findings

Instruction Vectors are localized in models.

IVs exhibit linear separability and non-linear causal interactions.

Different information pathways are activated in later layers based on early task representations.

Abstract

Despite the recent success of instruction-tuned language models and their ubiquitous usage, very little is known of how models process instructions internally. In this work, we address this gap from a mechanistic point of view by investigating how instruction-specific representations are constructed and utilized in different stages of post-training: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Via causal mediation, we identify that instruction representation is fairly localized in models. These representations, which we call Instruction Vectors (IVs), demonstrate a curious juxtaposition of linear separability along with non-linear causal interaction, broadly questioning the scope of the linear representation hypothesis commonplace in mechanistic interpretability. To disentangle the non-linear causal interaction, we propose a novel method to localize information processing in language models that is free from the implicit linear assumptions of patching-based techniques. We find that, conditioned on the task representations formed in the early layers, different information pathways are selected in the later layers to solve that task, i.e., IVs act as circuit selectors.