@inproceedings{kim-etal-2026-hierarchy,
title = "Hierarchy Aware Preference Optimization for the Safety of {K}orean Small Language Models",
author = "Kim, Carl S. and
Koh, Junyoung and
Kim, Kyeonghun and
Hong, Seunghyeok",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.aacl-long.78/",
pages = "1709--1722",
ISBN = "979-8-89176-439-2",
abstract = "Safety alignment of small language models (SLMs) typically incurs an alignment tax: benign prompts are refused along with truly harmful ones. The problem is sharper for Korean SLMs, where limited capacity meets a language whose intent signal is blurred by honorifics. We propose KoSCoPe, which combines a Hierarchical Safety Curriculum, a level-conditional adaptive margin (HARM-DPO), and Counterfactual Consistency Regularization, with a reasoning internalization schedule that recovers 1.5 to 3.2{\texttimes} inference speed. Across five Korean SLMs of 1.7B to 4B, KoSCoPe achieves the strongest Safety/Over-refusal trade-off among the baselines under our matched protocol and remains competitive with general-purpose models several times larger. Under paired-bootstrap tests, its Safety gains over the matched preference-optimization baselines survive multiple-comparison correction for all but the closest competitor, Cat-DPO, where the gain is smaller but directionally consistent. Preliminary cross-lingual validation on Japanese, Mandarin Chinese, and English suggests the same design can transfer beyond Korean, with larger gains where surface politeness markers are more salient."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-etal-2026-hierarchy">
<titleInfo>
<title>Hierarchy Aware Preference Optimization for the Safety of Korean Small Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Carl</namePart>
<namePart type="given">S</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Junyoung</namePart>
<namePart type="family">Koh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kyeonghun</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seunghyeok</namePart>
<namePart type="family">Hong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-439-2</identifier>
</relatedItem>
<abstract>Safety alignment of small language models (SLMs) typically incurs an alignment tax: benign prompts are refused along with truly harmful ones. The problem is sharper for Korean SLMs, where limited capacity meets a language whose intent signal is blurred by honorifics. We propose KoSCoPe, which combines a Hierarchical Safety Curriculum, a level-conditional adaptive margin (HARM-DPO), and Counterfactual Consistency Regularization, with a reasoning internalization schedule that recovers 1.5 to 3.2× inference speed. Across five Korean SLMs of 1.7B to 4B, KoSCoPe achieves the strongest Safety/Over-refusal trade-off among the baselines under our matched protocol and remains competitive with general-purpose models several times larger. Under paired-bootstrap tests, its Safety gains over the matched preference-optimization baselines survive multiple-comparison correction for all but the closest competitor, Cat-DPO, where the gain is smaller but directionally consistent. Preliminary cross-lingual validation on Japanese, Mandarin Chinese, and English suggests the same design can transfer beyond Korean, with larger gains where surface politeness markers are more salient.</abstract>
<identifier type="citekey">kim-etal-2026-hierarchy</identifier>
<location>
<url>https://aclanthology.org/2026.aacl-long.78/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1709</start>
<end>1722</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Hierarchy Aware Preference Optimization for the Safety of Korean Small Language Models
%A Kim, Carl S.
%A Koh, Junyoung
%A Kim, Kyeonghun
%A Hong, Seunghyeok
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-439-2
%F kim-etal-2026-hierarchy
%X Safety alignment of small language models (SLMs) typically incurs an alignment tax: benign prompts are refused along with truly harmful ones. The problem is sharper for Korean SLMs, where limited capacity meets a language whose intent signal is blurred by honorifics. We propose KoSCoPe, which combines a Hierarchical Safety Curriculum, a level-conditional adaptive margin (HARM-DPO), and Counterfactual Consistency Regularization, with a reasoning internalization schedule that recovers 1.5 to 3.2× inference speed. Across five Korean SLMs of 1.7B to 4B, KoSCoPe achieves the strongest Safety/Over-refusal trade-off among the baselines under our matched protocol and remains competitive with general-purpose models several times larger. Under paired-bootstrap tests, its Safety gains over the matched preference-optimization baselines survive multiple-comparison correction for all but the closest competitor, Cat-DPO, where the gain is smaller but directionally consistent. Preliminary cross-lingual validation on Japanese, Mandarin Chinese, and English suggests the same design can transfer beyond Korean, with larger gains where surface politeness markers are more salient.
%U https://aclanthology.org/2026.aacl-long.78/
%P 1709-1722
Markdown (Informal)
[Hierarchy Aware Preference Optimization for the Safety of Korean Small Language Models](https://aclanthology.org/2026.aacl-long.78/) (Kim et al., AACL-IJCNLP 2026)
ACL
- Carl S. Kim, Junyoung Koh, Kyeonghun Kim, and Seunghyeok Hong. 2026. Hierarchy Aware Preference Optimization for the Safety of Korean Small Language Models. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1709–1722, Hengqin, China. Association for Computational Linguistics.