Patches of Nonlinearity: Instruction Vectors in Large Language Models

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Bigoulaeva, Irina, Rohweder, Jonas, Dutta, Subhabrata, Gurevych, Iryna
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917257570942976
author Bigoulaeva, Irina
Rohweder, Jonas
Dutta, Subhabrata
Gurevych, Iryna
author_facet Bigoulaeva, Irina
Rohweder, Jonas
Dutta, Subhabrata
Gurevych, Iryna
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2602_07930
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Patches of Nonlinearity: Instruction Vectors in Large Language Models
Bigoulaeva, Irina
Rohweder, Jonas
Dutta, Subhabrata
Gurevych, Iryna
Computation and Language
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.
title Patches of Nonlinearity: Instruction Vectors in Large Language Models
topic Computation and Language
url https://arxiv.org/abs/2602.07930