@inproceedings{han-heo-2026-cultureprm,
title = "{C}ulture{PRM}: A Process Reward Model for Mitigating Cultural Overriding in Cultural Reasoning",
author = "Han, Ji-Eun and
Heo, Yoonseok",
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.1228/",
pages = "28023--28040",
ISBN = "979-8-89176-459-0",
abstract = "Recent large reasoning models produce explicit reasoning traces before final answers, yet their behavior in culturally grounded reasoning remains insufficiently understood. In this work, we identify cultural overriding, a step-level failure mode in which a model initially recognizes a relevant local cultural norm but later weakens or replaces it with a more general normative frame, resulting in an incorrect judgment. To address this failure, we introduce CulturePRM, a 1B-parameter process reward model trained on step-annotated synthetic cultural reasoning traces to verify whether intermediate steps preserve, override, or are unrelated to the target cultural norm. At inference time, CulturePRM aggregates step-level rewards to rerank candidate reasoning traces via Best-of-N selection.\par We evaluated CulturePRM across multiple cultural reasoning benchmarks and base reasoning models and observed consistent improvements, with gains of up to 12.06 pp. These gains were not uniform across classes and were concentrated in lower-recall classes, while performance in already strong classes was largely preserved, thereby reducing class-wise disparities. A finer-grained error analysis further revealed substantial recovery among baseline errors alongside a marked reduction in cultural overriding. These findings suggest that culturally aligned reasoning requires models not only to recognize local cultural norms, but also to preserve them throughout the reasoning process."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="han-heo-2026-cultureprm">
<titleInfo>
<title>CulturePRM: A Process Reward Model for Mitigating Cultural Overriding in Cultural Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ji-Eun</namePart>
<namePart type="family">Han</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yoonseok</namePart>
<namePart type="family">Heo</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>Recent large reasoning models produce explicit reasoning traces before final answers, yet their behavior in culturally grounded reasoning remains insufficiently understood. In this work, we identify cultural overriding, a step-level failure mode in which a model initially recognizes a relevant local cultural norm but later weakens or replaces it with a more general normative frame, resulting in an incorrect judgment. To address this failure, we introduce CulturePRM, a 1B-parameter process reward model trained on step-annotated synthetic cultural reasoning traces to verify whether intermediate steps preserve, override, or are unrelated to the target cultural norm. At inference time, CulturePRM aggregates step-level rewards to rerank candidate reasoning traces via Best-of-N selection.\par We evaluated CulturePRM across multiple cultural reasoning benchmarks and base reasoning models and observed consistent improvements, with gains of up to 12.06 pp. These gains were not uniform across classes and were concentrated in lower-recall classes, while performance in already strong classes was largely preserved, thereby reducing class-wise disparities. A finer-grained error analysis further revealed substantial recovery among baseline errors alongside a marked reduction in cultural overriding. These findings suggest that culturally aligned reasoning requires models not only to recognize local cultural norms, but also to preserve them throughout the reasoning process.</abstract>
<identifier type="citekey">han-heo-2026-cultureprm</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1228/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>28023</start>
<end>28040</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T CulturePRM: A Process Reward Model for Mitigating Cultural Overriding in Cultural Reasoning
%A Han, Ji-Eun
%A Heo, Yoonseok
%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 han-heo-2026-cultureprm
%X Recent large reasoning models produce explicit reasoning traces before final answers, yet their behavior in culturally grounded reasoning remains insufficiently understood. In this work, we identify cultural overriding, a step-level failure mode in which a model initially recognizes a relevant local cultural norm but later weakens or replaces it with a more general normative frame, resulting in an incorrect judgment. To address this failure, we introduce CulturePRM, a 1B-parameter process reward model trained on step-annotated synthetic cultural reasoning traces to verify whether intermediate steps preserve, override, or are unrelated to the target cultural norm. At inference time, CulturePRM aggregates step-level rewards to rerank candidate reasoning traces via Best-of-N selection.\par We evaluated CulturePRM across multiple cultural reasoning benchmarks and base reasoning models and observed consistent improvements, with gains of up to 12.06 pp. These gains were not uniform across classes and were concentrated in lower-recall classes, while performance in already strong classes was largely preserved, thereby reducing class-wise disparities. A finer-grained error analysis further revealed substantial recovery among baseline errors alongside a marked reduction in cultural overriding. These findings suggest that culturally aligned reasoning requires models not only to recognize local cultural norms, but also to preserve them throughout the reasoning process.
%U https://aclanthology.org/2026.emnlp-main.1228/
%P 28023-28040
Markdown (Informal)
[CulturePRM: A Process Reward Model for Mitigating Cultural Overriding in Cultural Reasoning](https://aclanthology.org/2026.emnlp-main.1228/) (Han & Heo, EMNLP 2026)
ACL