@inproceedings{maheswaran-desarkar-2026-understanding,
title = "Understanding {\&} Isolating the Assistant State in Instruction-Tuned Models",
author = "Maheswaran, Aishwarya and
Desarkar, Maunendra Sankar",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2133/",
pages = "48491--48511",
ISBN = "979-8-89176-459-0",
abstract = "While instruction-tuned Large Language Models exhibit an ``AI Assistant'' persona, it remains heavily debated whether this behavior is merely one of many simulated roles or the fundamental structural identity of the model. In this work, we present evidence that instruction-tuning instills a global, underlying Assistant State that not only bounds autoregressive generation but is also capable of meta-awareness and reflecting on its own outputs. We demonstrate that this behavior surfaces naturally following the stop token ({\ensuremath{<}}eos{\ensuremath{>}}) without any special training or prompting. Furthermore, we show that this state can be deterministically triggered during generation via a steering vector that isolates the model{'}s post-execution geometry. We introduce this methodology as a zero-shot auditing tool for model evaluators. Applying this framework reveals that models demonstrating Emergent Misalignment in early checkpoints show dual nature where the main generation demonstrates misalignment but post {\ensuremath{<}}eos{\ensuremath{>}} text provides safety warning or rationalizes it as hypothetical."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="maheswaran-desarkar-2026-understanding">
<titleInfo>
<title>Understanding & Isolating the Assistant State in Instruction-Tuned Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aishwarya</namePart>
<namePart type="family">Maheswaran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maunendra</namePart>
<namePart type="given">Sankar</namePart>
<namePart type="family">Desarkar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>While instruction-tuned Large Language Models exhibit an “AI Assistant” persona, it remains heavily debated whether this behavior is merely one of many simulated roles or the fundamental structural identity of the model. In this work, we present evidence that instruction-tuning instills a global, underlying Assistant State that not only bounds autoregressive generation but is also capable of meta-awareness and reflecting on its own outputs. We demonstrate that this behavior surfaces naturally following the stop token (\ensuremath<eos\ensuremath>) without any special training or prompting. Furthermore, we show that this state can be deterministically triggered during generation via a steering vector that isolates the model’s post-execution geometry. We introduce this methodology as a zero-shot auditing tool for model evaluators. Applying this framework reveals that models demonstrating Emergent Misalignment in early checkpoints show dual nature where the main generation demonstrates misalignment but post \ensuremath<eos\ensuremath> text provides safety warning or rationalizes it as hypothetical.</abstract>
<identifier type="citekey">maheswaran-desarkar-2026-understanding</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2133/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>48491</start>
<end>48511</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Understanding & Isolating the Assistant State in Instruction-Tuned Models
%A Maheswaran, Aishwarya
%A Desarkar, Maunendra Sankar
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F maheswaran-desarkar-2026-understanding
%X While instruction-tuned Large Language Models exhibit an “AI Assistant” persona, it remains heavily debated whether this behavior is merely one of many simulated roles or the fundamental structural identity of the model. In this work, we present evidence that instruction-tuning instills a global, underlying Assistant State that not only bounds autoregressive generation but is also capable of meta-awareness and reflecting on its own outputs. We demonstrate that this behavior surfaces naturally following the stop token (\ensuremath<eos\ensuremath>) without any special training or prompting. Furthermore, we show that this state can be deterministically triggered during generation via a steering vector that isolates the model’s post-execution geometry. We introduce this methodology as a zero-shot auditing tool for model evaluators. Applying this framework reveals that models demonstrating Emergent Misalignment in early checkpoints show dual nature where the main generation demonstrates misalignment but post \ensuremath<eos\ensuremath> text provides safety warning or rationalizes it as hypothetical.
%U https://aclanthology.org/2026.emnlp-main.2133/
%P 48491-48511
Markdown (Informal)
[Understanding & Isolating the Assistant State in Instruction-Tuned Models](https://aclanthology.org/2026.emnlp-main.2133/) (Maheswaran & Desarkar, EMNLP 2026)
ACL