@inproceedings{yang-zhao-2026-country,
title = "Country Name or Cultural Evidence? A Causal Audit of Population-Grounded Steering in Large Language Models",
author = "Yang, Ziyan and
Zhao, Kuo",
editor = "Xiong, Deyi and
Branco, Ant{\'o}nio and
Zhang, Hongming and
Dong, Yue and
Wang, Benyou and
Zhang, Wenxuan and
Zhou, Li and
Zheng, Jingting",
booktitle = "Proceedings of the First Workshop on Pluralistic Value Alignment of {LLM}s",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.plurvallm-1.13/",
pages = "176--186",
ISBN = "979-8-89176-453-8",
abstract = "When a country name and an observed population distribution point in different directions, which does a large language model follow? We introduce a four-condition counterfactual audit comprising a no-cue baseline, a country label, matched population evidence, and a conflict condition in which the label and evidence refer to different populations. On a quality-controlled diagnostic subset of GlobalOpinionQA, we select 100 population pairs and evaluate both reciprocal directions, yielding 200 directed contrasts and 3,200 likelihood-based predictions from four instruction-tuned models in the 7B-9B range. Country labels yield modest mean gains in fit to the named population (Country Influence: 0.028-0.071), whereas direct population evidence yields substantially larger gains (Evidence Influence: 0.184-0.284). The paired evidence advantage is positive for every model. Under conflict, normalized Evidence Override ranges from 0.543 to 0.728, 85.5-97.5{\%} of predictions are closer to the evidence population, and all four models agree on the evidence side for 161 of 200 contrasts. These results establish a clear hierarchy of steering signals: country names influence model behavior, but explicit population distributions exert far greater influence and usually prevail when the two disagree. The pattern is consistent with evidence-weighted updating from a weaker, defeasible country cue. Our audit provides an evaluation protocol for pluralistic alignment, connecting steerable and distributional pluralism by testing not merely whether a model can be steered toward different populations, but whether its behavior tracks the observed plurality within those populations."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-zhao-2026-country">
<titleInfo>
<title>Country Name or Cultural Evidence? A Causal Audit of Population-Grounded Steering in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ziyan</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kuo</namePart>
<namePart type="family">Zhao</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 First Workshop on Pluralistic Value Alignment of LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Deyi</namePart>
<namePart type="family">Xiong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">António</namePart>
<namePart type="family">Branco</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongming</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Benyou</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenxuan</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Li</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingting</namePart>
<namePart type="family">Zheng</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">979-8-89176-453-8</identifier>
</relatedItem>
<abstract>When a country name and an observed population distribution point in different directions, which does a large language model follow? We introduce a four-condition counterfactual audit comprising a no-cue baseline, a country label, matched population evidence, and a conflict condition in which the label and evidence refer to different populations. On a quality-controlled diagnostic subset of GlobalOpinionQA, we select 100 population pairs and evaluate both reciprocal directions, yielding 200 directed contrasts and 3,200 likelihood-based predictions from four instruction-tuned models in the 7B-9B range. Country labels yield modest mean gains in fit to the named population (Country Influence: 0.028-0.071), whereas direct population evidence yields substantially larger gains (Evidence Influence: 0.184-0.284). The paired evidence advantage is positive for every model. Under conflict, normalized Evidence Override ranges from 0.543 to 0.728, 85.5-97.5% of predictions are closer to the evidence population, and all four models agree on the evidence side for 161 of 200 contrasts. These results establish a clear hierarchy of steering signals: country names influence model behavior, but explicit population distributions exert far greater influence and usually prevail when the two disagree. The pattern is consistent with evidence-weighted updating from a weaker, defeasible country cue. Our audit provides an evaluation protocol for pluralistic alignment, connecting steerable and distributional pluralism by testing not merely whether a model can be steered toward different populations, but whether its behavior tracks the observed plurality within those populations.</abstract>
<identifier type="citekey">yang-zhao-2026-country</identifier>
<location>
<url>https://aclanthology.org/2026.plurvallm-1.13/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>176</start>
<end>186</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Country Name or Cultural Evidence? A Causal Audit of Population-Grounded Steering in Large Language Models
%A Yang, Ziyan
%A Zhao, Kuo
%Y Xiong, Deyi
%Y Branco, António
%Y Zhang, Hongming
%Y Dong, Yue
%Y Wang, Benyou
%Y Zhang, Wenxuan
%Y Zhou, Li
%Y Zheng, Jingting
%S Proceedings of the First Workshop on Pluralistic Value Alignment of LLMs
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-453-8
%F yang-zhao-2026-country
%X When a country name and an observed population distribution point in different directions, which does a large language model follow? We introduce a four-condition counterfactual audit comprising a no-cue baseline, a country label, matched population evidence, and a conflict condition in which the label and evidence refer to different populations. On a quality-controlled diagnostic subset of GlobalOpinionQA, we select 100 population pairs and evaluate both reciprocal directions, yielding 200 directed contrasts and 3,200 likelihood-based predictions from four instruction-tuned models in the 7B-9B range. Country labels yield modest mean gains in fit to the named population (Country Influence: 0.028-0.071), whereas direct population evidence yields substantially larger gains (Evidence Influence: 0.184-0.284). The paired evidence advantage is positive for every model. Under conflict, normalized Evidence Override ranges from 0.543 to 0.728, 85.5-97.5% of predictions are closer to the evidence population, and all four models agree on the evidence side for 161 of 200 contrasts. These results establish a clear hierarchy of steering signals: country names influence model behavior, but explicit population distributions exert far greater influence and usually prevail when the two disagree. The pattern is consistent with evidence-weighted updating from a weaker, defeasible country cue. Our audit provides an evaluation protocol for pluralistic alignment, connecting steerable and distributional pluralism by testing not merely whether a model can be steered toward different populations, but whether its behavior tracks the observed plurality within those populations.
%U https://aclanthology.org/2026.plurvallm-1.13/
%P 176-186
Markdown (Informal)
[Country Name or Cultural Evidence? A Causal Audit of Population-Grounded Steering in Large Language Models](https://aclanthology.org/2026.plurvallm-1.13/) (Yang & Zhao, PlurVA-LLM 2026)
ACL