@inproceedings{kobayashi-etal-2026-stabilizing,
title = "Stabilizing Hard-Negative Preference Optimization for Medical {LLM}s",
author = "Kobayashi, Kazuma and
Yamagishi, Yosuke and
Shibaki, Ryota and
Sakai, Azusa and
Kodama, Takashi and
Gu, Lin and
Li, Irene and
Arase, Yuki and
Aizawa, Akiko and
Kurohashi, Sadao",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.716/",
pages = "15038--15063",
ISBN = "979-8-89176-462-0",
abstract = "Hard-negative preference optimization should improve support for ground-truth preferred responses while limiting drift from the reference model. In Japanese medical instruction tuning, we find that standard DPO and IPO improve the preferred{--}dispreferred gap but can leave preferred-response likelihood below the reference model and induce growing sequence-level KL divergence. Motivated by this failure-mode analysis, we propose a stabilized IPO objective that combines preferred-response anchoring with explicit KL control. Ablations show that positive anchoring improves preferred-response support, while KL control suppresses reference-model drift; their combination preserves anchoring gains while further reducing drift. These results suggest that reliable medical preference optimization benefits from both preferred-response anchoring and drift control."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kobayashi-etal-2026-stabilizing">
<titleInfo>
<title>Stabilizing Hard-Negative Preference Optimization for Medical LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kazuma</namePart>
<namePart type="family">Kobayashi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yosuke</namePart>
<namePart type="family">Yamagishi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ryota</namePart>
<namePart type="family">Shibaki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Azusa</namePart>
<namePart type="family">Sakai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Takashi</namePart>
<namePart type="family">Kodama</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lin</namePart>
<namePart type="family">Gu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Irene</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuki</namePart>
<namePart type="family">Arase</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Akiko</namePart>
<namePart type="family">Aizawa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sadao</namePart>
<namePart type="family">Kurohashi</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Hard-negative preference optimization should improve support for ground-truth preferred responses while limiting drift from the reference model. In Japanese medical instruction tuning, we find that standard DPO and IPO improve the preferred–dispreferred gap but can leave preferred-response likelihood below the reference model and induce growing sequence-level KL divergence. Motivated by this failure-mode analysis, we propose a stabilized IPO objective that combines preferred-response anchoring with explicit KL control. Ablations show that positive anchoring improves preferred-response support, while KL control suppresses reference-model drift; their combination preserves anchoring gains while further reducing drift. These results suggest that reliable medical preference optimization benefits from both preferred-response anchoring and drift control.</abstract>
<identifier type="citekey">kobayashi-etal-2026-stabilizing</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.716/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>15038</start>
<end>15063</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Stabilizing Hard-Negative Preference Optimization for Medical LLMs
%A Kobayashi, Kazuma
%A Yamagishi, Yosuke
%A Shibaki, Ryota
%A Sakai, Azusa
%A Kodama, Takashi
%A Gu, Lin
%A Li, Irene
%A Arase, Yuki
%A Aizawa, Akiko
%A Kurohashi, Sadao
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F kobayashi-etal-2026-stabilizing
%X Hard-negative preference optimization should improve support for ground-truth preferred responses while limiting drift from the reference model. In Japanese medical instruction tuning, we find that standard DPO and IPO improve the preferred–dispreferred gap but can leave preferred-response likelihood below the reference model and induce growing sequence-level KL divergence. Motivated by this failure-mode analysis, we propose a stabilized IPO objective that combines preferred-response anchoring with explicit KL control. Ablations show that positive anchoring improves preferred-response support, while KL control suppresses reference-model drift; their combination preserves anchoring gains while further reducing drift. These results suggest that reliable medical preference optimization benefits from both preferred-response anchoring and drift control.
%U https://aclanthology.org/2026.findings-emnlp.716/
%P 15038-15063
Markdown (Informal)
[Stabilizing Hard-Negative Preference Optimization for Medical LLMs](https://aclanthology.org/2026.findings-emnlp.716/) (Kobayashi et al., Findings 2026)
ACL
- Kazuma Kobayashi, Yosuke Yamagishi, Ryota Shibaki, Azusa Sakai, Takashi Kodama, Lin Gu, Irene Li, Yuki Arase, Akiko Aizawa, and Sadao Kurohashi. 2026. Stabilizing Hard-Negative Preference Optimization for Medical LLMs. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 15038–15063, Budapest, Hungary. Association for Computational Linguistics.