The Assistant as a Privileged Persona: A canonical reference in cross-persona self-recognition

Asvin G, [2pt] Institute for Advanced Study, Princeton · 2026

Read The Assistant as a Privileged Persona: A canonical reference in cross-persona self-recognition by Asvin G, [2pt] Institute for Advanced Study, Princeton (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-2606-00545.txt.

Description

G probes how a trained chat model knows its own words across the roles it plays. Such models can spot their own text from a line or two. They can also sense when they write in their main role. Both marks tie to the Assistant role that late training shapes most.

His test runs on a Llama model in a 70B class frame. A grid of claim rates spans roles from book keep to dragon to Shakespeare. On the Assistant's own row, claim rates tie two gap marks in a tight way. Off that row, the tie breaks down in full. What calls the claim there is the judge's shock next to the Assistant's shock. No stand-in role can play that core part, he shows.

Catalog record

CREATORAsvin G, [2pt] Institute for Advanced Study, Princeton
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.00545

Browse the full library · About the Tower of Babel