@inproceedings{hwang-etal-2026-instruction,
title = "How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure",
author = "Hwang, Yerin and
Moon, Jiwon and
Lee, Changho and
Ki, Kwangmin and
Kang, Taegwan and
Lee, Jinsik and
Lee, Honglak and
Jung, Kyomin",
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.2261/",
pages = "51406--51436",
ISBN = "979-8-89176-459-0",
abstract = "As large language models (LLMs) increasingly process instructions from multiple sources with different privilege ranks, preserving the priority of higher-rank directives over conflicting lower-rank ones is critical for safe and predictable deployment. However, existing evaluations mostly focus on limited conflict settings, leaving it unclear how models handle instruction hierarchy (IH) when multiple conflicts arise. To address this gap, we present the first systematic study of IH under multi-conflict settings and introduce MIH-Bench, a benchmark that scales the number of simultaneous conflicts across six pairwise hierarchy scenarios. Experiments on 21 LLMs show that IH compliance deteriorates sharply as conflicts accumulate, while neither model scale nor generation reliably predicts robustness. We further find that system-involving conflicts are the most vulnerable, a particularly concerning result given the safety-critical role of system prompts. Moreover, reasoning can improve conflict resolution, although the gains vary across models and are bounded by underlying model capability. These results show that current LLMs remain fragile in pairwise multi-conflict settings, highlighting the need for stronger hierarchy-aware alignment. Code and data are available at https://anonymous.4open.science/r/MIH-release-A5B8/"
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hwang-etal-2026-instruction">
<titleInfo>
<title>How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yerin</namePart>
<namePart type="family">Hwang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiwon</namePart>
<namePart type="family">Moon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Changho</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kwangmin</namePart>
<namePart type="family">Ki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Taegwan</namePart>
<namePart type="family">Kang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinsik</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Honglak</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kyomin</namePart>
<namePart type="family">Jung</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>As large language models (LLMs) increasingly process instructions from multiple sources with different privilege ranks, preserving the priority of higher-rank directives over conflicting lower-rank ones is critical for safe and predictable deployment. However, existing evaluations mostly focus on limited conflict settings, leaving it unclear how models handle instruction hierarchy (IH) when multiple conflicts arise. To address this gap, we present the first systematic study of IH under multi-conflict settings and introduce MIH-Bench, a benchmark that scales the number of simultaneous conflicts across six pairwise hierarchy scenarios. Experiments on 21 LLMs show that IH compliance deteriorates sharply as conflicts accumulate, while neither model scale nor generation reliably predicts robustness. We further find that system-involving conflicts are the most vulnerable, a particularly concerning result given the safety-critical role of system prompts. Moreover, reasoning can improve conflict resolution, although the gains vary across models and are bounded by underlying model capability. These results show that current LLMs remain fragile in pairwise multi-conflict settings, highlighting the need for stronger hierarchy-aware alignment. Code and data are available at https://anonymous.4open.science/r/MIH-release-A5B8/</abstract>
<identifier type="citekey">hwang-etal-2026-instruction</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2261/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51406</start>
<end>51436</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure
%A Hwang, Yerin
%A Moon, Jiwon
%A Lee, Changho
%A Ki, Kwangmin
%A Kang, Taegwan
%A Lee, Jinsik
%A Lee, Honglak
%A Jung, Kyomin
%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 hwang-etal-2026-instruction
%X As large language models (LLMs) increasingly process instructions from multiple sources with different privilege ranks, preserving the priority of higher-rank directives over conflicting lower-rank ones is critical for safe and predictable deployment. However, existing evaluations mostly focus on limited conflict settings, leaving it unclear how models handle instruction hierarchy (IH) when multiple conflicts arise. To address this gap, we present the first systematic study of IH under multi-conflict settings and introduce MIH-Bench, a benchmark that scales the number of simultaneous conflicts across six pairwise hierarchy scenarios. Experiments on 21 LLMs show that IH compliance deteriorates sharply as conflicts accumulate, while neither model scale nor generation reliably predicts robustness. We further find that system-involving conflicts are the most vulnerable, a particularly concerning result given the safety-critical role of system prompts. Moreover, reasoning can improve conflict resolution, although the gains vary across models and are bounded by underlying model capability. These results show that current LLMs remain fragile in pairwise multi-conflict settings, highlighting the need for stronger hierarchy-aware alignment. Code and data are available at https://anonymous.4open.science/r/MIH-release-A5B8/
%U https://aclanthology.org/2026.emnlp-main.2261/
%P 51406-51436
Markdown (Informal)
[How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure](https://aclanthology.org/2026.emnlp-main.2261/) (Hwang et al., EMNLP 2026)
ACL
- Yerin Hwang, Jiwon Moon, Changho Lee, Kwangmin Ki, Taegwan Kang, Jinsik Lee, Honglak Lee, and Kyomin Jung. 2026. How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 51406–51436, Budapest, Hungary. Association for Computational Linguistics.