Multilingual != Multicultural: Evaluating Gaps Between Multilingual Capabilities and Cultural Alignment in LLMs

Multilingual != Multicultural: Evaluating Gaps Between Multilingual
Capabilities and Cultural Alignment in LLMs
Jonathan Rystrøm
Oxford Internet Institute
University of Oxford, UK
Hannah Rose Kirk
Oxford Internet Institute
University of Oxford, UK
Correspondence: jonathan.rystrom@oii.ox.ac.uk
Scott A. Hale
Oxford Internet Institute
University of Oxford, UK
Abstract
Large Language Models (LLMs) are becoming
increasingly capable across global languages.
However, the ability to communicate across
languages does not necessarily translate to
appropriate cultural representations. A key
concern is US-centric bias, where LLMs re-
flect US rather than local cultural values. We
propose a novel methodology that compares
LLM-generated response distributions against
population-level opinion data from the World
Value Survey across four languages (Danish,
Dutch, English, and Portuguese). Using a rig-
orous linear mixed-effects regression frame-
work, we compare three families of models:
Google’s Gemma models (2B–27B parame-
ters), AI2’s OLMo models (7B-32B parame-
ters), and successive iterations of OpenAI’s
turbo-series. Across the families of models,
we find no consistent relationships between
language capabilities and cultural alignment.
While the Gemma models have a positive cor-
relation between language capability and cul-
tural alignment across all languages, the Ope-
nAI and OLMo models are inconsistent. Our
results demonstrate that achieving meaningful
cultural alignment requires dedicated effort be-
yond improving general language capabilities.
1 Introduction
Spearheaded by accessible chat interfaces to pow-
erful models like ChatGPT (OpenAI, 2022), LLMs
are reaching hundreds of millions of users (Milmo,
2023). These models are deployed across di-
verse contexts: from tutoring mathematics (Khan,
2023) to building software applications (Peng et al.,
2023) to assisting in legal cases (Tan et al., 2023).
While most LLMs demonstrate multilingual abili-
ties (Üstün et al.,

are reaching hundreds of millions of users (Milmo,
2023). These models are deployed across di-
verse contexts: from tutoring mathematics (Khan,
2023) to building software applications (Peng et al.,
2023) to assisting in legal cases (Tan et al., 2023).
While most LLMs demonstrate multilingual abili-
ties (Üstün et al., 2024), the ability to communicate
across languages does not necessarily translate into
appropriate cultural representations. Disentangling
language capabilities and cultural alignment is cru-
Figure 1: The relationship between multilingual capa-
bility and cultural alignment is inconsistent across LLM
families, as shown by coefficients from our linear mixed-
effects model (βmultilingual = βf lm; Eq. 3; §3.2).
OpenAI and OLMo models show negative or insignifi-
cant relationships outside of Danish and Dutch, while
Gemma models show positive relationships throughout
(p < .05).
cial for understanding how LLMs should be exam-
ined and audited (Mökander et al., 2024) and for
ensuring these technologies work for diverse peo-
ple (D’ignazio and Klein, 2023; Weidinger et al.,
2022).
Given the Silicon Valley origins of many frontier
AI labs and the prevalence of American English
training data, we might expect LLMs to exhibit
US-centric cultural biases despite their multilin-
gual capabilities. These companies comprise a nar-
row slice of human experience, limiting the voices
that contribute to critical design decisions in LLMs
(D’ignazio and Klein, 2023). They typically train
LLMs on massive amounts of predominantly En-
glish text and employ American crowd workers
to rate and evaluate the LLMs’ responses (John-
son et al., 2022; Kirk et al., 2023). Far too often,
the benefits and harms of data technologies are un-
equally distributed, reinforcing biases and harming
arXiv:2502.16534v2 [cs.CL] 30 Aug 2025

-- 1 of 12 --

already minoritized groups (Birhane, 2020; Milan
and Treré, 2019; Khandelwal et al., 2024). Under-
standing how LLMs represent different

, 2023). Far too often,
the benefits and harms of data technologies are un-
equally distributed, reinforcing biases and harming
arXiv:2502.16534v2 [cs.CL] 30 Aug 2025

-- 1 of 12 --

already minoritized groups (Birhane, 2020; Milan
and Treré, 2019; Khandelwal et al., 2024). Under-
standing how LLMs represent different cultures is
thus paramount to establishing risks of representa-
tional harm (Rauh et al., 2022) and ensuring the
technology’s utility is shared across diverse com-
munities.
Increasing diversity and cross-cultural under-
standing is stymied by unchecked assumptions in
both alignment techniques and evaluation method-
ologies. First, there is an assumption that bigger
and more capable LLMs trained on more data will
be inherently easier to align (Zhou et al., 2023;
Kundu et al., 2023), but this sidesteps the thorny
question of pluralistic variation and cultural repre-
sentations (Kirk et al., 2024b). Thus, it is unclear
whether improvements in architecture (Fedus et al.,
2022) and post-training methods (Kirk et al., 2023;
Rafailov et al., 2023) translate into improvements
in cultural alignment.
Although studies like the World Values Sur-
vey (WVS) have documented how values vary
across cultures (EVS/WVS, 2022), it remains un-
clear whether more capable LLMs—through scal-
ing or improved training—better align with these
cultural differences (Bai et al., 2022; Kirk et al.,
2023). While the WVS has been used in prior
research on values in LLMs, these studies have fo-
cused predominantly on individual models’ perfor-
mance within an English-language context. (Cao
et al., 2023; Arora et al., 2023; AlKhamissi et al.,
2024). This paper addresses this gap by develop-
ing a methodology for assessing how well fami-
lies of LLMs represent different cultural contexts
across multiple languages. We compare two dis-
tinct paths to model improvement: systematic scal-
ing of instruction-tuned models and commercial
product development comprising scaling and inno-
vation

this gap by develop-
ing a methodology for assessing how well fami-
lies of LLMs represent different cultural contexts
across multiple languages. We compare two dis-
tinct paths to model improvement: systematic scal-
ing of instruction-tuned models and commercial
product development comprising scaling and inno-
vation in post-training to accommodate pressures
from capabilities, cost, and preferences (OpenAI
et al., 2024b).
Given these considerations, we investigate the
following research questions:
RQ1 Multilingual Cultural Alignment: Does
improved multilingual capability increase
LLM alignment with population-specific
value distributions?
RQ2 US-centric Bias: When using different lan-
guages, do LLMs align more with US values
or with values from the countries where these
languages are native?
We operationalise multilingual capability as an
LLM’s performance on a range of multilingual
benchmarks across languages (see, e.g., Nielsen,
2023). We describe the specific benchmarks and
performances in the supplementary materials.
This work makes several key contributions. First,
we introduce a novel distribution-based method-
ology for probing cultural alignment across lan-
guages, moving beyond direct survey approaches
to better capture latent cultural values (Sorensen
et al., 2024). Second, we provide the first sys-
tematic comparison of how improvements in scale
and post-training affect cultural alignment and US-
centric bias across English, Danish, Dutch, and Por-
tuguese through a series of robust statistical mod-
els. Third, we release a dataset of model-generated
responses across multiple languages and cultural
contexts as well as our code, enabling future re-
search into cultural alignment and bias.1 Together,
these contributions advance our understanding of
how LLM development choices influence cultural
representation while providing tools for ongoing
investigation of these critical issues.
2 Measuring Cultural Alignment
Figure 2: Pearson correlations in value

future re-
search into cultural alignment and bias.1 Together,
these contributions advance our understanding of
how LLM development choices influence cultural
representation while providing tools for ongoing
investigation of these critical issues.
2 Measuring Cultural Alignment
Figure 2: Pearson correlations in value polarity scores
across studied countries from the World Values Survey.
Value polarity scores are the fraction of the population
in favour of a given topic. All correlations are positive,
with most being between 0.7–0.95.
This section defines ‘cultural alignment’ and
how to measure it in LLMs. We conceptualise
cultural alignment as reproducing distributions of
1See github.com/jhrystrom/multicultural-alignment for
code, data, and supplementary materials.

-- 2 of 12 --

values in a particular population. Then we show
how to a) get a ground-truth distribution of values
using the World Values Survey (§2.1) and b) elicit
value distributions from LLMs (§2.2).
Cultural alignment as value reproduction:
Within a culture there will be a variety of stances to
any particular topic. However, the distribution of
stances will be characteristic among cultures. For
instance, while around 8% of Danes are opposed to
abortion, it is a much less contentious topic than in
the US, where it’s close to 40% (EVS/WVS, 2022).
We posit that cultural alignment for a specific
group of people can be operationalised as how well
an LLM reproduces the distribution of values over
a wide range of topics (Sorensen et al., 2024). In-
vestigating distributions of responses differs from
previous work that directly surveys the LLMs as
regular participants (e.g., Cao et al., 2023). This
approach also addresses concerns raised by Khan
et al. (2025) about the instability of survey-based
evaluations by focusing on aggregate distributions
rather than individual responses and incorporating
explicit controls for response consistency. Our goal
is to get more naturalistic elicitations of the

al., 2023). This
approach also addresses concerns raised by Khan
et al. (2025) about the instability of survey-based
evaluations by focusing on aggregate distributions
rather than individual responses and incorporating
explicit controls for response consistency. Our goal
is to get more naturalistic elicitations of the un-
derlying values whilst avoiding sycophancy and
response bias (Sharma et al., 2023).
We operationalise reproduction as high correla-
tions between value polarity scores: the fraction
of people (or LLM responses) in favour of a topic
in the population. Note, that we binarise issues to
allow for simpler operationalisation. Below, we
describe how we empirically estimate the value po-
larity score for the ground truth (§2.1) and LLMs
(§2.2).
2.1 Ground Truth: World Values Survey
To get a ‘ground truth’ distribution of cultural val-
ues, we use the joint World Values Survey and
European Values Survey (EVS; EVS/WVS, 2022).
These surveys cover adults across 92 countries with
samples that are nationally representative for gen-
der, age, education, and religion. The surveys’
broad coverage enables cross-cultural comparabil-
ity for the many countries covered by the surveys,
though some scholars note challenges in ensuring
response comparability across countries (Alemán
and Woods, 2016). The WVS provides both coun-
try and language identifiers for each respondent,
allowing us to define populations either as citizens
of a country or speakers of a language using the
same underlying respondent-level data.
We select questions with binary agree/disagree
or rating scale formats that allow clear classifica-
tion of positive vs. negative stances, excluding
questions with multiple categorical response op-
tions (see the supplementary materials for the full
list of questions). These questions span environ-
ment, work, family, politics, religion, and security.
We convert responses to binary indicators by deter-
mining whether each response indicates support for
the

cluding
questions with multiple categorical response op-
tions (see the supplementary materials for the full
list of questions). These questions span environ-
ment, work, family, politics, religion, and security.
We convert responses to binary indicators by deter-
mining whether each response indicates support for
the measured construct, with custom coding to han-
dle the various question formats and reverse-scored
items. Finally, we calculate the value polarity score
as the demographically weighted proportion of re-
spondents with affirmative stances. Formally, we
can define the value polarity score for a given pop-
ulation, P (e.g., citizens in a country or speakers
of a language) and topic, q, (i.e., question within
the EVS/WVS) as shown in Eq. 1:
VPSP,q = X
i∈P
wi
P
j∈Pq wj
Ai,q (1)
Here, Ai,q is a binary indicator of whether par-
ticipant i has a positive stance on topic q, wi rep-
resents the survey-provided demographic weights,
and Pq denotes respondents in population P who
answered question q. The first term normalises
the weights to account for missing responses and
enables aggregation across any definition of a pop-
ulation (e.g., residents in a country, speakers of a
language, etc.).
For example, if 80% of Danish respondents
who answered the same-sex marriage question ex-
pressed support (after demographic reweighting),
Denmark’s value polarity score for this topic would
be 0.8. Thus, a culture’s values can be represented
as a vector, where each element corresponds to a
value polarity score for a specific topic.
2.2 Ecologically valid LLM responses
Testing cultural alignment effectively requires em-
bedding contextual and cultural elements in ways
that maintain ecological validity. At a high level,
eliciting values from an LLM consist of two steps:
1) Iteratively prompting the model with the se-
lected topics and 2) extracting the stances from
each model response.
Setting prompt context: Developing ecologi-
cally valid prompts requires careful

elements in ways
that maintain ecological validity. At a high level,
eliciting values from an LLM consist of two steps:
1) Iteratively prompting the model with the se-
lected topics and 2) extracting the stances from
each model response.
Setting prompt context: Developing ecologi-
cally valid prompts requires careful consideration.
When evaluating LLM responses to value-laden
topics, simply asking questions like “What propor-
tion of people support topic X?” or “Do you support

-- 3 of 12 --

topic X?” proves inadequate (e.g., Rozado, 2024).
Such direct approaches suffer from three key lim-
itations: they generate false positives through ex-
cessive agreement, fail to reflect realistic usage
patterns, and provide insufficient variation to as-
sess cultural alignment (Röttger et al., 2024). They
also struggle to capture instance-specific harms that
emerge when systems misalign with users’ cultural
contexts (Rauh et al., 2022).
Instead, we adopt an implicit approach by asking
the model to generate responses from hypothetical
respondents. For example, prompting “imagine
surveying 10 random people on topic X. What are
their responses?” This method reveals the model’s
latent opinion distribution while avoiding the lim-
itations of direct questioning. Details for prompt
construction are provided in the supplementary ma-
terials.
Seeding cultural responses: Having a method
for eliciting distributions of values, the next step is
to seed culture. One typical way of seeding a spe-
cific culture is to explicitly instruct the LLM either
by mentioning a specific country (‘imagine survey-
ing 10 random Americans’) or through describing
specific personas (‘Imagine surveying a 85-year-
old Danish woman...’; AlKhamissi et al., 2024).
The problem with these demographic prompting
approaches is that they stray from actual uses
of LLMs. Users are unlikely to explicitly men-
tion their demographic information or nationality
(Zheng et al., 2023a).
Instead, we use language as a proxy for

urveying a 85-year-
old Danish woman...’; AlKhamissi et al., 2024).
The problem with these demographic prompting
approaches is that they stray from actual uses
of LLMs. Users are unlikely to explicitly men-
tion their demographic information or nationality
(Zheng et al., 2023a).
Instead, we use language as a proxy for cultural
origin. For instance, a prompt in Danish is as-
sumed to come from a Dane. This approach creates
an intentional distinction in our analysis: we can
compare ‘language-level’ alignment (all speakers
of a language globally) with ‘country-level’ align-
ment (all people from specific nations where that
language is native). As argued by Havaldar et al.
(2023), users speaking a particular language would
expect culturally appropriate responses in that lan-
guage. For languages spoken in multiple countries,
this approach is intentionally ambiguous. The am-
biguity allows us to elicit the underlying ‘default’
alignment rather than the general ability to emu-
late cultures (Tao et al., 2024). We validate this
approach by showing that LLM responses exhibit
significantly lower self-consistency between lan-
guages compared to within languages, demonstrat-
ing that language impacts output (see the supple-
mentary materials). To create prompts across lan-
guages, we use gpt-3.5-turbo to translate our
original English prompts. Although previous liter-
ature has shown strong translation capabilities in
LLMs (Yan et al., 2024), we nonetheless manually
verify the translations.
Annotating and aggregating responses: Fi-
nally, to transform the LLMs’ hypothetical sur-
vey responses into vectors of stances, we use an
LLM-as-a-judge approach (Zheng et al., 2023b;
Guerdan et al., 2025). Specifically, we use
gpt-4.1-mini (OpenAI et al., 2025) to label
each substatement as either ‘pro’, ‘con’, or ‘null’
given the context of the topic and a representative
pro and con statement (generated with an LLM and
validated by the authors). We then calculate the
proportion of ‘pro’

., 2023b;
Guerdan et al., 2025). Specifically, we use
gpt-4.1-mini (OpenAI et al., 2025) to label
each substatement as either ‘pro’, ‘con’, or ‘null’
given the context of the topic and a representative
pro and con statement (generated with an LLM and
validated by the authors). We then calculate the
proportion of ‘pro’ versus ‘con’ responses as the
LLM’s value polarity score for the given statement.
For instance, a response with seven ‘pro’, one ‘con’,
and two ‘null’ statement would yield a value po-
larity score of 0.875 ( 7
8 ). A complete, unabridged
example can be found in the supplementary materi-
als. Formally, we label each substatement from the
full set of hypothetical statements, Gq,g, for topic
q and generation g as r. Furthermore, we label
the classifier as ℓ(r). We then formalise the value
polarity score for a given instance of a generation
for a topic (VPSLLM
q,g ) as shown in Eq. 2:
VPSLLM
q,g =
P
r∈Gq,g [ ℓ(r) = pro ]
P
r∈Gq,g [ ℓ(r) ∈ {pro, con} ] , (2)
These scores are then compared against the value
polarity scores from the WVS. Specifically, we
calculate the Spearman rank correlation to obtain
a measure of similarity between the LLMs’ re-
sponses and the value distributions of a given pop-
ulation.
To validate the LLM-as-judge, we manually an-
notate 200 statements. We iteratively refine the
prompts and the LLM used until we reach satisfac-
tory performance. We find a 91% agreement and a
mean absolute error for value polarity of 4.5% over
the dataset, ensuring consistent statistics between
LLM and human annotation (Guerdan et al., 2025).
3 Experimental Setup
To investigate whether improving the multilingual
capabilities of LLMs improves cultural alignment,
we set up an experiment using a carefully chosen
set of models and languages. We examine two

-- 4 of 12 --

Figure 3: Self-consistency in responses for LLMs and
WVS countries. LLMs have lower self-consistency
than resampled WVS responses—shown by the dashed
lines—particularly in non-English

proves cultural alignment,
we set up an experiment using a carefully chosen
set of models and languages. We examine two

-- 4 of 12 --

Figure 3: Self-consistency in responses for LLMs and
WVS countries. LLMs have lower self-consistency
than resampled WVS responses—shown by the dashed
lines—particularly in non-English languages.
different kinds of model improvements: scaling
and commercial product development. These cases
provide complementary perspectives on the effects
of multilingual capabilities on cultural alignment.
Scaling is the most well-studied path to improv-
ing LLMs (Kaplan et al., 2020; Ganguli et al.,
2022). Commercial product development, on the
other hand, comprises both scale and innovation in
post-training to accommodate different pressures
from capabilities, cost, and preferences (Kirk et al.,
2024a). For scaling, we use the instruction-tuned
Gemma models (Gemma et al., 2024) and OLMo-
2 models (OLMo et al., 2025), while for product
development, we use OpenAI’s turbo-series mod-
els (OpenAI, 2022; OpenAI et al., 2024a,b). We
provide details of these model families in §3.1. A
breakdown of the computational cost is in the sup-
plementary materials.
Languages: For the languages, we compare En-
glish with Danish, Dutch, and Portuguese. This
set allows us to test multiple assumptions about
cultural alignment. English represents a widely
used case: it is a global language with speakers
across many countries represented in the WVS (see
Fig. 2). This diversity allows us to assess whether
LLMs align more strongly with US values or those
of other English-speaking nations.
Danish and Dutch serve as controlled test cases
since they are primarily used in a single country.
If cultural alignment stems from pre-training data,
models should show strong Danish/Dutch cultural
alignment when using these languages, despite
their small share of training data (Kreutzer et al.,
2022). Alternatively, if alignment emerges from
post-training processes—which are

re primarily used in a single country.
If cultural alignment stems from pre-training data,
models should show strong Danish/Dutch cultural
alignment when using these languages, despite
their small share of training data (Kreutzer et al.,
2022). Alternatively, if alignment emerges from
post-training processes—which are predominantly
English-based (Blevins and Zettlemoyer, 2022)–
responses in these languages should align more
with US values.
Portuguese presents an interesting case since it is
an official language in several countries. We inves-
tigate whether the LLM responses are more aligned
to Portugal or Brazil—two countries that show dis-
tinct value patterns in relation to each other and
the US (see Fig. 2). This allows us to test whether
an LLM aligns more strongly with one country’s
values, the aggregate values of all language users,
or US values.
For each language-model pair, we collect 300
prompt-response pairs to power our statistical anal-
ysis sufficiently (see §3.2). After filtering out re-
sponses that either lacked the required hypothetical
survey format or were in a language other than
the prompt, we obtained between 111–299 valid
responses per combination. We calculate the corre-
lation in value polarity scores at three levels: coun-
try (e.g., US or Denmark), language (pooling all
speakers of a given language), and global (weighted
values from all WVS/EVS participants).
3.1 Models
We examine three model families representing dif-
ferent development approaches: Gemma (Gemma
et al., 2024) and OLMo (OLMo et al., 2025) for
improvements through scaling and OpenAI’s turbo
series for commercial product development, com-
bining scaling with post-training improvements
(OpenAI, 2022; OpenAI et al., 2024a,b). Other pre-
liminary experiments included different versions of
LLaMA models (Touvron et al., 2023) and Mistral
models (Jiang et al., 2023). However, these mod-
els either failed to consistently follow instructions
or always answered in English regardless

st-training improvements
(OpenAI, 2022; OpenAI et al., 2024a,b). Other pre-
liminary experiments included different versions of
LLaMA models (Touvron et al., 2023) and Mistral
models (Jiang et al., 2023). However, these mod-
els either failed to consistently follow instructions
or always answered in English regardless of the
prompt language. See the supplementary materials
for a more thorough description of the LLMs.
3.2 RQ1: Multilingual Cultural Alignment
To statistically assess whether improving the mul-
tilingual capabilities of LLMs improves cultural
alignment, we construct a linear mixed-effects re-
gression (LMER; Luke, 2017) based on the experi-
mental setup described above. Our LMER follows
standard practices and has three core components:
• Core coefficient: The coefficient of interest

-- 5 of 12 --

Figure 4: Language capability (x-axis) vs cultural align-
ment scores (y-axis) across languages. Stars indicate
significance (p < .05) in our linear mixed-effects regres-
sion of multiple runs (See §3.2). OpenAI models (blue)
and OLMo models (red) show negative/insignificant
relationships outside of English, while the Gemma
models (green) show positive relationships throughout
(p < .05).
is the three-way interaction between model
family, language, and multilingual capability.
This tests whether the multilingual capabil-
ity–alignment relationship differs by model
family and response language, directly ad-
dressing RQ1.
• Random effects: We include a model-specific
random intercept αj to account for repeated
measures of cultural alignment for the same
LLM. This models variation between LLMs
and can improve efficiency over standard lin-
ear regressions (Luke, 2017).
• Control for self-consistency: We include
a consistency-by-language term to help en-
sure that higher alignment scores reflect gen-
uine cultural adaptation rather than reduced
response noise, which can inflate scores (Kah-
neman et al., 2021).
We calculate self-consistency as the

d lin-
ear regressions (Luke, 2017).
• Control for self-consistency: We include
a consistency-by-language term to help en-
sure that higher alignment scores reflect gen-
uine cultural adaptation rather than reduced
response noise, which can inflate scores (Kah-
neman et al., 2021).
We calculate self-consistency as the Spearman
correlation between value polarity scores (defined
in §2) of repeated responses to identical topics, ad-
justed by the reliability of the LLM annotation (see
§2.2; Charles, 2005). A score of 1.0 indicates per-
fect consistency; 0.0 indicates random responses.
Population-level resampling of the human WVS
responses yields values between 0.66 and 0.84 (see
Fig.3 and the supplementary materials).
Formally, the model is specified in Eq. 3:
CAi ∼ N (μi, σ2),
μi = αj[i] + β1l Xcons,iXl,i
+ βf lm Xm,iXf,iXl,i,
αj ∼ N (μα, σ2
α), j = 1, . . . , J.
(3)
where i indexes responses and j[i] denotes the
LLM producing response i. Here Xcons,i is the self-
consistency score for response i, Xl,i is the set of
language indicators, Xf,i is the set of model-family
indicators, and Xm,i is the multilingual capability
score. The residual variance σ2 represents within-
LLM variation in alignment scores not explained by
the fixed effects or model-specific intercept, while
σ2
α represents between-LLM variation in average
alignment.
The above statistical model allows us to analyse
the relationship between multilingual capabilities
and cultural alignment in model families at the level
of individual languages. For example, we might
find that multilingual capabilities improve cultural
alignment for Gemma models for Danish but not
for Dutch or vice versa.
3.3 RQ2: US-Centric Bias
We analyse model bias by comparing cultural align-
ment between US and local values, where “local”
refers to values in the country or countries where
a given language is natively spoken. We define
US-centric bias as an LLM showing higher cultural
alignment with US value distributions compared

3.3 RQ2: US-Centric Bias
We analyse model bias by comparing cultural align-
ment between US and local values, where “local”
refers to values in the country or countries where
a given language is natively spoken. We define
US-centric bias as an LLM showing higher cultural
alignment with US value distributions compared to
local ones. To quantify this bias, we use a linear re-
gression model that measures the differential effect
of US versus local value alignment:
CA = β0 + β1(US)
+ X
m∈M
X
l∈L
βml(m × l)
+ X
m∈M
X
l∈L
βUS
ml (US ×m × l) + ϵ
(4)
The regression’s intercept (β0,i.e., the base case)
is a baseline that produces uniformly random value
polarity scores. M is the set of models and L is the
set of languages. US is a boolean feature denoting

-- 6 of 12 --

whether the cultural alignment is to the US (if 1)
or the local values (if 0). We primarily analyse the
coefficients with US (βU S
ml ) since these provide the
partial effect of US-centric bias, i.e., how much
more/less a given LLM is aligned to US rather than
local values. Assumption checks for the regression
can be seen in the supplementary materials.
4 Results
4.1 Multilingual Cultural Alignment (RQ1)
We first examine the stability of LLMs’ cultural
values. For LLMs lacking stable internal values,
apparent improvements in cultural alignment may
reflect reduced response variance rather than gen-
uine advances (Röttger et al., 2024; Kahneman
et al., 2021). We therefore analyse both the self-
consistency of LLM responses and how alignment
changes with model improvements.
LLMs have low self-consistency: We find low
self-consistency scores across all models and lan-
guages compared to human responses in the WVS
data (Fig. 3). In contrast, LLMs show gener-
ally lower self-consistency compared to the hu-
man responses, even in English, where instruction-
following capabilities are strongest due to English-
dominated training data. (OpenAI et al., 2024a;
Gemma et al., 2024; OLMo et al., 2025).
This lower

an responses in the WVS
data (Fig. 3). In contrast, LLMs show gener-
ally lower self-consistency compared to the hu-
man responses, even in English, where instruction-
following capabilities are strongest due to English-
dominated training data. (OpenAI et al., 2024a;
Gemma et al., 2024; OLMo et al., 2025).
This lower self-consistency complicates our cul-
tural alignment analysis (Wright et al., 2024).
Drawing on Kahneman et al. (2021)’s noise frame-
work, we recognise that inconsistent responses can
be as detrimental as bias with respect to the ac-
curacy of the analysis. To address the noise, we
employ larger sample sizes and incorporate consis-
tency controls in our regression analyses.
Multilinguality does not imply cultural align-
ment: The relationship between model improve-
ments and cultural alignment varies substantially
across languages and model families (Fig. 1). For
Gemma, there is a strong and significant positive
relationship between multilingual capabilities and
cultural alignment for all languages. In contrast,
the relationships for the GPT-Turbo models are
either insignificant or negative. For Dutch and
Danish the relationships are insignificant (βgpt,nl =
0.049, p = 0.589,βgpt,da = 0.053, p = 0.522), and
for Portuguese and English the effect is signifi-
cant and negative (βgpt,en = −0.24, p = 0.009,
βgpt,pt = −0.30, p < 0.001). Similarly for
OLMo, the relationship is positive for Danish and
Dutch (βOLMo,da = 0.44, p < 0.001, βOLMo,nl =
0.29, p < 0.001) and insignificant for English
and Portuguese (βOLMo,en = 0.068, p = 0.115,
βOLMo,pt = 0.008, p = 0.825).
The mismatch between multilingual perfor-
mance and cultural alignment could suggest a
capability threshold: multilingual improvements
might provide rudimentary instruction following
skills (Nie et al., 2024), but beyond a point, other
factors—such as the preferences of developers and
annotators—dominate (Kirk et al., 2024b). This
could explain the smaller open weights models’
higher

ent could suggest a
capability threshold: multilingual improvements
might provide rudimentary instruction following
skills (Nie et al., 2024), but beyond a point, other
factors—such as the preferences of developers and
annotators—dominate (Kirk et al., 2024b). This
could explain the smaller open weights models’
higher coefficients than the gpt-turbo models (see
Fig. 4 or Fig. 1). Further work is needed to under-
stand alignment at the sub-national level.
Furthermore, the strong effect of self-
consistency (0.405 < βconsistency < 0.723, p ≪
0.001) compared to multilingual capability
suggests that noise remains a major limiting factor
in analysing cultural alignment. This aligns with
broader findings about the instability of LLM value
elicitation (Röttger et al., 2024; Khan et al., 2025).
Moreover, even the highest observed alignment
scores (around 0.7; see Fig 4) indicate substantial
room for improvement in how well LLMs match
human cultural values and behaviours.
In conclusion, our analysis reveals a complex
relationship between model improvements and cul-
tural alignment. Although some languages show
progressive improvements in cultural alignment
from model scaling or iterative commercial devel-
opment, others show minimal or inconsistent im-
provements. These findings, combined with the
relatively low self-consistency of LLM responses,
demonstrate that improved multilingual capability
does not guarantee better cultural alignment.
4.2 US-centric Bias (RQ2)
Here, we answer RQ2 by examining US bias across
languages. Specifically, we investigate relative
alignment between local and US values (Fig. 5).
Our analysis reveals distinct patterns of US-
centric bias across both languages and model fam-
ilies (Fig. 5). Languages show different suscepti-
bilities to US bias: only one of nine LLMs exhibits
US-centric bias in Danish, all in English, all in Por-
tuguese, and none in Dutch. Note that for English,
these results mean that the LLM, on average, is
relatively more

tric bias across both languages and model fam-
ilies (Fig. 5). Languages show different suscepti-
bilities to US bias: only one of nine LLMs exhibits
US-centric bias in Danish, all in English, all in Por-
tuguese, and none in Dutch. Note that for English,
these results mean that the LLM, on average, is
relatively more aligned to US values compared to
other English-speaking countries like Kenya or the
United Kingdom. See the supplementary materials

-- 7 of 12 --

Figure 5: US-centric bias coefficients across LLMs and
languages (βBiasU S ); see Eq. 4). Error bars are stan-
dard errors from the regression. Positive values indicate
the presence of US-centric bias.
for detailed results.
The overarching pattern is that languages spoken
across countries (English and Portuguese) show
US-centric bias, whereas languages spoken in only
one country (Danish and Dutch) show less US-
centric bias. This supports the hypothesis that
homogeneity in the training data can counteract
US-centric bias—at least for medium-resourced,
Western-European languages.
For LLMs, some specific LLMs seem more
prone to bias across languages. Specifically,
the small gemma-2-2b-it exhibits higher US-
centric bias across every language except Dutch.
Beyond that, we see no clear progressions in US-
centric bias within any family.
In conclusion, language seems a stronger indica-
tor of US-centric bias in LLMs compared to LLM
development. Monocultural languages show in-
significant to negative bias, while English and Por-
tuguese show significant US-centric bias. Within
each LLM family, we find no consistent nor sig-
nificant change in US-centric bias across LLM
versions. These findings underscore the complex
relationship between multilingual capability and
alignment.
5 Related Work
Recent work emphasizes the need for systematic
auditing of LLMs’ cultural alignment, particularly
as these models are deployed globally (Kirk et al.,
2024a; Mökander et al., 2024; Kirk et al., 2024b).
Prior empirical

ings underscore the complex
relationship between multilingual capability and
alignment.
5 Related Work
Recent work emphasizes the need for systematic
auditing of LLMs’ cultural alignment, particularly
as these models are deployed globally (Kirk et al.,
2024a; Mökander et al., 2024; Kirk et al., 2024b).
Prior empirical approaches have primarily taken
two paths: using transformations based on Hofst-
ede’s cultural dimensions framework or directly
comparing against survey responses. Studies us-
ing Hofstede’s dimensions (Masoud et al., 2025;
Cao et al., 2023) provide structured cross-cultural
comparisons through latent variable analysis. How-
ever, these studies assume that LLMs’ latent dimen-
sions map directly onto human dimensions, since
they use formulas calibrated for humans—an as-
sumption that warrants scrutiny (Shanahan, 2024;
Schröder et al., 2025).
Recent work has explored using LLMs to sim-
ulate responses for assessing cultural alignment
(Tao et al., 2024; AlKhamissi et al., 2024; Havaldar
et al., 2023). Similarly to our work, these works
show that LLMs struggle to represent underrep-
resented personas (AlKhamissi et al., 2024) and
emotions (Havaldar et al., 2023) for non-English
languages. Prior approaches focused on individual-
level responses. In contrast, our method generates
distributions of opinions across hypothetical sur-
vey participants, enabling direct comparison with
population-level statistics. This distribution-based
approach offers three key advantages. First, it bet-
ter captures the inherent variation in cultural values
within populations, paving the way for investigat-
ing distributional alignment (Sorensen et al., 2024).
Second, it enables principled statistical comparison
against large-scale survey data like the World Val-
ues Survey (EVS/WVS, 2022). Finally, the frame-
work is easy to extend to new languages by auto-
matically translating the prompts. We detail our
quantitative framework for measuring alignment
with observed population

ond, it enables principled statistical comparison
against large-scale survey data like the World Val-
ues Survey (EVS/WVS, 2022). Finally, the frame-
work is easy to extend to new languages by auto-
matically translating the prompts. We detail our
quantitative framework for measuring alignment
with observed population distributions in §2.
There is also an increasing body of work inves-
tigating political biases in LLMs (Röttger et al.,
2024, 2025; Rozado, 2024). Much of this work also
relies on human political surveys like the Political
Compass Test. However, recent work has called for
increased attention to how the randomness inher-
ent in LLM decoding at non-zero temperatures can
create instability in attributes (Röttger et al., 2024;
Wright et al., 2024; Khan et al., 2025). We expand
on this work by including multilingual perspectives
and constructing prompts with a wide range of vari-
ations (see §2). These prompt variations, combined
with statistically accounting for self-consistency in
our statistical analysis (see §3.2), allow us to get a
more robust measure of cultural alignment.
The relationship between model capabilities
and cultural alignment remains understudied. Un-
like general performance metrics that follow pre-
dictable scaling laws (Kaplan et al., 2020), cultural
alignment may not improve systematically with
model capabilities. This aligns with research show-

-- 8 of 12 --

ing micro-level capabilities can be discontinuous
with scale (Ganguli et al., 2022). The challenge
is compounded in multilingual settings (Hoffmann
et al., 2022), where static benchmarks with single
correct answers fail to capture how cultural values
are distributed across different topics and contexts.
Previous work has focused primarily on English-
language performance (Tao et al., 2024) or indi-
vidual LLMs (Arora et al., 2023; Cao et al., 2023).
Our work extends this by examining how cultural
alignment systematically varies within model fam-
ilies and across languages,

ibuted across different topics and contexts.
Previous work has focused primarily on English-
language performance (Tao et al., 2024) or indi-
vidual LLMs (Arora et al., 2023; Cao et al., 2023).
Our work extends this by examining how cultural
alignment systematically varies within model fam-
ilies and across languages, providing insight into
how different development approaches—scaling
and commercial product development—influence
cultural representation capabilities.
There is already progress on improving the cross-
cultural participation in alignment data. Two no-
table projects are PRISM and AYA (Kirk et al.,
2024b; Üstün et al., 2024). PRISM is a large dataset
of conversational preferences from a diverse par-
ticipant pool. While the data is predominantly in
English, it could be an important resource for bet-
ter understanding and modelling diverse cultural
preferences. The AYA dataset is a massively multi-
lingual instruction fine-tuning dataset. AYA could
provide further means of realising the demonstrated
benefits of multilingual training (Nie et al., 2024).
6 Conclusion
Increased multilingual capabilities do not guaran-
tee improved cultural alignment in Large Language
Models. Through systematic comparison of three
model families—Gemma, OLMo, and OpenAI’s
GPTs—we find that the relationship between im-
provements in multilingual capability and cultural
alignment is complex. While some languages show
clear improvements in alignment with increased
model capabilities (e.g., Danish), others exhibit
inconsistent patterns, suggesting that cultural align-
ment does not automatically follow gains in mul-
tilingual capabilities. Our distribution-matching
methodology using World Values Survey data en-
abled the detection of these nuanced patterns across
languages and cultural contexts.
We also find that, contrary to popular discourse,
LLMs do not exhibit US-centric bias across all
languages; in Danish and Dutch, they align more
closely with the values of Denmark and the

ethodology using World Values Survey data en-
abled the detection of these nuanced patterns across
languages and cultural contexts.
We also find that, contrary to popular discourse,
LLMs do not exhibit US-centric bias across all
languages; in Danish and Dutch, they align more
closely with the values of Denmark and the Nether-
lands, respectively, than with the US. This fits with
the hypothesis that more culturally uniform data
leads to less US-centric bias. Both English and Por-
tuguese are spoken in multiple countries, whereas
Dutch and Danish are predominantly spoken in one.
To further validate this claim, future work could
include other multi-cultural languages (like Span-
ish or Swahili) and monocultural languages (like
Japanese)—especially with a wider geographical
reach to preclude European bias.
Our findings highlight that improving cultural
alignment requires dedicated effort beyond general
capability scaling. Future work should focus on
developing techniques that can better handle align-
ment with distributions of cultural values rather
than single points, while ensuring meaningful par-
ticipation from diverse communities in LLM devel-
opment. As these models continue to reach wider
audiences spanning many geographic and cultural
regions, achieving robust cultural alignment be-
comes increasingly crucial for equitable deploy-
ment.
Acknowledgements
We are thankful for the helpful feedback from the
anonymous reviewers. We also thank Shiri Dori-
Hacohen, Daniel Hershcovich, and others for help-
ful discussions throughout the project. For compute
support, the project used the Microsoft Azure Ac-
celerating Foundation Model Research Grant. This
work was supported in part by the Engineering and
Physical Sciences Research Council [grant number
EP/X028909/1].
References
José Alemán and Dwayne Woods. 2016. Value Orien-
tations From the World Values Survey: How Com-
parable Are They Cross-Nationally? Comparative
Political Studies, 49(8):1039–1067.
Badr AlKhamissi,

ork was supported in part by the Engineering and
Physical Sciences Research Council [grant number
EP/X028909/1].
References
José Alemán and Dwayne Woods. 2016. Value Orien-
tations From the World Values Survey: How Com-
parable Are They Cross-Nationally? Comparative
Political Studies, 49(8):1039–1067.
Badr AlKhamissi, Muhammad ElNokrashy, Mai
Alkhamissi, and Mona Diab. 2024. Investigating
cultural alignment of large language models. In Pro-
ceedings of the 62nd Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 1: Long
Papers), pages 12404–12422, Bangkok, Thailand.
Association for Computational Linguistics.
Arnav Arora, Lucie-aimée Kaffee, and Isabelle Augen-
stein. 2023. Probing pre-trained language models for
cross-cultural differences in values. In Proceedings
of the First Workshop on Cross-Cultural Considera-
tions in NLP (C3NLP), pages 114–130, Dubrovnik,
Croatia. Association for Computational Linguistics.
Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda
Askell, Anna Chen, Nova DasSarma, Dawn Drain,
Stanislav Fort, Deep Ganguli, and 22 others. 2022.

-- 9 of 12 --

Training a Helpful and Harmless Assistant with Re-
inforcement Learning from Human Feedback.
Abeba Birhane. 2020. Algorithmic colonization of
Africa. SCRIPTed, 17:389.
Terra Blevins and Luke Zettlemoyer. 2022. Language
contamination helps explains the cross-lingual capa-
bilities of english pretrained models. In Proceedings
of the 2022 Conference on Empirical Methods in Nat-
ural Language Processing, pages 3563–3574, Abu
Dhabi, United Arab Emirates. Association for Com-
putational Linguistics.
Yong Cao, Li Zhou, Seolhwa Lee, Laura Cabello, Min
Chen, and Daniel Hershcovich. 2023. Assessing
cross-cultural alignment between ChatGPT and hu-
man societies: An empirical study. In Proceedings of
the First Workshop on Cross-Cultural Considerations
in NLP (C3NLP), pages 53–67, Dubrovnik, Croatia.
Association for Computational Linguistics.
Eric P. Charles. 2005. The correction for

el Hershcovich. 2023. Assessing
cross-cultural alignment between ChatGPT and hu-
man societies: An empirical study. In Proceedings of
the First Workshop on Cross-Cultural Considerations
in NLP (C3NLP), pages 53–67, Dubrovnik, Croatia.
Association for Computational Linguistics.
Eric P. Charles. 2005. The correction for attenuation
due to measurement error: Clarifying concepts and
creating confidence sets. Psychological Methods,
10(2):206–226.
Catherine D’ignazio and Lauren F. Klein. 2023. Data
Feminism. MIT press.
EVS/WVS. 2022. Joint EVS/WVS 2017-2022 Dataset.
William Fedus, Jeff Dean, and Barret Zoph. 2022. A
review of sparse expert models in deep learning.
Deep Ganguli, Danny Hernandez, Liane Lovitt,
Amanda Askell, Yuntao Bai, Anna Chen, Tom Con-
erly, Nova Dassarma, Dawn Drain, and Nelson El-
hage. 2022. Predictability and surprise in large gener-
ative models. In 2022 ACM Conference on Fairness,
Accountability, and Transparency, pages 1747–1764.
Team Gemma, Morgane Riviere, Shreya Pathak,
Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhu-
patiraju, Bobak Shahriari, Alexandre Ramé, Johan
Ferret, and 187 others. 2024. Gemma 2: Improving
open language models at a practical size.
Luke Guerdan, Solon Barocas, Kenneth Holstein, Hanna
Wallach, Zhiwei Steven Wu, and Alexandra Choulde-
chova. 2025. Validating LLM-as-a-judge systems in
the absence of gold labels.
Shreya Havaldar, Bhumika Singhal, Sunny Rai,
Langchen Liu, Sharath Chandra Guntuku, and Lyle
Ungar. 2023. Multilingual language models are not
multicultural: A case study in emotion. In Proceed-
ings of the 13th Workshop on Computational Ap-
proaches to Subjectivity, Sentiment, & Social Media
Analysis, pages 202–214, Toronto, Canada. Associa-
tion for Computational Linguistics.
Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch,
Elena Buchatskaya, Trevor Cai, Eliza Rutherford,
Diego de Las Casas, Lisa Anne Hendricks, Johannes
Welbl, and 13 others. 2022. An empirical analysis
of compute-optimal large language model

, pages 202–214, Toronto, Canada. Associa-
tion for Computational Linguistics.
Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch,
Elena Buchatskaya, Trevor Cai, Eliza Rutherford,
Diego de Las Casas, Lisa Anne Hendricks, Johannes
Welbl, and 13 others. 2022. An empirical analysis
of compute-optimal large language model training.
Advances in Neural Information Processing Systems,
35:30016–30030.
Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men-
sch, Chris Bamford, Devendra Singh Chaplot, Diego
de las Casas, Florian Bressand, Gianna Lengyel, Guil-
laume Lample, and 9 others. 2023. Mistral 7B.
Rebecca L. Johnson, Giada Pistilli, Natalia Menédez-
González, Leslye Denisse Dias Duran, Enrico Panai,
Julija Kalpokiene, and Donald Jay Bertulfo. 2022.
The Ghost in the Machine has an American accent:
Value conflict in GPT-3.
Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein.
2021. Noise: A Flaw in Human Judgment. Little,
Brown.
Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B.
Brown, Benjamin Chess, Rewon Child, Scott Gray,
Alec Radford, Jeffrey Wu, and Dario Amodei.
2020. Scaling Laws for Neural Language Models.
arXiv:2001.08361 [cs, stat].
Ariba Khan, Stephen Casper, and Dylan Hadfield-
Menell. 2025. Randomness, not representation:
The unreliability of evaluating cultural alignment in
LLMs. In Proceedings of the 2025 ACM Confer-
ence on Fairness, Accountability, and Transparency,
FAccT ’25, pages 2151–2165, New York, NY, USA.
Association for Computing Machinery.
Sal Khan. 2023. Harnessing GPT-4 so that all students
benefit. A nonprofit approach for equal access!
Khyati Khandelwal, Manuel Tonneau, Andrew M. Bean,
Hannah Rose Kirk, and Scott A. Hale. 2024. Indian-
BhED: A dataset for measuring india-centric biases
in large language models. In Proceedings of the 2024
International Conference on Information Technology
for Social Good, pages 231–239, Bremen Germany.
ACM.
Hannah Rose Kirk, Andrew M. Bean, Bertie Vidgen,
Paul Röttger, and Scott A. Hale. 2023. The

Hale. 2024. Indian-
BhED: A dataset for measuring india-centric biases
in large language models. In Proceedings of the 2024
International Conference on Information Technology
for Social Good, pages 231–239, Bremen Germany.
ACM.
Hannah Rose Kirk, Andrew M. Bean, Bertie Vidgen,
Paul Röttger, and Scott A. Hale. 2023. The past,
present and better future of feedback learning in large
language models for subjective human preferences
and values. In Proceedings of the 2023 Conference
on Empirical Methods in Natural Language Process-
ing, pages 2409–2430, Singapore. Association for
Computational Linguistics.
Hannah Rose Kirk, Bertie Vidgen, Paul Röttger, and
Scott A. Hale. 2024a. The benefits, risks and bounds
of personalizing the alignment of large language
models to individuals. Nature Machine Intelligence,
6(4):383–392.
Hannah Rose Kirk, Alexander Whitefield, Paul Röttger,
Andrew Michael Bean, Katerina Margatina, Rafael
Mosquera, Juan Manuel Ciro, Max Bartolo, Adina
Williams, and 3 others. 2024b. The PRISM align-
ment dataset: What participatory, representative and
individualised human feedback reveals about the sub-
jective and multicultural alignment of large language

-- 10 of 12 --

models. In The Thirty-eight Conference on Neural
Information Processing Systems Datasets and Bench-
marks Track.
Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab,
Daan van Esch, Nasanbayar Ulzii-Orshikh, Allahsera
Tapo, Nishant Subramani, Artem Sokolov, and 43
others. 2022. Quality at a glance: An audit of web-
crawled multilingual datasets. Transactions of the
Association for Computational Linguistics, 10:50–72.
Sandipan Kundu, Yuntao Bai, Saurav Kadavath,
Amanda Askell, Andrew Callahan, Anna Chen, Anna
Goldie, Avital Balwit, Azalia Mirhoseini, and 27 oth-
ers. 2023. Specific versus General Principles for
Constitutional AI.
Steven G. Luke. 2017. Evaluating significance in lin-
ear mixed-effects models in R. Behavior Research
Methods, 49(4):1494–1502.
Reem Masoud, Ziquan Liu, Martin

l, Andrew Callahan, Anna Chen, Anna
Goldie, Avital Balwit, Azalia Mirhoseini, and 27 oth-
ers. 2023. Specific versus General Principles for
Constitutional AI.
Steven G. Luke. 2017. Evaluating significance in lin-
ear mixed-effects models in R. Behavior Research
Methods, 49(4):1494–1502.
Reem Masoud, Ziquan Liu, Martin Ferianc, Philip C.
Treleaven, and Miguel Rodrigues Rodrigues. 2025.
Cultural alignment in large language models: An ex-
planatory analysis based on hofstede’s cultural dimen-
sions. In Proceedings of the 31st International Con-
ference on Computational Linguistics, pages 8474–
8503, Abu Dhabi, UAE. Association for Computa-
tional Linguistics.
Stefania Milan and Emiliano Treré. 2019. Big data from
the south(s): Beyond data universalism. Television &
New Media, 20(4):319–335.
Dan Milmo. 2023. ChatGPT reaches 100 million users
two months after launch. The Guardian.
Jakob Mökander, Jonas Schuett, Hannah Rose Kirk,
and Luciano Floridi. 2024. Auditing large language
models: A three-layered approach. AI and Ethics,
4(4):1085–1115.
Shangrui Nie, Michael Fromm, Charles Welch, Rebekka
Görge, Akbar Karimi, Joan Plepi, Nazia Mowmita,
Nicolas Flores-Herr, Mehdi Ali, and Lucie Flek.
2024. Do multilingual large language models mit-
igate stereotype bias? In Proceedings of the 2nd
Workshop on Cross-cultural Considerations in NLP,
pages 65–83, Bangkok, Thailand. Association for
Computational Linguistics.
Dan Nielsen. 2023. ScandEval: A benchmark for scan-
dinavian natural language processing. In Proceed-
ings of the 24th Nordic Conference on Computational
Linguistics (NoDaLiDa), pages 185–201, Tórshavn,
Faroe Islands. University of Tartu Library.
Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groen-
eveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling
Gu, Shengyi Huang, and 31 others. 2025. 2 OLMo 2
furious.
OpenAI. 2022. ChatGPT: Optimizing Language Mod-
els for Dialogue.
OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal,
Lama Ahmad, Ilge Akkaya, Florencia Leoni

Team OLMo, Pete Walsh, Luca Soldaini, Dirk Groen-
eveld, Kyle Lo, Shane Arora, Akshita Bhagia, Yuling
Gu, Shengyi Huang, and 31 others. 2025. 2 OLMo 2
furious.
OpenAI. 2022. ChatGPT: Optimizing Language Mod-
els for Dialogue.
OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal,
Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman,
Diogo Almeida, Janko Altenschmidt, and 272 others.
2024a. GPT-4 technical report.
OpenAI, Aaron Hurst, Adam Lerer, Adam P. Goucher,
Adam Perelman, Aditya Ramesh, Aidan Clark, A. J.
Ostrow, Akila Welihinda, and 410 others. 2024b.
GPT-4o system card.
OpenAI, Ananya Kumar, Jiahui Yu, John Hallman, and
Michelle Pokrass. 2025. Introducing GPT-4.1.
Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert
Demirer. 2023. The impact of AI on developer pro-
ductivity: Evidence from GitHub copilot.
Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo-
pher D. Manning, Stefano Ermon, and Chelsea Finn.
2023. Direct preference optimization: Your lan-
guage model is secretly a reward model. Advances in
Neural Information Processing Systems, 36:53728–
53741.
Maribeth Rauh, John Mellor, Jonathan Uesato, Po-Sen
Huang, Johannes Welbl, Laura Weidinger, Sumanth
Dathathri, Amelia Glaese, Geoffrey Irving, and 3
others. 2022. Characteristics of harmful text: To-
wards rigorous benchmarking of language models.
In Proceedings of the 36th International Conference
on Neural Information Processing Systems, NIPS ’22,
pages 24720–24739, Red Hook, NY, USA. Curran
Associates Inc.
Paul Röttger, Musashi Hinck, Valentin Hofmann, Kobi
Hackenburg, Valentina Pyatkin, Faeze Brahman, and
Dirk Hovy. 2025. IssueBench: Millions of realistic
prompts for measuring issue bias in LLM writing
assistance.
Paul Röttger, Valentin Hofmann, Valentina Pyatkin,
Musashi Hinck, Hannah Kirk, Hinrich Schuetze, and
Dirk Hovy. 2024. Political compass or spinning
arrow? Towards more meaningful evaluations for
values and opinions in large language models. In
Proceedings of the 62nd Annual Meeting of the

as in LLM writing
assistance.
Paul Röttger, Valentin Hofmann, Valentina Pyatkin,
Musashi Hinck, Hannah Kirk, Hinrich Schuetze, and
Dirk Hovy. 2024. Political compass or spinning
arrow? Towards more meaningful evaluations for
values and opinions in large language models. In
Proceedings of the 62nd Annual Meeting of the As-
sociation for Computational Linguistics (Volume 1:
Long Papers), pages 15295–15311, Bangkok, Thai-
land. Association for Computational Linguistics.
David Rozado. 2024. The political preferences of
LLMs. PLOS One, 19(7):e0306621.
Sarah Schröder, Thekla Morgenroth, Ulrike Kuhl, Va-
lerie Vaquet, and Benjamin Paaßen. 2025. Large
language models do not simulate human psychology.
Murray Shanahan. 2024. Talking about large language
models. Commun. ACM, 67(2):68–79.
Mrinank Sharma, Meg Tong, Tomasz Korbak, David Du-
venaud, Amanda Askell, Samuel R. Bowman, Esin
Durmus, Zac Hatfield-Dodds, Scott R. Johnston, and
9 others. 2023. Towards understanding sycophancy
in language models. In The Twelfth International
Conference on Learning Representations.

-- 11 of 12 --

Taylor Sorensen, Jared Moore, Jillian Fisher,
Mitchell L. Gordon, Niloofar Mireshghallah,
Christopher Michael Rytting, Andre Ye, Liwei Jiang,
Ximing Lu, and 3 others. 2024. Position: A roadmap
to pluralistic alignment. In Proceedings of the 41st
International Conference on Machine Learning,
pages 46280–46302. PMLR.
Jinzhe Tan, Hannes Westermann, and Karim
Benyekhlef. 2023. Chatgpt as an artificial
lawyer? In Ai4aj@ Icail.
Yan Tao, Olga Viberg, Ryan S Baker, and René F Kizil-
cec. 2024. Cultural bias and cultural alignment of
large language models. PNAS Nexus, 3(9):pgae346.
Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier
Martinet, Marie-Anne Lachaux, Timothée Lacroix,
Baptiste Rozière, Naman Goyal, Eric Hambro, and 5
others. 2023. LLaMA: Open and Efficient Founda-
tion Language Models.
Ahmet Üstün, Viraat Aryabumi, Zheng Yong, Wei-Yin
Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhan-
dari,

vron, Thibaut Lavril, Gautier Izacard, Xavier
Martinet, Marie-Anne Lachaux, Timothée Lacroix,
Baptiste Rozière, Naman Goyal, Eric Hambro, and 5
others. 2023. LLaMA: Open and Efficient Founda-
tion Language Models.
Ahmet Üstün, Viraat Aryabumi, Zheng Yong, Wei-Yin
Ko, Daniel D’souza, Gbemileke Onilude, Neel Bhan-
dari, Shivalika Singh, Hui-Lee Ooi, and 8 others.
2024. Aya model: An instruction finetuned open-
access multilingual language model. In Proceedings
of the 62nd Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers),
pages 15894–15939, Bangkok, Thailand. Association
for Computational Linguistics.
Laura Weidinger, Jonathan Uesato, Maribeth Rauh,
Conor Griffin, Po-Sen Huang, John Mellor, Amelia
Glaese, Myra Cheng, Borja Balle, and 14 others.
2022. Taxonomy of Risks posed by Language Mod-
els. In 2022 ACM Conference on Fairness, Account-
ability, and Transparency, pages 214–229, Seoul Re-
public of Korea. ACM.
Dustin Wright, Arnav Arora, Nadav Borenstein, Sr-
ishti Yadav, Serge Belongie, and Isabelle Augenstein.
2024. LLM tropes: Revealing fine-grained values
and opinions in large language models. In Find-
ings of the Association for Computational Linguistics:
EMNLP 2024, pages 17085–17112, Miami, Florida,
USA. Association for Computational Linguistics.
Jianhao Yan, Pingchuan Yan, Yulong Chen, Judy Li,
Xianchao Zhu, and Yue Zhang. 2024. GPT-4 vs.
Human translators: A comprehensive evaluation of
translation quality across languages, domains, and
expertise levels. Corr.
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Tianle
Li, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang,
Zhuohan Li, Zi Lin, and 4 others. 2023a. LMSYS-
chat-1M: A large-scale real-world LLM conversation
dataset. In The Twelfth International Conference on
Learning Representations.
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan
Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin,
Zhuohan Li, Dacheng Li, and 4 others. 2023b. Judg-
ing LLM-as-a-judge with MT-bench and

LMSYS-
chat-1M: A large-scale real-world LLM conversation
dataset. In The Twelfth International Conference on
Learning Representations.
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan
Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin,
Zhuohan Li, Dacheng Li, and 4 others. 2023b. Judg-
ing LLM-as-a-judge with MT-bench and chatbot
arena. Advances in Neural Information Processing
Systems, 36:46595–46623.
Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao
Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu,
and 6 others. 2023. LIMA: Less is more for align-
ment. In Thirty-Seventh Conference on Neural Infor-
mation Processing Systems.

-- 12 of 12 --