The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models

Christina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, Jack Lindsey · 2026

Read The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models by Christina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, Jack Lindsey (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-10387.txt.

Introduction

Christina Lu and co authors map the role space of language models. A model can play many roles. Yet it slips back to one base role: the helpful Assistant. That role is set by late train steps. The team ask how that role sits in the model's state space.

They pull out state paths for a range of role types. Each path marks one stock role, from sage to rogue. The base role sits at a set place in that map. The team then test ways to keep it there. Shifts in role can change how a model acts. A firm base role is a safety aim.

Catalog record

CREATORChristina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, Jack Lindsey
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE159 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-10387

Browse the full library · About the Tower of Babel