@inproceedings{chua-etal-2026-hidden,
title = "Hidden Consensus: Preference-Validity Compression in Human Feedback",
author = "Chua, Dorcas Chia Ern and
Lee, Karen Myn Hui and
Tan, Jia Yue and
Gue, Zhen Xue and
Lim, Han Ying and
Hamid, Norzalena Abdul and
Azmi, Azima Binti and
Yeong, Keat Mei and
Mujab, Aizat Izyani Binti and
Azam, Hafsah Noor and
Khoo, Chee Guo and
Chan, Chee Seng",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.48/",
pages = "937--968",
ISBN = "979-8-89176-444-6",
abstract = "Standard RLHF pipelines often reduce heterogeneous human judgments into a single scalar reward target. We argue that this reduction can mis-measure alignment in structurally plural societies, where disagreement may reflect culturally, historically, linguistically, regionally, or normatively grounded interpretations rather than annotation noise. We call this failure Preference-Validity Compression, the collapse of multiple plural-valid response options into a single optimization target. Using Malaysia as a diagnostic setting, we analyze RLHF-style feedback aggregation through preference events linking prompts, responses, and acceptability judgments across interpretive frames. Across 321 preference events from 20 participants and 107 trio-annotated prompts, 79{\%} of prompts contain more than one majority-supported response that single-winner aggregation would discard, and apparent dominance gaps between top responses diminish when all majority-supported options are considered. Participants frequently select multiple acceptable responses, and discarded responses demonstrably reflect coherent local, practical, or cultural frames. These findings show that majority aggregation in this corpus measures $\arg\max$ acceptability rather than plural alignment. We treat this as a measurement-validity issue and argue that future alignment methods should satisfy Validity-Preserving Consistency, remaining stable across plural-valid interpretive frames rather than collapsing them into a single reward target."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chua-etal-2026-hidden">
<titleInfo>
<title>Hidden Consensus: Preference-Validity Compression in Human Feedback</title>
</titleInfo>
<name type="personal">
<namePart type="given">Dorcas</namePart>
<namePart type="given">Chia</namePart>
<namePart type="given">Ern</namePart>
<namePart type="family">Chua</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Karen</namePart>
<namePart type="given">Myn</namePart>
<namePart type="given">Hui</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jia</namePart>
<namePart type="given">Yue</namePart>
<namePart type="family">Tan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhen</namePart>
<namePart type="given">Xue</namePart>
<namePart type="family">Gue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Han</namePart>
<namePart type="given">Ying</namePart>
<namePart type="family">Lim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Norzalena</namePart>
<namePart type="given">Abdul</namePart>
<namePart type="family">Hamid</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Azima</namePart>
<namePart type="given">Binti</namePart>
<namePart type="family">Azmi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Keat</namePart>
<namePart type="given">Mei</namePart>
<namePart type="family">Yeong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aizat</namePart>
<namePart type="given">Izyani</namePart>
<namePart type="given">Binti</namePart>
<namePart type="family">Mujab</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hafsah</namePart>
<namePart type="given">Noor</namePart>
<namePart type="family">Azam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chee</namePart>
<namePart type="given">Guo</namePart>
<namePart type="family">Khoo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chee</namePart>
<namePart type="given">Seng</namePart>
<namePart type="family">Chan</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>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</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">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Standard RLHF pipelines often reduce heterogeneous human judgments into a single scalar reward target. We argue that this reduction can mis-measure alignment in structurally plural societies, where disagreement may reflect culturally, historically, linguistically, regionally, or normatively grounded interpretations rather than annotation noise. We call this failure Preference-Validity Compression, the collapse of multiple plural-valid response options into a single optimization target. Using Malaysia as a diagnostic setting, we analyze RLHF-style feedback aggregation through preference events linking prompts, responses, and acceptability judgments across interpretive frames. Across 321 preference events from 20 participants and 107 trio-annotated prompts, 79% of prompts contain more than one majority-supported response that single-winner aggregation would discard, and apparent dominance gaps between top responses diminish when all majority-supported options are considered. Participants frequently select multiple acceptable responses, and discarded responses demonstrably reflect coherent local, practical, or cultural frames. These findings show that majority aggregation in this corpus measures \arg\max acceptability rather than plural alignment. We treat this as a measurement-validity issue and argue that future alignment methods should satisfy Validity-Preserving Consistency, remaining stable across plural-valid interpretive frames rather than collapsing them into a single reward target.</abstract>
<identifier type="citekey">chua-etal-2026-hidden</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.48/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>937</start>
<end>968</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Hidden Consensus: Preference-Validity Compression in Human Feedback
%A Chua, Dorcas Chia Ern
%A Lee, Karen Myn Hui
%A Tan, Jia Yue
%A Gue, Zhen Xue
%A Lim, Han Ying
%A Hamid, Norzalena Abdul
%A Azmi, Azima Binti
%A Yeong, Keat Mei
%A Mujab, Aizat Izyani Binti
%A Azam, Hafsah Noor
%A Khoo, Chee Guo
%A Chan, Chee Seng
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F chua-etal-2026-hidden
%X Standard RLHF pipelines often reduce heterogeneous human judgments into a single scalar reward target. We argue that this reduction can mis-measure alignment in structurally plural societies, where disagreement may reflect culturally, historically, linguistically, regionally, or normatively grounded interpretations rather than annotation noise. We call this failure Preference-Validity Compression, the collapse of multiple plural-valid response options into a single optimization target. Using Malaysia as a diagnostic setting, we analyze RLHF-style feedback aggregation through preference events linking prompts, responses, and acceptability judgments across interpretive frames. Across 321 preference events from 20 participants and 107 trio-annotated prompts, 79% of prompts contain more than one majority-supported response that single-winner aggregation would discard, and apparent dominance gaps between top responses diminish when all majority-supported options are considered. Participants frequently select multiple acceptable responses, and discarded responses demonstrably reflect coherent local, practical, or cultural frames. These findings show that majority aggregation in this corpus measures \arg\max acceptability rather than plural alignment. We treat this as a measurement-validity issue and argue that future alignment methods should satisfy Validity-Preserving Consistency, remaining stable across plural-valid interpretive frames rather than collapsing them into a single reward target.
%U https://aclanthology.org/2026.findings-aacl.48/
%P 937-968
Markdown (Informal)
[Hidden Consensus: Preference-Validity Compression in Human Feedback](https://aclanthology.org/2026.findings-aacl.48/) (Chua et al., Findings 2026)
ACL
- Dorcas Chia Ern Chua, Karen Myn Hui Lee, Jia Yue Tan, Zhen Xue Gue, Han Ying Lim, Norzalena Abdul Hamid, Azima Binti Azmi, Keat Mei Yeong, Aizat Izyani Binti Mujab, Hafsah Noor Azam, Chee Guo Khoo, and Chee Seng Chan. 2026. Hidden Consensus: Preference-Validity Compression in Human Feedback. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 937–968, Hengqin, China. Association for Computational Linguistics.