Anthropogenic Regional Adaptation in Multimodal Vision-Language Model

Anthropogenic Regional Adaptation in
Multimodal Vision-Language Model
Samuel CahyawijayaQ1,2, Peerat LimkonchotiwatQ3,2, Tack Hwa WongQ4,2,
Hitesh Laxmichand PatelQ5, Amit AgarwalQ5, Manuel Antonio RufinoQ6,
Carlos Rafael CatalanQ6, Muhammad Reza QoribR7, Vicky FelirenR8,
Holy LoveniaR2, Aye Hninn KhineR9,2, Frederikus HudiR10,2,
David AnugrahaR11, Alham Fikri AjiR12,2, Romrawin ChumpuR13,
Viet-Thanh PhamN14, Minghan WangN14, Mohamed Fazli ImamN12,
Ruochen ZhangN15,2, Joseph Marvin ImperialN16,31, Khumaisa Nur’ainiN8,
Do Xuan LongN13, Musa Izzanardi WijanarkoN8, Joel Ruben Antony
MonizN17, Patrick Amadeus IrawanN12, Hanif Muhammad ZhafranN18,
Isaiah FloresN19, Salsabila Zahirah PranidaN12, Jun KevinN20,
Jostin Jerico RosalN21, Patricia Nicole MonderinN6, Kun Kerdthaisong22,
Ahmad Mustafid23, My Chiffon Nguyen2, Natchapon Jongwiriyanurak24,
Siva Worajitwannakul16, Haochen Li25, Adrian Xuan Wei Lim13,
Bin Wang26, Muhammad Ravi Shulthan Habibi27,2, Lynnette Hui Xian Ng7,
Mithil Bangera28, Yeshil Bangera28, Priyaranjan Pattnayak23, Dun Li Chan29,
Sherissa Caren Djuniwar30, Cho Chan Myei Oo32, and Hee Ming Shan12
1Cohere 2SEACrowd 3AI Singapore 4Universiti Teknologi PETRONAS
5Oracle 6Samsung R&D Institute Philippines 7Carnegie Mellon University
8Monash University, Indonesia 9King Mongkut’s University of Technology
Thonburi 10Nara Institute of Science and Technology 11Stanford University
12MBZUAI 13National University of Singapore 14Monash University, Australia
15Brown University 16University of Bath 17Mila - Quebec AI Institute
18Institut Teknologi Bandung 19Ateneo de Manila University 20Universitas Pelita
Harapan 21Seoul National University of Science and Technology 22Thammasat
University 23Independent 24University College London 25Nanyang
Technological University 26MiroMind AI 27University of Indonesia 28University
of New Haven 29INTI International University and Colleges 30Binus University
31National University Philippines 32ThoughtFull
Abstract. While the field of

d Technology 22Thammasat
University 23Independent 24University College London 25Nanyang
Technological University 26MiroMind AI 27University of Indonesia 28University
of New Haven 29INTI International University and Colleges 30Binus University
31National University Philippines 32ThoughtFull
Abstract. While the field of vision-language (VL) has achieved remark-
able success in integrating visual and textual information across multiple
languages and domains, there is still no dedicated framework for assessing
human-centric alignment in vision-language systems. We offer two con-
tributions to address this gap. First, we introduce Anthropogenic Re-
gional Adaptation: a novel paradigm that aims to optimize model rele-
vance to specific regional contexts while ensuring the retention of global
generalization capabilities. Second, we present a simple, but effective
Q Main Contributors, R Major Contributors, N Notable Contributors
We release all artifacts generated within our work including training corpora, eval-
uation datasets, and models at https://huggingface.co/collections/SEACrowd/
sea-vl-phase-2-multimodal-vision-language-models-for-sea.
arXiv:2604.11490v2 [cs.AI] 16 Apr 2026

-- 1 of 35 --

2 Cahyawijaya et al.
adaptation method named Geographical-generalization-made-easy (GG-
EZ), which utilizes regional data filtering and model merging. Through
comprehensive experiments on 3 VL architectures: large vision-language
models, text-to-image diffusion models, and vision-language embedding
models, and a case study in Southeast Asia (SEA) regional adaptation,
we demonstrate the importance of Anthropogenic Regional Adaptation
and the effectiveness of GG-EZ, showing 5-15% gains in cultural rel-
evance metrics across SEA while maintaining over 98% of global per-
formance and even occasionally surpassing it. Our findings establish An-
thropogenic Regional Alignment as a foundational paradigm towards ap-
plicability of multimodal vision-language models in diverse regions

EZ, showing 5-15% gains in cultural rel-
evance metrics across SEA while maintaining over 98% of global per-
formance and even occasionally surpassing it. Our findings establish An-
thropogenic Regional Alignment as a foundational paradigm towards ap-
plicability of multimodal vision-language models in diverse regions and
demonstrate a simple-yet-effective baseline method that optimizes re-
gional value alignment while preserving global generalization.
1 Introduction
Representation alignment in underrepresented regional contexts of foundation
models has been a longstanding problem in AI [6,29,62], limiting their effective-
ness and applicability across diverse global populations. This challenge is partic-
ularly acute in under-developed and developing regions – such as African [1, 57],
Indian [32, 55], Middle Eastern, Southeast Asian (SEA) [12, 15, 38, 60], etc –
where existing resources exhibit severe underrepresentation. Building on this
critical gap, we observe that existing vision-language (VL) AI solutions, despite
their capabilities to generalize across different languages and domains, often ex-
hibit cultural insensitivity [11,42,48], stereotypical outputs [16,36], and reduced
task performance when deployed in underrepresented regions [3, 17, 31]. This
misalignment stems from training data dominated by certain regions with lim-
ited exposure to diverse contextual nuances prevalent in other regions [2, 10, 13],
undermining their real-world applicability.
To bridge this divide, we propose Anthropogenic Regional Adaptation, a
foundational paradigm designed to systematically evaluate human-centric align-
ment of VL models across regional contexts while preserving global generaliza-
tion capabilities. Within this framework, we identify two primary model archetypes:
(1) global models that achieve strong performance across broad geographic con-
texts but struggle with underrepresented regions, and (2) regionally specialized
models that excel on locale-specific

ts while preserving global generaliza-
tion capabilities. Within this framework, we identify two primary model archetypes:
(1) global models that achieve strong performance across broad geographic con-
texts but struggle with underrepresented regions, and (2) regionally specialized
models that excel on locale-specific metrics, yet falter when confronted with
broader global contexts. Neither archetype is ideal, creating a critical gap in the
utilization of existing systems for targeted regional applications.
Overcoming this limitation, we specifically designed a simple yet effective
method, Geographical-generalization-made-easy (GG-EZ), which adapts an ex-
isting global model to a regional-specific context with minimal degradation on
the global context. Inspired by recent advancements in training strategies of
large language models (LLMs) [18, 23, 33, 67], GG-EZ operationalizes regional
adaptation through a two-level approach: (1) regional data filtering to curate
culturally relevant training subsets, and (2) model merging to integrate region-
specific adaptations without catastrophic forgetting of global knowledge.

-- 2 of 35 --

Anthropogenic Regional Adaptation 3
We validate Anthropogenic Regional Adaptation and GG-EZ through rig-
orous experiments across three VL architectures—large vision-language models
(Gemma-3 27B), text-to-image diffusion models (SDXL), and vision-language
embedding models (SigLIP-2) – focusing on SEA as a case study. Our results
demonstrate that GG-EZ achieves 5-15% improvements in cultural relevance
metrics, including cultural context accuracy and local visual context understand-
ing, while retaining over 98% of global performance on standard benchmarks.
These findings establish Anthropogenic Regional Adaptation as a critical step
toward equitable vision-language solutions and position GG-EZ as a practical
baseline for anthropogenic regional adaptation in vision-language systems.
2 Related Work
Regional-specific models have been reported

rmance on standard benchmarks.
These findings establish Anthropogenic Regional Adaptation as a critical step
toward equitable vision-language solutions and position GG-EZ as a practical
baseline for anthropogenic regional adaptation in vision-language systems.
2 Related Work
Regional-specific models have been reported to outperform one-size-fits-all mod-
els on regional context [45]. Several datasets [15,38] and language models [43,44]
have been developed for the Southeast Asia (SEA) region. However, in computer
vision, work on regional adaptation has focused more on remote sensing [47,56],
which results in limited understanding of a region’s unique cultural and anthro-
pomorphic characteristics. Cahyawijaya et al. [15] report that existing multi-
lingual vision-language models such as MAYA-8B [5], PaliGemma-2-10B [50],
Pangea-7B [66], Qwen2-VL-7B [58], Gemma-3 [52] fail in generating culturally-
relevant responses when benchmarked against regional-specific cultural bench-
marks such as SEAVQA [54], CVQA [40] and World Cuisines [61]. On the other
hand, recent development of regional-specific vision-language models such as VI-
OLET [41] and Baseer [27], VARCO-VISION [30], and SEA-LION-VL [49], have
strong performance on regional-specific benchmark, but fall short on a broader
and more general global context. Motivated by these limitations, we present
a generic framework to systematically assess the gap of anthropogenic regional
alignment in different models and introduce a simple-yet-effective approach, GG-
EZ, to reduce the regional gap of existing multi-modal vision-language models,
while maintaining its performance on the broader global contexts. 1
3 Anthropogenic Regional Adaptation
3.1 Overview
Let Rglobal denote the comprehensive spatial domain representing our global
context of interest. Within this universal space, we establish a discrete parti-
tioning into meaningful geographical or conceptual units through the definition
of R, a finite collection of k

1
3 Anthropogenic Regional Adaptation
3.1 Overview
Let Rglobal denote the comprehensive spatial domain representing our global
context of interest. Within this universal space, we establish a discrete parti-
tioning into meaningful geographical or conceptual units through the definition
of R, a finite collection of k distinct regions. Formally, R = {R1, R2, . . . , Rk}
where each individual region Ri constitutes a measurable subset of Rglobal. This
decomposition enables systematic analysis of adaptation strategies across het-
erogeneous spatial contexts, recognizing that the world is rarely homogeneous in
its characteristics or response patterns to anthropogenic influences.
1 We provide comparison of our SEA-VLM models with other models in Appendix D.

-- 3 of 35 --

4 Cahyawijaya et al.
Well-aligned
Low performance
Global models:
Broad coverage, 	gaps in underrepresented regions
Regionally specialized models:
Local excellence, 	failure in broader global contexts
Well-aligned
Low performance
Fig. 1: Through anthropogenic regional adaptation, we identify two primary model
archetypes: (left) Global model with strong overall global performance, but struggle
to represent certain regions appropriately, and (right) Regional-specific model that has
a strong representation towards certain regions, but fall short on the global context.
Building upon this foundational regional partitioning, we introduce a critical
distinction between regions based on strategic importance. The target region
Rregional is defined as a specific, non-empty subset of Rglobal representing our
primary focus region where adaptation efforts should be maximized. Comple-
mentarily, the remaining regions are collectively designated as Rothers, defined
through set difference as Rothers = Rglobal \ Rreg. This ensures complete domain
coverage while maintaining clear analytical separation between priority and non-
priority regions, with each point in Rglobal belonging to exactly one region.
Building upon

remaining regions are collectively designated as Rothers, defined
through set difference as Rothers = Rglobal \ Rreg. This ensures complete domain
coverage while maintaining clear analytical separation between priority and non-
priority regions, with each point in Rglobal belonging to exactly one region.
Building upon this theoretical framework, as depicted in Figure 1, exist-
ing vision-language models predominantly follow two archetypal approaches.
Global models capture broad, universal patterns across Rglobal but sacrifice nu-
anced regional specificity. Conversely, regional-specific models focus on particu-
lar Rregional subsets with high local performance but reduced global coherence.
This dichotomy reflects a fundamental trade-off in model design between uni-
versal applicability and specialized excellence. Most existing models, commit to
one of the paradigm: global optimization for comprehensive coverage or regional
specialization for enhanced regional performance.
3.2 Global-Regional Parity (GRP) Optimization
For each region Ri, we have a set of evaluation metrics ERi = {eRi
1 , eRi
2 , . . . , eRi
m }.
We define a quality metric qRi
j for each evaluation metric eRi
j , and the collection
of these quality metrics forms the set QRi = {qRi
1 , qRi
2 , . . . , qRi
m }. This multi-
dimensional approach captures the complexity of regional characteristics influ-
encing adaptation outcomes. The evaluation framework extends to our parti-
tioned structure, yielding aggregated evaluation sets QRglobal
and QRregional
that
consolidate metrics across regions while maintaining connections to individual
characteristics. To balance competing objectives, we introduce a globalization
factor α ∈ [0, 1] that explicitly controls the trade-off between optimizing the
target region and maintaining global generalization. The formal optimization

-- 4 of 35 --

Anthropogenic Regional Adaptation 5
Translate
Global VL data
in English
Region-specific
cultural VL data
Global VL data

introduce a globalization
factor α ∈ [0, 1] that explicitly controls the trade-off between optimizing the
target region and maintaining global generalization. The formal optimization

-- 4 of 35 --

Anthropogenic Regional Adaptation 5
Translate
Global VL data
in English
Region-specific
cultural VL data
Global VL data in
regional languages
Worldwide
cultural VL data
Regional filter
Quality filter
Region-specific
SFT VL corpora
+
Supervised
fine-tuning
GG-EZ
Merge
Global model
Regional model
Anthropogenic
regional adapted
model
Phase 2: Global-Regional Refinement	Phase 1: Regional Quality Filtering
Fig. 2: Overview of our Geographical-generalization-made-easy (GG-EZ) framework.
Our framework consist of 3 constituents: (1) High-quality regional data filtering
pipeline; (2) supervised fine-tuning to create a high quality regional-specific model;
and (3) model merging to capture the best combination between regional-specific and
global represention while also maintaining the generalization capabilities of the model.
objective is constructed as a weighted combination:
max
θ
h
α · QRglobal
+ (1 − α) · QRregional i
This formulation creates a single scalar objective that can be optimized using
standard techniques, with α serves as a critical control mechanism that can
be adjusted based on the specific focus and characteristics of each region. For
regions with strong global connections or dependencies, α can be set closer to
1 to emphasize maintenance of Rglobal conditions, while for regions where local
impacts dominate, α can approach 0 to prioritize Rregional optimization. The
solution achieves "best parity" between regions, representing the most favorable
equilibrium where neither aspect is disproportionately favored according to our
specified weighting scheme.
4 Geographical Generalization Made Easy (GG-EZ)
Geographical-generalization-made-easy (GG-EZ) is a general framework for re-
gional adaptation of global multimodal models that achieves effective region-
specific

equilibrium where neither aspect is disproportionately favored according to our
specified weighting scheme.
4 Geographical Generalization Made Easy (GG-EZ)
Geographical-generalization-made-easy (GG-EZ) is a general framework for re-
gional adaptation of global multimodal models that achieves effective region-
specific performance while preserving original generalization capabilities. As il-
lustrated in Figure 2, the core mechanism of GG-EZ consists of two phases:
regional quality filtering and global-regional refinement. GG-EZ not only
enables adaptation across local textual and imagery contexts while maintaining
base model capabilities, but also better yet being architecture-agnostic, allowing
effective regional adaptation across different vision-language architectures.
4.1 Regional Quality Filtering
The regional quality filtering phase refines the training dataset to focus on region-
ally relevant examples while preserving general knowledge. Let D = {Dreg, Dgen}

-- 5 of 35 --

6 Cahyawijaya et al.
represent the complete dataset, where Dreg contains region-specific examples and
Dgen contains general-domain examples.
Regional Filter (Frf): We define a boolean regional filter function that
selects examples from the target region:
Frf(x, r) =
(
1 if region(x) = r
0 otherwise
where r denotes the target region and region(x) returns the geographical region
associated with example x.
Multilingual Reward (Frm): To further refine data filtering, we employ
reward models that scores data quality and relevance on a continuous scale:
Frm(x) = θreward(x)
where θreward is the weight of the reward model and θreward(x) measure the
quality of the data x using the reward model θreward. The filtered regional-
specific dataset is then obtained by applying a threshold τ :
Dfiltered = {x ∈ Dreg | Frf(x, r) = 1 and Frm(x) > τ }
Language Translation Augmentation: Beyond regional filtering of exist-
ing cultural-relevant corpora, a further data enrichment is done through trans-
lating

reward model θreward. The filtered regional-
specific dataset is then obtained by applying a threshold τ :
Dfiltered = {x ∈ Dreg | Frf(x, r) = 1 and Frm(x) > τ }
Language Translation Augmentation: Beyond regional filtering of exist-
ing cultural-relevant corpora, a further data enrichment is done through trans-
lating English high-quality datasets to multiple target regional languages. Let T
represent the translation function mapping English to regional languages Lr :
Dtranslation = {T (x, English → l) | x ∈ Dfiltered, l ∈ Lr }
4.2 Global-Regional Refinement
Using a pre-trained global model and the resulting corpora from the regional
quality filtering phase, two steps of global-regional refinements are incorpo-
rated: 1) supervised fine-tuning, which turns the global model into a highlt
performant region-specific model, and 2) model merging to ensure the preser-
vation of the original global generalization capabilities. We first trained the
global model θglobal into a quality regional-specific model θregional using Dsft ⊆
{Dfiltered ∪ Dtranslation}. After supervised fine-tuning on Dsft, a linear model
merging technique is applied to combine the global model with the region-specific
model. The merged model parameters θmerged are defined as:
θmerged(β) = β · θregional + (1 − β) · θglobal
where β controls the interpolation between the region-specific and the original
global model. For beta interpolation, we first select a certain value for the global-
regional parity factor α (e.g., α = 0.65) to balance region-specific performance
and general capabilities. The optimal β value is selected based on evaluation
metrics:
β∗ = argmax
α∈[0.0,1.0]
h
α · QRregional
θmerged(β) + (1 − α) · QRglobal
θmerged(β)
i

-- 6 of 35 --

Anthropogenic Regional Adaptation 7
where QRregional
θmerged(β) measures regional-specific performance from the evaluation
sets ERregional
and QRglobal
θmerged(β) measures global performance from the evaluation
sets ERglobal
. This approach ensures minimal

l
θmerged(β) + (1 − α) · QRglobal
θmerged(β)
i

-- 6 of 35 --

Anthropogenic Regional Adaptation 7
where QRregional
θmerged(β) measures regional-specific performance from the evaluation
sets ERregional
and QRglobal
θmerged(β) measures global performance from the evaluation
sets ERglobal
. This approach ensures minimal degradation of base model capa-
bilities while achieving strong regional adaptation.
5 Case Study on Southeast Asian (SEA) Adaptation
We explore GG-EZ on 3 distinct multimodal architectures: a large-scale Vision-
Language Model (VLM) with 27B parameters, a contrastive Vision-Language
Embedding Model (VL-Embed) with 1B parameters, and a Contextualized Dif-
fusion Model (SEA-ImageGen) with 3.5B parameters. Our experiments focus
on the SEA region, encompassing 11 countries—Singapore, Indonesia, Malaysia,
Brunei, Thailand, Philippines, Vietnam, Myanmar, Cambodia, Laos, and East
Timor—with a total population of approximately 700 million. In this section, we
provide detailed configurations for applying GG-EZ and the evaluation strategy
used to assess its effectiveness.
5.1 Data Curation
For data curation, we leveraged prefiltered SEA-specific content from SEA-VL,
culturally relevant imagery from CulturalGround [46], and translated instruction
data from MAmmoTH-VL [24] (5 shards converted to major SEA languages,
including Indonesian, Malaysian, Thai, Vietnamese, Filipino, Khmer, Lao, Chi-
nese, and Tamil). To ensure high-quality translation on the underrepresented
languages like Khmer and Lao, we ablate different translation models for each
language and select the one with optimal quality and cost 2. The data filtering
process applied a regional filter to isolate SEA-specific examples, utilized the
UnifiedReward [59] model for quality assessment with a threshold set at 3 or
above 3, and combined regional and translated instruction data to create the
final fine-tuning dataset. This dataset curation is crucial to enable adaptation of
a general global model

onal filter to isolate SEA-specific examples, utilized the
UnifiedReward [59] model for quality assessment with a threshold set at 3 or
above 3, and combined regional and translated instruction data to create the
final fine-tuning dataset. This dataset curation is crucial to enable adaptation of
a general global model into a strong regional-specific model as explained in §6.2.
5.2 Regional Supervised Fine-tuning
All models underwent supervised fine-tuning using the AdamW optimizer [37]
with linear learning rate decay. The SEA-adapted VLM (SEA-VLM) was trained
from the Gemma-3 model as the initial checkpoint with a batch size of 64,
learning rate of 2e-5, weight decay of 0.01, and 3 epochs. Before the supervised
fine-tuning, we perform continuous pre-training on the Gemma-3 checkpoint
using SEA-VL, XM3600, and Flickr30k, where we translated these datasets into
2 See Appendix A for more detail.
3 We pick the best reward model based on our ablations on a human-annotated SEA-
specific human preference dataset. The comparison of different reward model quality
on the SEA regional test set is shown on Appendix B.

-- 7 of 35 --

8 Cahyawijaya et al.
SEA languages similar to Section 5.1. For the contextualized diffusion model
(SEA-ImageGen), the model was fine-tuned from SDXL with a batch size of
32, learning rate of 1e-5, weight decay of 0.01, and 4 epochs. We only fine-
tuned the UNet module the model while keeping the VAE module as is. For the
VL embedding model (SEA-VL Embed), the model was fully fine-tuned from
SigLIP2-SO400m with a batch size of 128, learning rate of 5e-6, weight decay
of 0.001, and 2 epochs. The resulting regionally-adapted models are denoted
as SEA-Gemma-3, SEA-SDXL, and SEA-SigLIP2 for the VLM, diffusion, and
embedding architectures respectively.
5.3 Global-Regional Merging
After fine-tuning, we performed linear model merging to combine region-specific
adaptations with the original base models. For VL-Embed and SEA-ImageGen,
we explored

odels are denoted
as SEA-Gemma-3, SEA-SDXL, and SEA-SigLIP2 for the VLM, diffusion, and
embedding architectures respectively.
5.3 Global-Regional Merging
After fine-tuning, we performed linear model merging to combine region-specific
adaptations with the original base models. For VL-Embed and SEA-ImageGen,
we explored interpolation weights β in the range [0.25, 0.5, 0.75], while for the
SEA-VLM we tested a broader range of β values [0.05, 0.10, 0.5, 0.7]. The optimal
β was determined by maximizing a weighted combination of regional and general
performance metrics, where RegionAcc measures performance on SEA-specific
validation data and GeneralAcc evaluates performance on broad-domain bench-
marks. This formulation prioritizes regional adaptation while ensuring minimal
degradation of foundational capabilities. We denote the resulting merged model
as “<MODEL> (X%)” – e.g., SEA-Gemma-3 10% – where X% denotes the
weight percentage β of the regionally-specific model used in the linear merging.
5.4 Evaluation
For the SEA-VLM model, evaluation included the SEAVQA regional visual ques-
tion answering benchmark, CVQA for culturally nuanced queries, WorldCuisine
for food identification tasks, and human evaluation to assess cultural relevance
and accuracy. 4. The SEA-VL Embed was evaluated on SEAVQA and CVQA
to measure regional and cultural performance in embedding space. The SEA-
ImageGen model underwent assessment using a DPGBench benchmark [28] as a
proxy for the global quality of the models and diffusion model human evaluation
set from SEA-VL [16] to gauge cultural relevance and visual quality on the SEA
region. All evaluations were conducted on held-out test sets to ensure fair com-
parison across architectures and adaptation strategies. For human evaluation we
use a 3-level likert scoring with native speaker annotators for each of the evalu-
ated regional language. Finally, to compute a fair and practical global-regional
parity (GRP) score from both of the global and

est sets to ensure fair com-
parison across architectures and adaptation strategies. For human evaluation we
use a 3-level likert scoring with native speaker annotators for each of the evalu-
ated regional language. Finally, to compute a fair and practical global-regional
parity (GRP) score from both of the global and region-specific evaluations, we
set the globalization factor α = 0.43 5 derived from 2023 KOF Globalization
Index [25, 26] which measures the level of globalization across 190 countries.
4 Due to limited evaluation resources for the region, we develop two human evaluation
sets: (1) SEA-AYA a human translated AyaVisionBench [19]; and (2) SEA-VL VQA,
a human annotated VQA prompts from SEA imagery. Details in Appendix C.
5 α = 0.43 indicates a moderate tendency toward SEA-focused quality QRregional
with
a fair amount of consideration for global quality QRglobal
.

-- 8 of 35 --

Anthropogenic Regional Adaptation 9
Table 1: SEA-VLM evaluation results on CVQA, SEA-VQA, and WorldCuisine (WC).
<MODEL> (X%) denotes a merged model where X% is the weight β use in the linear
merging process of the region-specific model.
Model GRP Global SEA-Specific
Avg. WC CVQA Avg. SEAVQA WC CVQA
Google Gemma-3 59.4 63.5 59.8 67.2 56.3 41.0 60.1 67.8
SEA-Gemma-3 5% 64.0 64.3 60.0 68.7 63.7 61.2 60.3 69.5
SEA-Gemma-3 10% 64.1 64.4 60.0 68.8 63.8 61.7 60.2 69.5
SEA-Gemma-3 50% 57.3 56.7 51.6 61.8 57.8 59.5 51.4 62.6
SEA-Gemma-3 70% 56.1 56.3 51.9 60.6 56.0 54.0 52.6 61.3
SEA-Gemma-3 (w/o merge) 42.2 42.1 48.5 35.6 42.2 41.9 48.6 36.2
Table 2: SEA-VLM human evaluation on SEA-AYA and SEA-VL VQA datasets. We
report the average rank across different models (Higher is better). <MODEL> (X%)
denotes a linear merge where X% is the weight β of the region-specific model.
Model Overall Language Breakdown
GRP Global SEA fil ind tha vie zsm
Google Gemma-3 2.29 2.54 2.09 1.69 2.15 2.17 2.37 2.07
SEA-Gemma-3 10% 2.31 2.42 2.22 1.88 2.07 2.29 2.61 2.25
SEA-Gemma-3 (w/o merge) 1.74 1.18 2.23

s better). <MODEL> (X%)
denotes a linear merge where X% is the weight β of the region-specific model.
Model Overall Language Breakdown
GRP Global SEA fil ind tha vie zsm
Google Gemma-3 2.29 2.54 2.09 1.69 2.15 2.17 2.37 2.07
SEA-Gemma-3 10% 2.31 2.42 2.22 1.88 2.07 2.29 2.61 2.25
SEA-Gemma-3 (w/o merge) 1.74 1.18 2.23 2.75 2.29 2.33 1.76 2.00
6 Result and Discussion
6.1 Overall Results
GG-EZ on SEA-VLM We observe two major trends in SEA-Gemma-3 per-
formance, as illustrated in Table 1. First, without any model merging, SEA-
Gemma-3 generally performs worse compared to the original Gemma-3 model
as evidenced in both the CVQA and World-Cuisine benchmarks, possibly due to
some loss of task generalization capability after the supervised fine-tuning with
regional data . Nevertheless, SEA-Gemma-3 is still able to outperform the orig-
inal Gemma-3 model when faced with a challenging SEA-specific regional eval-
uation, i.e., SEA-VQA. This signifies the only applying supervised fine-tuning
to perform regional adaptation might not results in an ideal performance as
pre-trained general VLMs are usually already trained on large subset of diverse
high-quality datasets across languages, regions, and tasks, and fine-tuning them
for with smaller regional-specific data might harm their generalization capability.
Interestingly, when model merging is applied into SEA-Gemma-3„ we observe
significant performance improvements across all evaluation scenarios with 50:50
mix of the original Gemma-3 and SEA-Gemma-3 models (SEA-Gemma-3 50%),
we attain a turning point, where the global performance is only slightly lower,
while the SEA-specific performance surpasses the original Gemma-3 model. Push-
ing the merging weight β further to 10% brings further improvement, with the
SEA-Gemma-3 10% outperforms the original Gemma-3 model on all evaluation

-- 9 of 35 --

10 Cahyawijaya et al.
Table 3: SEA-SDXL Performance on Image Generation Benchmark DPGBench
(Higher is better). <MODEL> (X%) denotes a

emma-3 model. Push-
ing the merging weight β further to 10% brings further improvement, with the
SEA-Gemma-3 10% outperforms the original Gemma-3 model on all evaluation

-- 9 of 35 --

10 Cahyawijaya et al.
Table 3: SEA-SDXL Performance on Image Generation Benchmark DPGBench
(Higher is better). <MODEL> (X%) denotes a linear merge where X% is the weight
β of the region-specific model.
Model Overall Aspects
Score Attribute Relation Entity Other Global
StabilityAI SDXL 73.75 79.18 86.38 81.07 60.00 83.89
SEA-SDXL 25% 74.75 80.43 86.34 81.86 61.60 85.41
SEA-SDXL 50% 74.61 80.93 87.12 81.15 60.00 86.02
SEA-SDXL 75% 74.39 79.98 85.96 81.36 63.20 82.67
SEA-SDXL (w/o merge) 74.32 80.79 86.92 81.57 65.20 81.16
sets, improving the averaged global performance by 1% while significantly im-
prove the average SEA-specific performance by 7.5% leading to a much higher
GRP compared the source global and region-specific models. This result un-
derscores the advantage of GG-EZ to enable regional-adapted models that are
strong on both region-specific aspect and generalization capability.
The human evaluation results in Table 2 further validate the automatic eval-
uation trends of GG-EZ. The result reveals that SEA-Gemma-3 (w/o merge)
achieves superior regional specialization, securing the best average rank across
Southeast Asian languages—including Filipino (2.75), Indonesian (2.29), Thai
(2.33) and the broader SEA region (2.23). These findings corroborate our ear-
lier observation that SEA-Gemma-3 excels on region-specific evaluations where
Gemma-3’s global optimization proves to be less effective. Conversely, Google
Gemma-3 maintains dominance in global performance (2.54), confirming that
SEA-Gemma-3’s regional advantages come at the cost of broader generalizability.
The SEA-Gemma-3 10% variant occupies an interesting middle ground, achiev-
ing the best Vietnamese score (2.61) and Malaysian performance (2.25) with
minimal degradation on the global performance (2.42) compared to

(2.54), confirming that
SEA-Gemma-3’s regional advantages come at the cost of broader generalizability.
The SEA-Gemma-3 10% variant occupies an interesting middle ground, achiev-
ing the best Vietnamese score (2.61) and Malaysian performance (2.25) with
minimal degradation on the global performance (2.42) compared to (2.54) on
the original Gemma-3 model yielding highest GRP score of 2.31. This suggests
that even modest SEA integration yields measurable regional benefits. 6.
GG-EZ on SEA-ImageGen As illustrated in Table 3, the resulting SEA-
SDXL models also demonstrate strong regional adaptation while preserving
strong general image generation performance on DPGBench which is used as the
proxy of the global performance. the fully fine-tuned SEA-SDXL model shows a
slight improvement in most aspects, achieving an overall performance of 74.32
compared to the original SDXL with 73.75. Similar with SEA-VLM, the linear
merging incorporated in GG-EZ further improve the image generation capabil-
ity, with the SEA-SDXL 25% and 50% models surpassing the general model and
the regional-specific SEA-SDXL model achieving 74.75 and 74.61 on DPGBench,
respectively, with competitive scores across all aspect categories.
6 We provide more detailed results and samples outputs from different models from
our experiments in Appendix E and Appendix F.

-- 10 of 35 --

Anthropogenic Regional Adaptation 11
Table 4: SEA-SDXL human evaluation results on 3 distinct cultural aspects. “T”
denotes Tradition, “L” denotes Landmark, and “C” denotes Cuisine
Model Correctness Naturalness
Overall T L C Overall T L C
StabilityAI SDXL 1.491 1.470 1.636 1.387 1.675 1.436 2.023 1.613
SEA-SDXL 25% 1.569 1.587 1.729 1.413 1.767 1.473 2.124 1.753
SEA-SDXL (w/o merge) 1.431 1.473 1.527 1.307 1.557 1.340 1.806 1.560
The regional-specific human evaluation results (Table 4) further validate the
effectiveness of the merging strategy for regional context. The SEA-SDXL 25%
model demonstrates superior performance

DXL 25% 1.569 1.587 1.729 1.413 1.767 1.473 2.124 1.753
SEA-SDXL (w/o merge) 1.431 1.473 1.527 1.307 1.557 1.340 1.806 1.560
The regional-specific human evaluation results (Table 4) further validate the
effectiveness of the merging strategy for regional context. The SEA-SDXL 25%
model demonstrates superior performance across all cultural aspects (Tradition,
Landmark, Cuisine) in both correctness and naturalness metrics, outperforming
both the original SDXL and the fully fine-tuned SEA-SDXL This indicates that
the merging process successfully recovers the generalization capability lost during
regional specialization, allowing the resulting SEA-SDXL models to maintain
or enhance their performance on general image generation benchmarks while
simultaneously excelling in regional context evaluations. 7
GG-EZ on SEA-VL Embed Aligned to the result of SEA-ImageGen, the
developed SEA-SigLIP2 demonstrate an interesting performance trend where,
as shown in Table 5, the regional fine-tuned SEA-SigLIP2 model perform bet-
ter on both SEA and non-SEA (other) regions in both CVQA and SEA-VQA
evaluations compared to the original Google SigLIP2 model. Moreover, merging
the SEA-SigLIP2 model back with the original SigLIP2 model brings additional
boost of the evaluation performance. The SEA-SigLIP2 50% model excels in gen-
eral knowledge transfer, achieving the highest global CVQA score (27.52) and
strongest performance on non-SEA regions (27.97), While the SEA-SigLIP2 75%
model demonstrates superior regional specialization with the highest SEAVQA
overall score (29.66) and peak performance for Indonesia (30.05) and Vietnam
(28.75), while at the same time maintaining a high global performance (27.12)
yielding the highest GRP score of 27.96 compared to the original SigLIP2 model
(25.17) and the fine-tuned SEA-SingLIP2 (26.96). 8 The merging approach ef-
fectively balances the trade-off between regional adaptation and generalization,
allowing the models to preserve and enhance the

g a high global performance (27.12)
yielding the highest GRP score of 27.96 compared to the original SigLIP2 model
(25.17) and the fine-tuned SEA-SingLIP2 (26.96). 8 The merging approach ef-
fectively balances the trade-off between regional adaptation and generalization,
allowing the models to preserve and enhance the original SigLIP2 broad capa-
bilities while incorporating specialized regional knowledge from SEA-SigLIP2.
6.2 Impact of Regional Quality Filtering
The quality and relevance of data utilized for region-specific model adaptation
significantly dictate performance outcomes [13, 14]. Our ablation study on SEA-
7 We do not measure the GRP score for image generation as there is no comparable
test sets with the comparable value scale. However, it is clear that the resulting
model improves both regional-specific and general image generation scores.
8 We provide more detailed breakdown of all the evaluations in Appendix E.

-- 11 of 35 --

12 Cahyawijaya et al.
Table 5: SEA-VL Embedding Zero-Shot Performance on Cultural VQA Benchmarks.
<MODEL> (X%) denotes a linear merge where X% is the weight β of the region-
specific model. For SEA-VQA, we provide additional breakdown of 3 of the biggest
country data in the dataset with “Khm”=Cambodia, “Ind”=Indonesia, “Vie”=Vietnam.
Model GRP CVQA SEAVQA
Global SEA Others SEA Khm Ind Vie
Google SigLIP2 25.17 25.51 24.02 25.84 25.81 28.95 25.00 25.56
SEA-SigLIP2 25% 24.96 24.38 24.44 24.36 26.36 27.63 25.40 23.96
SEA-SigLIP2 50% 27.10 27.52 25.50 27.97 28.06 31.25 26.99 27.16
SEA-SigLIP2 75% 27.96 27.12 27.51 27.03 29.66 28.62 30.05 28.75
SEA-SigLIP2 (w/o merge) 26.96 26.75 25.29 27.07 28.96 28.29 28.59 29.07
Gemma-3 demonstrates that naive data augmentation does not guarantee im-
provements; rather, performance gains depend heavily on the specific structure,
scope, and provenance of the integrated data.
As shown in Figure 3, data volume acts as a primary bottleneck. Utilizing
only a 20% subset of the baseline dataset, which

mma-3 demonstrates that naive data augmentation does not guarantee im-
provements; rather, performance gains depend heavily on the specific structure,
scope, and provenance of the integrated data.
As shown in Figure 3, data volume acts as a primary bottleneck. Utilizing
only a 20% subset of the baseline dataset, which consists of the SEA-Mammoth
250k dataset translated into 10 regional languages, leads to a severe 70% degra-
dation in overall model performance relative to the full dataset, emphasizing the
necessity of maintaining adequate training scale. As for subsequent integration of
culture-specific knowledge, i.e., the SEA-filtered Cultural-Ground dataset [46],
reveals that performance is highly contingent on the VQA task format. As il-
lustrated in Figure 3, the open-ended VQA formulation effectively enhances the
model’s regional understanding from the baseline to 41.9%, albeit with high
sensitivity to training prompt design. Conversely, appending the multiple-choice
variant of the same data source actively degrades performance to 21.6%. Simi-
larly, incorporating the SEA-filtered WorldCuisine dataset [62] proves harmful,
resulting in a ∼42% performance reduction compared to the baseline. We hypoth-
esize that this degradation stems from WorldCuisine’s specialization in regional
cuisine, which excessively narrows the model’s representational focus and fails to
provide the comprehensive knowledge required for adaptation to other cultural
aspects, e.g., landmarks, history, traditional clothing, and others.
These findings highlight the critical importance of strategic data curation. In-
discriminately appending datasets without rigorous, format- and domain-aware
evaluation can compromise the model’s capacity for generalizable learning.
6.3 Globalization Factor in Anthropogenic Regional Adaptation
The globalization factor α plays a huge role in the GRP optimization of An-
thropogenic Regional Adaptation. This factor determine how a model should
be optimized to

and domain-aware
evaluation can compromise the model’s capacity for generalizable learning.
6.3 Globalization Factor in Anthropogenic Regional Adaptation
The globalization factor α plays a huge role in the GRP optimization of An-
thropogenic Regional Adaptation. This factor determine how a model should
be optimized to achieve the optimal characteristic of representing a particular
target region. An appropriate globalization factor α would reflect the actual con-
dition on how society within the targeted regions interact with others, whether

-- 12 of 35 --

Anthropogenic Regional Adaptation 13
10.15 	15.15 	20.15 	25.15 	30.15 	35.15 	40.15
(Baseline dataset)
45.15 	50.15
SEA-VQA Accuracy (%)
Only used a subset (20%)
Added Cultural-Ground
Multiple-Choice
Added World Cuisine
Added Cultural-Ground
Open-Ended
Modifications to
SEA-Mammoth 250k
 	Country
ID
KH
LA
MY
PH
SG
TH
VN
Fig. 3: Impact of regional-specific data curation strategy on SEA-Gemma-3.
individuals prefer a single unified global perspective or a more unique and lo-
calized regional perspective. The misalignment of globalization α could result in
undesirable behavior of the adapted model as shown in Figure 4 (left).
To accommodate different preferences across different regions, we derive the
globalization factor α based on globalization index [7, 21, 39] which has been
widely explored in the field of social science to measure the degree of globalization
across different regions or countries. More specifically, we employ one prominent
globalization index known as Konjunkturforschungsstelle Globalization Index
(KOF-GI) [25, 26] . This is done by calculating the expectation of the KOF-GI
across all countries within the target region Rregional. 9. This formulation of the
globalization factor α provides a timely, quantifiable, and standardized measure
of globalization intensity for each region. By doing so, we establish a fair global-
regional parity baseline that reflects how embedded globalization is within

all countries within the target region Rregional. 9. This formulation of the
globalization factor α provides a timely, quantifiable, and standardized measure
of globalization intensity for each region. By doing so, we establish a fair global-
regional parity baseline that reflects how embedded globalization is within that
specific region. This has significant implications: 1) as the globalization index
evolves over time – as shown in Figure 4 (right) –, the GRP value naturally
shifted, dynamically adjusting the optimization balance; and 2) the KOF-GI
derived α provides an objective, external metric that reduces subjectivity in
determining regional importance, making the process more data-driven.
The integration of KOF-GI-based α directly addresses the model archetype
dichotomy illustrated in Figure 1. By providing a globalization-informed weight-
ing scheme, it helps bridge the gap between purely global models (which struggle
with regional specificity) and purely regional models (which lack global coher-
ence). Models optimized with KOF-GI derived α can better adjust their fo-
cus—emphasizing regional performance in various regions with lower globaliza-
tion parity while maintaining stronger global consistency in highly intercon-
nected regions. This creates an adaptive system that can simultaneously satisfy
the competing demands of universal applicability and specialized excellence lead-
ing to solutions that perform robustly across heterogeneous regional contexts. 10
9 Specifically, we use the “de facto interpersonal” component of KOF Globalization
Index as our anchor for GRP, as this aspect measure the globalization in terms of
person-to-person interaction which is a significant aspect for human-centric systems.
10 To help future research working on Anthropogenic Regional Adaptation in vision-
language, we provide the breakdown of the Globalization Index in Appendix G.

-- 13 of 35 --

14 Cahyawijaya et al.
0 	0.2 	0.4 	0.6 	0.8

in terms of
person-to-person interaction which is a significant aspect for human-centric systems.
10 To help future research working on Anthropogenic Regional Adaptation in vision-
language, we provide the breakdown of the Globalization Index in Appendix G.

-- 13 of 35 --

14 Cahyawijaya et al.
0 	0.2 	0.4 	0.6 	0.8 	1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
Google Gemma 3 	SEA-Gemma 10% 	SEA-Gemma (w/o merge)
GRP Factor
GRP Score (Human Eval)
0 	0.2 	0.4 	0.6 	0.8 	1
56
57
58
59
60
61
62
63
64
65
Google Gemma 3 	SEA-Gemma 10% 	SEA-Gemma 50%
SEA-Gemma 70%
GRP Factor
GRP Score (Automatic Eval)	1993 	1998 	2003 	2008 	2013 	2018 	2023
30
40
50
60
70
80
East Asia & Pacific 	Europe & Central Asia
Latin America 	Middle East & North Africa
North America 	South Asia
Southeast Asian 	Sub-Saharan Africa
World
Year
Globalization Index
Fig. 4: (left) Impact of globalization factor alpha to GRP across different models.
Optimizing on a misaligned α can lead to suboptimal performance. (right) We derive
α from the KOF globalization index [25,26] to better reflect the degree of globalization
across regions. The globalization index is distinct across regions and evolves over time.
7 Conclusion
We propose Anthropogenic Regional Adaptation framework which addresses the
critical challenge of cultural insensitivity and reduced performance in under-
represented regions by establishing a foundational paradigm for human-centric
alignment in vision-language systems. The proposed approach, GG-EZ, effec-
tively implements this framework, demonstrating through rigorous testing across
three major architectures—Gemma-3 27B, SDXL, and SigLIP-2— on a compre-
hensive SEA case study, which specifically demonstrated 5-15% gains in cultural
relevance metrics across SEA contexts while maintaining or even improving its
generalization capabilities, enabling equitable deployment of multimodal mod-
els across diverse populations without sacrificing core generalization. Our work
demonstrates the importance of GRP

ecifically demonstrated 5-15% gains in cultural
relevance metrics across SEA contexts while maintaining or even improving its
generalization capabilities, enabling equitable deployment of multimodal mod-
els across diverse populations without sacrificing core generalization. Our work
demonstrates the importance of GRP in human-centric alignment allowing for a
data-driven and dynamically adjusted optimization.
8 Acknowledgements
We extend our gratitude to all affiliated institutions, particularly Oracle and
AISingapore, for providing the compute resources essential to this project. We
express our deepest appreciation to Sarana Nutanong, Genta Indra Winata, Fajri
Koto, Derry Wijaya, Muhamad Risqi U Saputra, Gholamreza Haffari, Enshiun
Annie Lee, and William Tjhi for their invaluable assistance throughout the vari-
ous phases of this project, including idea discussion, information dissemination,
financial support, and other invaluable supports. Additionally, we wish to ac-
knowledge Daryl Peralta, Robert Wijaya, Pume Tuchinda, and Ong Lip Wei
for their dedicated support and expertise in dataset annotations and human
evaluations. We acknowledge the support of the Google Cloud Research Credits

-- 14 of 35 --

Anthropogenic Regional Adaptation 15
program (Award No. GCP19980904), which provided cloud computing credits
used for data set translation. This research is also supported by the National
Research Foundation, Singapore, under its National Large Language Models
Funding Initiative. Any opinions, findings, and conclusions or recommendations
expressed in this material are those of the author(s) and do not reflect the views
of the National Research Foundation, Singapore.
References
1. Adelani, D.I., Abbott, J., Neubig, G., D’Souza, D., Kreutzer, J., Lignos, C., Palen-
Michel, C., Rijhwani, S., et al.: Masakhaner: Named entity recognition for african
languages. Transactions of the Association for Computational Linguistics 9, 1116–
1131 (2021)
2. Adilazuarda, M.F.,

oundation, Singapore.
References
1. Adelani, D.I., Abbott, J., Neubig, G., D’Souza, D., Kreutzer, J., Lignos, C., Palen-
Michel, C., Rijhwani, S., et al.: Masakhaner: Named entity recognition for african
languages. Transactions of the Association for Computational Linguistics 9, 1116–
1131 (2021)
2. Adilazuarda, M.F., Mukherjee, S., Lavania, P., Singh, S.S., Aji, A.F., O’Neill,
J., Modi, A., Choudhury, M.: Towards measuring and modeling “culture” in
LLMs: A survey. In: Al-Onaizan, Y., Bansal, M., Chen, Y.N. (eds.) Proceed-
ings of the 2024 Conference on Empirical Methods in Natural Language Process-
ing. pp. 15763–15784. Association for Computational Linguistics, Miami, Florida,
USA (Nov 2024). https://doi.org/10.18653/v1/2024.emnlp-main.882, https:
//aclanthology.org/2024.emnlp-main.882/
3. Agarwal, A., Meghwani, H., Patel, H.L., Sheng, T., Ravi, S., Roth, D.: Aligning
LLMs for multilingual consistency in enterprise applications. In: Potdar, S., Rojas-
Barahona, L., Montella, S. (eds.) Proceedings of the 2025 Conference on Empirical
Methods in Natural Language Processing: Industry Track. pp. 117–137. Association
for Computational Linguistics, Suzhou (China) (Nov 2025). https://doi.org/10.
18653/v1/2025.emnlp- industry.9, https://aclanthology.org/2025.emnlp-
industry.9/
4. Agrawal, P., Antoniak, S., Hanna, E.B., Bout, B., Chaplot, D., Chudnovsky, J.,
Costa, D., Monicault, B.D., Garg, S., Gervet, T., Ghosh, S., Héliou, A., Jacob,
P., Jiang, A.Q., Khandelwal, K., Lacroix, T., Lample, G., Casas, D.L., Lavril, T.,
Scao, T.L., Lo, A., Marshall, W., Martin, L., Mensch, A., Muddireddy, P., Nemy-
chnikova, V., Pellat, M., Platen, P.V., Raghuraman, N., Rozière, B., Sablayrolles,
A., Saulnier, L., Sauvestre, R., Shang, W., Soletskyi, R., Stewart, L., Stock, P.,
Studnia, J., Subramanian, S., Vaze, S., Wang, T., Yang, S.: Pixtral 12b (2024),
https://arxiv.org/abs/2410.07073
5. Alam, N., Kanjula, K.R., Guthikonda, S., Chung, T., Vegesna, B.K.S., Das, A.,
Susevski, A., Chan,

ère, B., Sablayrolles,
A., Saulnier, L., Sauvestre, R., Shang, W., Soletskyi, R., Stewart, L., Stock, P.,
Studnia, J., Subramanian, S., Vaze, S., Wang, T., Yang, S.: Pixtral 12b (2024),
https://arxiv.org/abs/2410.07073
5. Alam, N., Kanjula, K.R., Guthikonda, S., Chung, T., Vegesna, B.K.S., Das, A.,
Susevski, A., Chan, R.S.Y., Uddin, S.M.I., Islam, S.B., Santhosh, R., A, S., Sharma,
D., Liu, C., Chaturvedi, I., Winata, G.I., S, A., Mukherjee, S., Aji, A.F.: Maya:
An instruction finetuned multilingual multimodal model (2024), https://arxiv.
org/abs/2412.07112
6. Anugraha, D., Irawan, P.A., Singh, A., Lee, E.S.A., Winata, G.I.: M4-rag:
A massive-scale multilingual multi-cultural multimodal rag. arXiv preprint
arXiv:2512.05959 (2025)
7. Axel, D., Noel, G., Pim, M., Lotte, V.B.: Measuring globalization opening the black
box. a critical analysis of globalization indices. Journal of Globalization Studies
1(1), 166–185 (2010)
8. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao,
C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li,
Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D.,

-- 15 of 35 --

16 Cahyawijaya et al.
Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun,
Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P., Wang, Q., Wang, Y., Xie, T.,
Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F.,
Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y.,
Zhu, K.: Qwen3-vl technical report (2025), https://arxiv.org/abs/2511.21631
9. Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang,
S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W.,
Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H.,
Lin, J.: Qwen2.5-vl technical report (2025), https://arxiv.org/abs/2502.13923
10. Cahyawijaya, S.: Llm for

J., Ge, W., Song, S., Dang, K., Wang, P., Wang,
S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W.,
Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H.,
Lin, J.: Qwen2.5-vl technical report (2025), https://arxiv.org/abs/2502.13923
10. Cahyawijaya, S.: Llm for everyone: Representing the underrepresented in large
language models (2024), https://arxiv.org/abs/2409.13897
11. Cahyawijaya, S., Chen, D., Bang, Y., Khalatbari, L., Wilie, B., Ji, Z., Ishii, E.,
Fung, P.: High-dimension human value representation in large language models.
In: Chiruzzo, L., Ritter, A., Wang, L. (eds.) Proceedings of the 2025 Confer-
ence of the Nations of the Americas Chapter of the Association for Computa-
tional Linguistics: Human Language Technologies (Volume 1: Long Papers). pp.
5303–5330. Association for Computational Linguistics, Albuquerque, New Mex-
ico (Apr 2025). https://doi.org/10.18653/v1/2025.naacl-long.274, https:
//aclanthology.org/2025.naacl-long.274/
12. Cahyawijaya, S., Lovenia, H., Aji, A.F., Winata, G., Wilie, B., Koto, F., Ma-
hendra, R., Wibisono, C., Romadhony, A., Vincentio, K., Santoso, J., Moeljadi,
D., Wirawan, C., Hudi, F., Wicaksono, M.S., Parmonangan, I., Alfina, I., Putra,
I.F., Rahmadani, S., Oenang, Y., Septiandri, A., Jaya, J., Dhole, K., Suryani,
A., Putri, R.A., Su, D., Stevens, K., Nityasya, M.N., Adilazuarda, M., Hadiwi-
jaya, R., Diandaru, R., Yu, T., Ghifari, V., Dai, W., Xu, Y., Damapuspita, D.,
Wibowo, H., Tho, C., Karo Karo, I., Fatyanosa, T., Ji, Z., Neubig, G., Bald-
win, T., Ruder, S., Fung, P., Sujaini, H., Sakti, S., Purwarianti, A.: NusaCrowd:
Open source initiative for Indonesian NLP resources. In: Rogers, A., Boyd-Graber,
J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics:
ACL 2023. pp. 13745–13818. Association for Computational Linguistics, Toronto,
Canada (Jul 2023). https://doi.org/10.18653/v1/2023.findings-

A.: NusaCrowd:
Open source initiative for Indonesian NLP resources. In: Rogers, A., Boyd-Graber,
J., Okazaki, N. (eds.) Findings of the Association for Computational Linguistics:
ACL 2023. pp. 13745–13818. Association for Computational Linguistics, Toronto,
Canada (Jul 2023). https://doi.org/10.18653/v1/2023.findings- acl.868,
https://aclanthology.org/2023.findings-acl.868/
13. Cahyawijaya, S., Lovenia, H., Koto, F., Adhista, D., Dave, E., Oktavianti, S.,
Akbar, S., Lee, J., Shadieq, N., Cenggoro, T.W., Linuwih, H., Wilie, B., Muri-
dan, G., Winata, G., Moeljadi, D., Aji, A.F., Purwarianti, A., Fung, P.: Nu-
saWrites: Constructing high-quality corpora for underrepresented and extremely
low-resource languages. In: Park, J.C., Arase, Y., Hu, B., Lu, W., Wijaya, D.,
Purwarianti, A., Krisnadhi, A.A. (eds.) Proceedings of the 13th International
Joint Conference on Natural Language Processing and the 3rd Conference of the
Asia-Pacific Chapter of the Association for Computational Linguistics (Volume
1: Long Papers). pp. 921–945. Association for Computational Linguistics, Nusa
Dua, Bali (Nov 2023). https://doi.org/10.18653/v1/2023.ijcnlp-main.60,
https://aclanthology.org/2023.ijcnlp-main.60/
14. Cahyawijaya, S., Lovenia, H., Koto, F., Putri, R.A., Dave, E., Lee, J., Shadieq,
N., Cenggoro, W., Akbar, S.M., Mahendra, M.I., Putri, D.A., Wilie, B., Winata,
G.I., Aji, A.F., Purwarianti, A., Fung, P.: Cendol: Open instruction-tuned gen-
erative large language models for Indonesian languages. In: Ku, L.W., Martins,
A., Srikumar, V. (eds.) Proceedings of the 62nd Annual Meeting of the As-
sociation for Computational Linguistics (Volume 1: Long Papers). pp. 14899–
14914. Association for Computational Linguistics, Bangkok, Thailand (Aug 2024).

-- 16 of 35 --

Anthropogenic Regional Adaptation 17
https://doi.org/10.18653/v1/2024.acl- long.796, https://aclanthology.
org/2024.acl-long.796/
15. Cahyawijaya, S., Lovenia, H., Moniz, J.R.A., Wong, T.H., Farhansyah, M.R.,
Maung, T.T.,

14914. Association for Computational Linguistics, Bangkok, Thailand (Aug 2024).

-- 16 of 35 --

Anthropogenic Regional Adaptation 17
https://doi.org/10.18653/v1/2024.acl- long.796, https://aclanthology.
org/2024.acl-long.796/
15. Cahyawijaya, S., Lovenia, H., Moniz, J.R.A., Wong, T.H., Farhansyah, M.R.,
Maung, T.T., Hudi, F., Anugraha, D., Habibi, M.R.S., Qorib, M.R., Agarwal,
A., Imperial, J.M., Patel, H.L., Feliren, V., Nasution, B.I., Rufino, M.A., Winata,
G.I., Rajagede, R.A., Catalan, C.R., Imam, M.F.M., Pattnayak, P., Pranida, S.Z.,
Pratama, K., Bangera, Y., Na-Thalang, A., Monderin, P.N., Song, Y., Simon, C.,
Ng, L.H.X., Sapan, R.L., Rafi, T.H., Wang, B., Supryadi, Veerakanjana, K., It-
tichaiwong, P., Roque, M.T., Vincentio, K., Kreangphet, T., Artkaew, P., Palgu-
nadi, K.H., Yu, Y., Hastuti, R.P., Nixon, W., Bangera, M., Lim, A.X.W., Khine,
A.H., Zhafran, H.M., Ferdinan, T., Izzani, A.A., Singh, A., Evan, E., Krito, J.A.,
Anugraha, M., Ilasariya, F.A., Li, H., Daniswara, J.A., Tjiaranata, F.A., Yulianri-
fat, E.P., Udomcharoenchaikit, C., Ansori, F.R., Ihsani, M.K., Nguyen, G., Barik,
A.M., Velasco, D.J., Genadi, R.A., Saha, S., Wei, C., Flores, I.E.W., Han, K.C.K.,
Santos, A.G.D., Lim, W.S., Phyo, K.S., Santos, T., Dwiastuti, M., Luo, J., Cruz,
J.C.B., Hee, M.S., Hanif, I.A., Hakim, M.A., Sya’ban, M.R., Kerdthaisong, K., Mi-
randa, L.J.V., Koto, F., Fatyanosa, T.N., Aji, A.F., Rosal, J.J., Kevin, J., Wijaya,
R., Kampman, O.P., Zhang, R., Karlsson, B.F., Limkonchotiwat, P.: Crowdsource,
crawl, or generate? creating SEA-VL, a multicultural vision-language dataset for
Southeast Asia. In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T. (eds.) Pro-
ceedings of the 63rd Annual Meeting of the Association for Computational Linguis-
tics (Volume 1: Long Papers). pp. 18685–18717. Association for Computational Lin-
guistics, Vienna, Austria (Jul 2025). https://doi.org/10.18653/v1/2025.acl-
long.916, https://aclanthology.org/2025.acl-long.916/
16. Cahyawijaya,

s.) Pro-
ceedings of the 63rd Annual Meeting of the Association for Computational Linguis-
tics (Volume 1: Long Papers). pp. 18685–18717. Association for Computational Lin-
guistics, Vienna, Austria (Jul 2025). https://doi.org/10.18653/v1/2025.acl-
long.916, https://aclanthology.org/2025.acl-long.916/
16. Cahyawijaya, S., Lovenia, H., Moniz, J.R.A., Wong, T.H., Farhansyah, M.R.,
Maung, T.T., Hudi, F., Anugraha, D., Habibi, M.R.S., Qorib, M.R., et al.: Sea-vl:
A multicultural vision-language dataset for southeast asia. Proceedings of the 62nd
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long
Papers) pp. 18685–18717 (2025)
17. Cecilia Liu, C., Koto, F., Baldwin, T., Gurevych, I.: Are multilingual LLMs
culturally-diverse reasoners? an investigation into multicultural proverbs and say-
ings. In: Duh, K., Gomez, H., Bethard, S. (eds.) Proceedings of the 2024 Con-
ference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies (Volume 1: Long Papers). pp. 2016–
2039. Association for Computational Linguistics, Mexico City, Mexico (Jun 2024).
https://doi.org/10.18653/v1/2024.naacl-long.112, https://aclanthology.
org/2024.naacl-long.112/
18. Cohere, T., :, Aakanksha, Ahmadian, A., Ahmed, M., Alammar, J., Alizadeh,
M., Alnumay, Y., Althammer, S., Arkhangorodsky, A., Aryabumi, V., Aumiller,
D., Avalos, R., Aviv, Z., Bae, S., Baji, S., Barbet, A., Bartolo, M., Bebensee, B.,
Beladia, N., Beller-Morales, W., Bérard, A., Berneshawi, A., Bialas, A., Blunsom,
P., Bobkin, M., Bongale, A., Braun, S., Brunet, M., Cahyawijaya, S., Cairuz, D.,
Campos, J.A., Cao, C., Cao, K., Castagné, R., Cendrero, J., Currie, L.C., Chandak,
Y., Chang, D., Chatziveroglou, G., Chen, H., Cheng, C., Chevalier, A., Chiu, J.T.,
Cho, E., Choi, E., Choi, E., Chung, T., Cirik, V., Cismaru, A., Clavier, P., Conklin,
H., Crawhall-Stein, L., Crouse, D., Cruz-Salinas, A.F., Cyrus, B., D’souza, D.,
Dalla-Torre, H., Dang, J., Darling,

ndrero, J., Currie, L.C., Chandak,
Y., Chang, D., Chatziveroglou, G., Chen, H., Cheng, C., Chevalier, A., Chiu, J.T.,
Cho, E., Choi, E., Choi, E., Chung, T., Cirik, V., Cismaru, A., Clavier, P., Conklin,
H., Crawhall-Stein, L., Crouse, D., Cruz-Salinas, A.F., Cyrus, B., D’souza, D.,
Dalla-Torre, H., Dang, J., Darling, W., Domingues, O.D., Dash, S., Debugne, A.,
Dehaze, T., Desai, S., Devassy, J., Dholakia, R., Duffy, K., Edalati, A., Eldeib, A.,
Elkady, A., Elsharkawy, S., Ergün, I., Ermis, B., Fadaee, M., Fan, B., Fayoux, L.,
Flet-Berliac, Y., Frosst, N., Gallé, M., Galuba, W., Garg, U., Geist, M., Azar, M.G.,

-- 17 of 35 --

18 Cahyawijaya et al.
Gilsenan-McMahon, E., Goldfarb-Tarrant, S., Goldsack, T., Gomez, A., Gonzaga,
V.M., Govindarajan, N., Govindassamy, M., Grinsztajn, N., Gritsch, N., Gu, P.,
Guo, S., Haefeli, K., Hajjar, R., Hawes, T., He, J., Hofstätter, S., Hong, S., Hooker,
S., Hosking, T., Howe, S., Hu, E., Huang, R., Jain, H., Jain, R., Jakobi, N., Jenkins,
M., Jordan, J., Joshi, D., Jung, J., Kalyanpur, T., Kamalakara, S.R., Kedrzycki, J.,
Keskin, G., Kim, E., Kim, J., Ko, W.Y., Kocmi, T., Kozakov, M., Kryściński, W.,
Jain, A.K., Teru, K.K., Land, S., Lasby, M., Lasche, O., Lee, J., Lewis, P., Li, J., Li,
J., Lin, H., Locatelli, A., Luong, K., Ma, R., Mach, L., Machado, M., Magbitang,
J., Lopez, B.M., Mann, A., Marchisio, K., Markham, O., Matton, A., McKinney,
A., McLoughlin, D., Mokry, J., Morisot, A., Moulder, A., Moynehan, H., Mozes,
M., Muppalla, V., Murakhovska, L., Nagarajan, H., Nandula, A., Nasir, H., Nehra,
S., Netto-Rosen, J., Ohashi, D., Owers-Bardsley, J., Ozuzu, J., Padilla, D., Park,
G., Passaglia, S., Pekmez, J., Penstone, L., Piktus, A., Ploeg, C., Poulton, A., Qi,
Y., Raghvendra, S., Ramos, M., Ranjan, E., Richemond, P., Robert-Michon, C.,
Rodriguez, A., Roy, S., Ruder, S., Ruis, L., Rust, L., Sachan, A., Salamanca, A.,
Saravanakumar, K.K., Satyakam, I., Sebag, A.S., Sen, P., Sepehri, S., Seshadri,
P., Shen, Y., Sherborne, T.,

nstone, L., Piktus, A., Ploeg, C., Poulton, A., Qi,
Y., Raghvendra, S., Ramos, M., Ranjan, E., Richemond, P., Robert-Michon, C.,
Rodriguez, A., Roy, S., Ruder, S., Ruis, L., Rust, L., Sachan, A., Salamanca, A.,
Saravanakumar, K.K., Satyakam, I., Sebag, A.S., Sen, P., Sepehri, S., Seshadri,
P., Shen, Y., Sherborne, T., Shi, S.S., Shivaprasad, S., Shmyhlo, V., Shrinivason,
A., Shteinbuk, I., Shukayev, A., Simard, M., Snyder, E., Spataru, A., Spooner, V.,
Starostina, T., Strub, F., Su, Y., Sun, J., Talupuru, D., Tarassov, E., Tommasone,
E., Tracey, J., Trend, B., Tumer, E., Üstün, A., Venkitesh, B., Venuto, D., Verga,
P., Voisin, M., Wang, A., Wang, D., Wang, S., Wen, E., White, N., Willman, J.,
Winkels, M., Xia, C., Xie, J., Xu, M., Yang, B., Yi-Chern, T., Zhang, I., Zhao, Z.,
Zhao, Z.: Command a: An enterprise-ready large language model (2025), https:
//arxiv.org/abs/2504.00698
19. Dash, S., Nan, Y., Dang, J., Ahmadian, A., Singh, S., Smith, M., Venkitesh,
B., Shmyhlo, V., Aryabumi, V., Beller-Morales, W., Pekmez, J., Ozuzu, J.,
Richemond, P.H., Locatelli, A., Frosst, N., Blunsom, P., Gomez, A., Zhang,
I., Fadaee, M., Govindassamy, M., Roy, S., Gallé, M., Ermis, B., Üstün, A.,
Hooker, S.: Aya vision: Advancing the frontier of multilingual multimodality (2026),
https://openreview.net/forum?id=DinXMuw6ED
20. Deitke, M., Clark, C., Lee, S., Tripathi, R., Yang, Y., Park, J.S., Salehi, M., Muen-
nighoff, N., Lo, K., Soldaini, L., Lu, J., Anderson, T., Bransom, E., Ehsani, K.,
Ngo, H., Chen, Y., Patel, A., Yatskar, M., Callison-Burch, C., Head, A., Hen-
drix, R., Bastani, F., VanderBilt, E., Lambert, N., Chou, Y., Chheda, A., Sparks,
J., Skjonsberg, S., Schmitz, M., Sarnat, A., Bischoff, B., Walsh, P., Newell, C.,
Wolters, P., Gupta, T., Zeng, K.H., Borchardt, J., Groeneveld, D., Nam, C., Le-
brecht, S., Wittlif, C., Schoenick, C., Michel, O., Krishna, R., Weihs, L., Smith,
N.A., Hajishirzi, H., Girshick, R., Farhadi, A., Kembhavi, A.: Molmo and pixmo:
Open weights and

hmitz, M., Sarnat, A., Bischoff, B., Walsh, P., Newell, C.,
Wolters, P., Gupta, T., Zeng, K.H., Borchardt, J., Groeneveld, D., Nam, C., Le-
brecht, S., Wittlif, C., Schoenick, C., Michel, O., Krishna, R., Weihs, L., Smith,
N.A., Hajishirzi, H., Girshick, R., Farhadi, A., Kembhavi, A.: Molmo and pixmo:
Open weights and open data for state-of-the-art vision-language models (2024),
https://arxiv.org/abs/2409.17146
21. Figge, L., Martens, P.: Globalisation continues: The maastricht globalisation index
revisited and updated. Globalizations 11(6), 875–893 (2014). https://doi.org/
10.1080/14747731.2014.887389, https://doi.org/10.1080/14747731.2014.
887389
22. Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let-
man, A., Mathur, A., Schelten, A., Vaughan, A., Yang, A., Fan, A., Goyal, A.,
Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., Korenev, A., Hinsvark, A.,
Rao, A., Zhang, A., Rodriguez, A., Gregerson, A., Spataru, A., Roziere, B., Biron,
B., Tang, B., Chern, B., Caucheteux, C., Nayak, C., Bi, C., Marra, C., McConnell,
C., Keller, C., Touret, C., Wu, C., Wong, C., Ferrer, C.C., Nikolaidis, C., Allon-

-- 18 of 35 --

Anthropogenic Regional Adaptation 19
sius, D., Song, D., Pintz, D., Livshits, D., Wyatt, D., Esiobu, D., Choudhary, D.,
Mahajan, D., Garcia-Olano, D., Perino, D., Hupkes, D., Lakomkin, E., AlBadawy,
E., Lobanova, E., Dinan, E., Smith, E.M., Radenovic, F., Guzmán, F., Zhang, F.,
Synnaeve, G., Lee, G., Anderson, G.L., Thattai, G., Nail, G., Mialon, G., Pang,
G., Cucurell, G., Nguyen, H., Korevaar, H., Xu, H., Touvron, H., Zarov, I., Ibarra,
I.A., Kloumann, I., Misra, I., Evtimov, I., Zhang, J., Copet, J., Lee, J., Geffert, J.,
Vranes, J., Park, J., Mahadeokar, J., Shah, J., van der Linde, J., Billock, J., Hong,
J., Lee, J., Fu, J., Chi, J., Huang, J., Liu, J., Wang, J., Yu, J., Bitton, J., Spisak,
J., Park, J., Rocca, J., Johnstun, J., Saxe, J., Jia, J., Alwala, K.V., Prasad, K.,
Upasani, K., Plawiak, K., Li, K., Heafield,

J., Geffert, J.,
Vranes, J., Park, J., Mahadeokar, J., Shah, J., van der Linde, J., Billock, J., Hong,
J., Lee, J., Fu, J., Chi, J., Huang, J., Liu, J., Wang, J., Yu, J., Bitton, J., Spisak,
J., Park, J., Rocca, J., Johnstun, J., Saxe, J., Jia, J., Alwala, K.V., Prasad, K.,
Upasani, K., Plawiak, K., Li, K., Heafield, K., Stone, K., El-Arini, K., Iyer, K.,
Malik, K., Chiu, K., Bhalla, K., Lakhotia, K., Rantala-Yeary, L., van der Maaten,
L., Chen, L., Tan, L., Jenkins, L., Martin, L., Madaan, L., Malo, L., Blecher, L.,
Landzaat, L., de Oliveira, L., Muzzi, M., Pasupuleti, M., Singh, M., Paluri, M.,
Kardas, M., Tsimpoukelli, M., Oldham, M., Rita, M., Pavlova, M., Kambadur,
M., Lewis, M., Si, M., Singh, M.K., Hassan, M., Goyal, N., Torabi, N., Bashlykov,
N., Bogoychev, N., Chatterji, N., Zhang, N., Duchenne, O., Çelebi, O., Alrassy, P.,
Zhang, P., Li, P., Vasic, P., Weng, P., Bhargava, P., Dubal, P., Krishnan, P., Koura,
P.S., Xu, P., He, Q., Dong, Q., Srinivasan, R., Ganapathy, R., Calderer, R., Cabral,
R.S., Stojnic, R., Raileanu, R., Maheswari, R., Girdhar, R., Patel, R., Sauvestre,
R., Polidoro, R., Sumbaly, R., Taylor, R., Silva, R., Hou, R., Wang, R., Hosseini,
S., Chennabasappa, S., Singh, S., Bell, S., Kim, S.S., Edunov, S., Nie, S., Narang,
S., Raparthy, S., Shen, S., Wan, S., Bhosale, S., Zhang, S., Vandenhende, S., Batra,
S., Whitman, S., Sootla, S., Collot, S., Gururangan, S., Borodinsky, S., Herman,
T., Fowler, T., Sheasha, T., Georgiou, T., Scialom, T., Speckbacher, T., Mihaylov,
T., Xiao, T., Karn, U., Goswami, V., Gupta, V., Ramanathan, V., Kerkez, V.,
Gonguet, V., Do, V., Vogeti, V., Albiero, V., Petrovic, V., Chu, W., Xiong, W.,
Fu, W., Meers, W., Martinet, X., Wang, X., Wang, X., Tan, X.E., Xia, X., Xie,
X., Jia, X., Wang, X., Goldschlag, Y., Gaur, Y., Babaei, Y., Wen, Y., Song, Y.,
Zhang, Y., Li, Y., Mao, Y., Coudert, Z.D., Yan, Z., Chen, Z., Papakipos, Z., Singh,
A., Srivastava, A., Jain, A., Kelsey, A., Shajnfeld, A., Gangidi, A., Victoria,

, W., Meers, W., Martinet, X., Wang, X., Wang, X., Tan, X.E., Xia, X., Xie,
X., Jia, X., Wang, X., Goldschlag, Y., Gaur, Y., Babaei, Y., Wen, Y., Song, Y.,
Zhang, Y., Li, Y., Mao, Y., Coudert, Z.D., Yan, Z., Chen, Z., Papakipos, Z., Singh,
A., Srivastava, A., Jain, A., Kelsey, A., Shajnfeld, A., Gangidi, A., Victoria, A.,
Goldstand, A., Menon, A., Sharma, A., Boesenberg, A., Baevski, A., Feinstein, A.,
Kallet, A., Sangani, A., Teo, A., Yunus, A., Lupu, A., Alvarado, A., Caples, A.,
Gu, A., Ho, A., Poulton, A., Ryan, A., Ramchandani, A., Dong, A., Franco, A.,
Goyal, A., Saraf, A., Chowdhury, A., Gabriel, A., Bharambe, A., Eisenman, A.,
Yazdan, A., James, B., Maurer, B., Leonhardi, B., Huang, B., Loyd, B., Paola,
B.D., Paranjape, B., Liu, B., Wu, B., Ni, B., Hancock, B., Wasti, B., Spence, B.,
Stojkovic, B., Gamido, B., Montalvo, B., Parker, C., Burton, C., Mejia, C., Liu,
C., Wang, C., Kim, C., Zhou, C., Hu, C., Chu, C.H., Cai, C., Tindal, C., Feicht-
enhofer, C., Gao, C., Civin, D., Beaty, D., Kreymer, D., Li, D., Adkins, D., Xu,
D., Testuggine, D., David, D., Parikh, D., Liskovich, D., Foss, D., Wang, D., Le,
D., Holland, D., Dowling, E., Jamil, E., Montgomery, E., Presani, E., Hahn, E.,
Wood, E., Le, E.T., Brinkman, E., Arcaute, E., Dunbar, E., Smothers, E., Sun, F.,
Kreuk, F., Tian, F., Kokkinos, F., Ozgenel, F., Caggioni, F., Kanayet, F., Seide,
F., Florez, G.M., Schwarz, G., Badeer, G., Swee, G., Halpern, G., Herman, G.,
Sizov, G., Guangyi, Zhang, Lakshminarayanan, G., Inan, H., Shojanazeri, H., Zou,
H., Wang, H., Zha, H., Habeeb, H., Rudolph, H., Suk, H., Aspegren, H., Goldman,
H., Zhan, H., Damlaj, I., Molybog, I., Tufanov, I., Leontiadis, I., Veliche, I.E., Gat,
I., Weissman, J., Geboski, J., Kohli, J., Lam, J., Asher, J., Gaya, J.B., Marcus, J.,
Tang, J., Chan, J., Zhen, J., Reizenstein, J., Teboul, J., Zhong, J., Jin, J., Yang,
J., Cummings, J., Carvill, J., Shepard, J., McPhie, J., Torres, J., Ginsburg, J.,

-- 19 of 35 --

20 Cahyawijaya et al.
Wang,

, Veliche, I.E., Gat,
I., Weissman, J., Geboski, J., Kohli, J., Lam, J., Asher, J., Gaya, J.B., Marcus, J.,
Tang, J., Chan, J., Zhen, J., Reizenstein, J., Teboul, J., Zhong, J., Jin, J., Yang,
J., Cummings, J., Carvill, J., Shepard, J., McPhie, J., Torres, J., Ginsburg, J.,

-- 19 of 35 --

20 Cahyawijaya et al.
Wang, J., Wu, K., U, K.H., Saxena, K., Khandelwal, K., Zand, K., Matosich, K.,
Veeraraghavan, K., Michelena, K., Li, K., Jagadeesh, K., Huang, K., Chawla, K.,
Huang, K., Chen, L., Garg, L., A, L., Silva, L., Bell, L., Zhang, L., Guo, L., Yu,
L., Moshkovich, L., Wehrstedt, L., Khabsa, M., Avalani, M., Bhatt, M., Mankus,
M., Hasson, M., Lennie, M., Reso, M., Groshev, M., Naumov, M., Lathi, M., Ke-
neally, M., Liu, M., Seltzer, M.L., Valko, M., Restrepo, M., Patel, M., Vyatskov,
M., Samvelyan, M., Clark, M., Macey, M., Wang, M., Hermoso, M.J., Metanat,
M., Rastegari, M., Bansal, M., Santhanam, N., Parks, N., White, N., Bawa, N.,
Singhal, N., Egebo, N., Usunier, N., Mehta, N., Laptev, N.P., Dong, N., Cheng,
N., Chernoguz, O., Hart, O., Salpekar, O., Kalinli, O., Kent, P., Parekh, P., Saab,
P., Balaji, P., Rittner, P., Bontrager, P., Roux, P., Dollar, P., Zvyagina, P., Ratan-
chandani, P., Yuvraj, P., Liang, Q., Alao, R., Rodriguez, R., Ayub, R., Murthy,
R., Nayani, R., Mitra, R., Parthasarathy, R., Li, R., Hogan, R., Battey, R., Wang,
R., Howes, R., Rinott, R., Mehta, S., Siby, S., Bondu, S.J., Datta, S., Chugh, S.,
Hunt, S., Dhillon, S., Sidorov, S., Pan, S., Mahajan, S., Verma, S., Yamamoto, S.,
Ramaswamy, S., Lindsay, S., Lindsay, S., Feng, S., Lin, S., Zha, S.C., Patil, S.,
Shankar, S., Zhang, S., Zhang, S., Wang, S., Agarwal, S., Sajuyigbe, S., Chintala,
S., Max, S., Chen, S., Kehoe, S., Satterfield, S., Govindaprasad, S., Gupta, S.,
Deng, S., Cho, S., Virk, S., Subramanian, S., Choudhury, S., Goldman, S., Remez,
T., Glaser, T., Best, T., Koehler, T., Robinson, T., Li, T., Zhang, T., Matthews,
T., Chou, T., Shaked, T., Vontimitta, V., Ajayi, V., Montanez,

., Chintala,
S., Max, S., Chen, S., Kehoe, S., Satterfield, S., Govindaprasad, S., Gupta, S.,
Deng, S., Cho, S., Virk, S., Subramanian, S., Choudhury, S., Goldman, S., Remez,
T., Glaser, T., Best, T., Koehler, T., Robinson, T., Li, T., Zhang, T., Matthews,
T., Chou, T., Shaked, T., Vontimitta, V., Ajayi, V., Montanez, V., Mohan, V.,
Kumar, V.S., Mangla, V., Ionescu, V., Poenaru, V., Mihailescu, V.T., Ivanov, V.,
Li, W., Wang, W., Jiang, W., Bouaziz, W., Constable, W., Tang, X., Wu, X.,
Wang, X., Wu, X., Gao, X., Kleinman, Y., Chen, Y., Hu, Y., Jia, Y., Qi, Y., Li,
Y., Zhang, Y., Zhang, Y., Adi, Y., Nam, Y., Yu, Wang, Zhao, Y., Hao, Y., Qian,
Y., Li, Y., He, Y., Rait, Z., DeVito, Z., Rosnbrick, Z., Wen, Z., Yang, Z., Zhao, Z.,
Ma, Z.: The llama 3 herd of models (2024), https://arxiv.org/abs/2407.21783
23. Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma,
S., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z.F., Gou, Z., Shao, Z., Li, Z., Gao, Z.,
Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Ruan,
C., Dai, D., Chen, D., Ji, D., Li, E., Lin, F., Dai, F., Luo, F., Hao, G., Chen, G., Li,
G., Zhang, H., Xu, H., Ding, H., Gao, H., Qu, H., Li, H., Guo, J., Li, J., Chen, J.,
Yuan, J., Tu, J., Qiu, J., Li, J., Cai, J.L., Ni, J., Liang, J., Chen, J., Dong, K., Hu,
K., You, K., Gao, K., Guan, K., Huang, K., Yu, K., Wang, L., Zhang, L., Zhao, L.,
Wang, L., Zhang, L., Xu, L., Xia, L., Zhang, M., Zhang, M., Tang, M., Zhou, M.,
Li, M., Wang, M., Li, M., Tian, N., Huang, P., Zhang, P., Wang, Q., Chen, Q., Du,
Q., Ge, R., Zhang, R., Pan, R., Wang, R., Chen, R.J., Jin, R.L., Chen, R., Lu, S.,
Zhou, S., Chen, S., Ye, S., Wang, S., Yu, S., Zhou, S., Pan, S., Li, S.S., Zhou, S.,
Wu, S., Yun, T., Pei, T., Sun, T., Wang, T., Zeng, W., Liu, W., Liang, W., Gao,
W., Yu, W., Zhang, W., Xiao, W.L., An, W., Liu, X., Wang, X., Chen, X., Nie,
X., Cheng, X., Liu, X., Xie, X., Liu, X., Yang, X., Li, X., Su, X., Lin, X., Li, X.Q.,
Jin, X., Shen,

S., Wang, S., Yu, S., Zhou, S., Pan, S., Li, S.S., Zhou, S.,
Wu, S., Yun, T., Pei, T., Sun, T., Wang, T., Zeng, W., Liu, W., Liang, W., Gao,
W., Yu, W., Zhang, W., Xiao, W.L., An, W., Liu, X., Wang, X., Chen, X., Nie,
X., Cheng, X., Liu, X., Xie, X., Liu, X., Yang, X., Li, X., Su, X., Lin, X., Li, X.Q.,
Jin, X., Shen, X., Chen, X., Sun, X., Wang, X., Song, X., Zhou, X., Wang, X.,
Shan, X., Li, Y.K., Wang, Y.Q., Wei, Y.X., Zhang, Y., Xu, Y., Li, Y., Zhao, Y.,
Sun, Y., Wang, Y., Yu, Y., Zhang, Y., Shi, Y., Xiong, Y., He, Y., Piao, Y., Wang,
Y., Tan, Y., Ma, Y., Liu, Y., Guo, Y., Ou, Y., Wang, Y., Gong, Y., Zou, Y., He,
Y., Xiong, Y., Luo, Y., You, Y., Liu, Y., Zhou, Y., Zhu, Y.X., Huang, Y., Li, Y.,
Zheng, Y., Zhu, Y., Ma, Y., Tang, Y., Zha, Y., Yan, Y., Ren, Z.Z., Ren, Z., Sha, Z.,
Fu, Z., Xu, Z., Xie, Z., Zhang, Z., Hao, Z., Ma, Z., Yan, Z., Wu, Z., Gu, Z., Zhu, Z.,
Liu, Z., Li, Z., Xie, Z., Song, Z., Pan, Z., Huang, Z., Xu, Z., Zhang, Z., Zhang, Z.:
Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature

-- 20 of 35 --

Anthropogenic Regional Adaptation 21
645(8081), 633–638 (Sep 2025). https://doi.org/10.1038/s41586-025-09422-z,
http://dx.doi.org/10.1038/s41586-025-09422-z
24. Guo, J., Zheng, T., Bai, Y., Li, B., Wang, Y., Zhu, K., Li, Y., Neubig, G., Chen,
W., Yue, X.: Mammoth-vl: Eliciting multimodal reasoning with instruction tuning
at scale (2024), https://arxiv.org/abs/2412.05237
25. Gygli, S., Haelg, F., Potrafke, N., Sturm, J.E.: The kof globalisation index – re-
visited. The Review of International Organizations 14(3), 543–574 (Jan 2019).
https://doi.org/10.1007/s11558-019-09344-2, http://dx.doi.org/10.1007/
s11558-019-09344-2
26. Haelg, F.: The kof globalisation index – a multidimensional approach to glob-
alisation. Jahrbücher für Nationalökonomie und Statistik 240(5), 691–696 (Sep
2019). https://doi.org/10.1515/jbnst-2019-0045, http://dx.doi.org/10.
1515/jbnst-2019-0045
27. Hennara, K., Hreden, M., Hamed, M.M., Bastati, A.,

8-019-09344-2
26. Haelg, F.: The kof globalisation index – a multidimensional approach to glob-
alisation. Jahrbücher für Nationalökonomie und Statistik 240(5), 691–696 (Sep
2019). https://doi.org/10.1515/jbnst-2019-0045, http://dx.doi.org/10.
1515/jbnst-2019-0045
27. Hennara, K., Hreden, M., Hamed, M.M., Bastati, A., Aldallal, Z., Chrouf,
S., AlModhayan, S.: Baseer: A vision-language model for arabic document-to-
markdown ocr (2025), https://arxiv.org/abs/2509.18174
28. Hu, X., Wang, R., Fang, Y., Fu, B., Cheng, P., Yu, G.: Ella: Equip diffusion models
with llm for enhanced semantic alignment (2024), https://arxiv.org/abs/2403.
05135
29. Jain, A., et al.: Ai4bharat indicbert: A monolingual bert model for indian languages
pp. 10685–10706 (2024)
30. Ju, J., Kim, D., Park, S., Kim, Y.: Varco-vision: Expanding frontiers in korean
vision-language models (2024), https://arxiv.org/abs/2411.19103
31. Kabra, A., Liu, E., Khanuja, S., Aji, A.F., Winata, G., Cahyawijaya, S., Aremu,
A., Ogayo, P., Neubig, G.: Multi-lingual and multi-cultural figurative language
understanding. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of
the Association for Computational Linguistics: ACL 2023. pp. 8269–8284. As-
sociation for Computational Linguistics, Toronto, Canada (Jul 2023). https:
//doi.org/10.18653/v1/2023.findings-acl.525, https://aclanthology.org/
2023.findings-acl.525/
32. Khan, F., et al.: Ai4bharat indicbert: A monolingual bert model for indian lan-
guages pp. 10685–10706 (2024)
33. Kocmi, T., Arkhangorodsky, A., Berard, A., Blunsom, P., Cahyawijaya, S., De-
haze, T., Fadaee, M., Frosst, N., Galle, M., Gomez, A., Govindarajan, N., Ko,
W.Y., Kreutzer, J., Marchisio, K., Üstün, A., Vincent, S., Zhang, I.: Command-a-
translate: Raising the bar of machine translation with difficulty filtering. In: Had-
dow, B., Kocmi, T., Koehn, P., Monz, C. (eds.) Proceedings of the Tenth Confer-
ence on Machine Translation. pp. 789–799. Association for Computational Linguis-
tics,

, J., Marchisio, K., Üstün, A., Vincent, S., Zhang, I.: Command-a-
translate: Raising the bar of machine translation with difficulty filtering. In: Had-
dow, B., Kocmi, T., Koehn, P., Monz, C. (eds.) Proceedings of the Tenth Confer-
ence on Machine Translation. pp. 789–799. Association for Computational Linguis-
tics, Suzhou, China (Nov 2025). https://doi.org/10.18653/v1/2025.wmt-1.55,
https://aclanthology.org/2025.wmt-1.55/
34. Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C.H., Gonzalez, J.E.,
Zhang, H., Stoica, I.: Efficient memory management for large language model serv-
ing with pagedattention. In: Proceedings of the ACM SIGOPS 29th Symposium
on Operating Systems Principles (2023)
35. Labs, C.: Aya vision benchmark (2025), https://huggingface.co/datasets/
CohereLabs/AyaVisionBench
36. Liu, C.C., Gurevych, I., Korhonen, A.: Culturally aware and adapted NLP: A
taxonomy and a survey of the state of the art. Transactions of the Association for
Computational Linguistics 13, 652–689 (2025). https://doi.org/10.1162/tacl_
a_00760, https://aclanthology.org/2025.tacl-1.31/

-- 21 of 35 --

22 Cahyawijaya et al.
37. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv (2017).
https://doi.org/10.48550/arxiv.1711.05101
38. Lovenia, H., Mahendra, R., Akbar, S.M., Miranda, L.J.V., Santoso, J., Aco,
E., Fadhilah, A., Mansurov, J., Imperial, J.M., Kampman, O.P., Moniz, J.R.A.,
Habibi, M.R.S., Hudi, F., Montalan, R., Ignatius, R., Lopo, J.A., Nixon, W., Karls-
son, B.F., Jaya, J., Diandaru, R., Gao, Y., Amadeus, P., Wang, B., Cruz, J.C.B.,
Whitehouse, C., Parmonangan, I.H., Khelli, M., Zhang, W., Susanto, L., Ryanda,
R.A., Hermawan, S.L., Velasco, D.J., Kautsar, M.D.A., Hendria, W.F., Moslem, Y.,
Flynn, N., Adilazuarda, M.F., Li, H., Lee, J., Damanhuri, R., Sun, S., Qorib, M.R.,
Djanibekov, A., Leong, W.Q., Do, Q.V., Muennighoff, N., Pansuwan, T., Putra,
I.F., Xu, Y., Chia, T.N., Purwarianti, A., Ruder, S., Tjhi, W., Limkonchotiwat,
P., Aji, A.F.,

.L., Velasco, D.J., Kautsar, M.D.A., Hendria, W.F., Moslem, Y.,
Flynn, N., Adilazuarda, M.F., Li, H., Lee, J., Damanhuri, R., Sun, S., Qorib, M.R.,
Djanibekov, A., Leong, W.Q., Do, Q.V., Muennighoff, N., Pansuwan, T., Putra,
I.F., Xu, Y., Chia, T.N., Purwarianti, A., Ruder, S., Tjhi, W., Limkonchotiwat,
P., Aji, A.F., Keh, S., Winata, G.I., Zhang, R., Koto, F., Yong, Z.X., Cahyaw-
ijaya, S.: SEACrowd: A multilingual multimodal data hub and benchmark suite
for Southeast Asian languages. In: Al-Onaizan, Y., Bansal, M., Chen, Y.N. (eds.)
Proceedings of the 2024 Conference on Empirical Methods in Natural Language
Processing. pp. 5155–5203. Association for Computational Linguistics, Miami,
Florida, USA (Nov 2024). https://doi.org/10.18653/v1/2024.emnlp-main.296,
https://aclanthology.org/2024.emnlp-main.296/
39. Martens, P., Raza, M.: The maastricht globalisation index: An update, pp. 279–
309. Nova Science Publishers, United States (2009)
40. Mogrovejo, D.O.R., Lyu, C., Wibowo, H.A., Góngora, S., Mandal, A., Purkayastha,
S., Ortiz-Barajas, J.G., Cueva, E.V., Baek, J., Jeong, S., Hamed, I., Yong, Z.X.,
Lim, Z.W., Silva, P.M., Dunstan, J., Jouitteau, M., MEUR, D.L., Nwatu, J., Bat-
nasan, G., Otgonbold, M.E., Gochoo, M., Ivetta, G., Benotti, L., Alemany, L.A.,
Maina, H., Geng, J., Torrent, T.T., Belcavello, F., Viridiano, M., Cruz, J.C.B.,
Velasco, D.J., Ignat, O., Burzo, Z., Whitehouse, C., Abzaliev, A., Clifford, T.,
Caulfield, G., Lynn, T., Salamea-Palacios, C., Araujo, V., Kementchedjhieva, Y.,
Mihaylov, M.M., Azime, I.A., Ademtew, H.B., Balcha, B.F., Etori, N.A., Ade-
lani, D.I., Mihalcea, R., Tonja, A.L., Cabrera, M.C.B., Vallejo, G., Lovenia, H.,
Zhang, R., Estecha-Garitagoitia, M., Rodríguez-Cantelar, M., Ehsan, T., Chevi,
R., Adilazuarda, M.F., Diandaru, R., Cahyawijaya, S., Koto, F., Kuribayashi, T.,
Song, H., Khandavally, A.N.K., Jayakumar, T., Dabre, R., Imam, M.F.M., Na-
gasinghe, K.R.Y., Dragonetti, A., D’Haro, L.F., NIYOMUGISHA, O., Gala, J.,
Chitale, P.A.,

, R., Estecha-Garitagoitia, M., Rodríguez-Cantelar, M., Ehsan, T., Chevi,
R., Adilazuarda, M.F., Diandaru, R., Cahyawijaya, S., Koto, F., Kuribayashi, T.,
Song, H., Khandavally, A.N.K., Jayakumar, T., Dabre, R., Imam, M.F.M., Na-
gasinghe, K.R.Y., Dragonetti, A., D’Haro, L.F., NIYOMUGISHA, O., Gala, J.,
Chitale, P.A., Farooqui, F., Solorio, T., Aji, A.F.: CVQA: Culturally-diverse mul-
tilingual visual question answering benchmark. In: The Thirty-eight Conference on
Neural Information Processing Systems Datasets and Benchmarks Track (2024),
https://openreview.net/forum?id=E18kRXTGmV
41. Mohamed, A., Alwajih, F., Nagoudi, E.M.B., Inciarte, A., Abdul-Mageed, M.:
Violet: A vision-language model for Arabic image captioning with gemini decoder.
In: Sawaf, H., El-Beltagy, S., Zaghouani, W., Magdy, W., Abdelali, A., Tomeh, N.,
Abu Farha, I., Habash, N., Khalifa, S., Keleg, A., Haddad, H., Zitouni, I., Mrini,
K., Almatham, R. (eds.) Proceedings of ArabicNLP 2023. pp. 1–11. Association for
Computational Linguistics, Singapore (Hybrid) (Dec 2023). https://doi.org/10.
18653/v1/2023.arabicnlp-1.1, https://aclanthology.org/2023.arabicnlp-
1.1/
42. Naous, T., Ryan, M.J., Ritter, A., Xu, W.: Having beer after prayer? measuring cul-
tural bias in large language models. In: Ku, L.W., Martins, A., Srikumar, V. (eds.)
Proceedings of the 62nd Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers). pp. 16366–16393. Association for Computa-

-- 22 of 35 --

Anthropogenic Regional Adaptation 23
tional Linguistics, Bangkok, Thailand (Aug 2024). https://doi.org/10.18653/
v1/2024.acl-long.862, https://aclanthology.org/2024.acl-long.862/
43. Ng, R., Nguyen, T.N., Yuli, H., Chia, T.N., Yi, L.W., Leong, W.Q., Yong, X., Ngui,
J.G., Susanto, Y., Cheng, N., Rengarajan, H., Limkonchotiwat, P., Hulagadri, A.V.,
Teng, K.W., Tong, Y.Y., Siow, B., Teo, W.Y., Meng, T.C., Ong, B., Ong, Z.H.,
Montalan, J.R., Chan, A., Antonyrex, S., Lee, R., Choa, E., Tat-Wee, D.O.,

. Ng, R., Nguyen, T.N., Yuli, H., Chia, T.N., Yi, L.W., Leong, W.Q., Yong, X., Ngui,
J.G., Susanto, Y., Cheng, N., Rengarajan, H., Limkonchotiwat, P., Hulagadri, A.V.,
Teng, K.W., Tong, Y.Y., Siow, B., Teo, W.Y., Meng, T.C., Ong, B., Ong, Z.H.,
Montalan, J.R., Chan, A., Antonyrex, S., Lee, R., Choa, E., Tat-Wee, D.O., Liu,
B.J.D., Tjhi, W.C., Cambria, E., Teo, L.: SEA-LION: Southeast Asian languages
in one network. In: Inui, K., Sakti, S., Wang, H., Wong, D.F., Bhattacharyya,
P., Banerjee, B., Ekbal, A., Chakraborty, T., Singh, D.P. (eds.) Proceedings of
the 14th International Joint Conference on Natural Language Processing and the
4th Conference of the Asia-Pacific Chapter of the Association for Computational
Linguistics. pp. 512–526. The Asian Federation of Natural Language Processing
and The Association for Computational Linguistics, Mumbai, India (Dec 2025),
https://aclanthology.org/2025.ijcnlp-long.30/
44. Nguyen, T.S., Qorib, M.R., Ng, H.T.: Openseal: Good, fast, and cheap construction
of an open-source southeast asian llm via parallel data (2026), https://arxiv.
org/abs/2602.02266
45. Nguyen, X.P., Zhang, W., Li, X., Aljunied, M., Hu, Z., Shen, C., Chia, Y.K.,
Li, X., Wang, J., Tan, Q., Cheng, L., Chen, G., Deng, Y., Yang, S., Liu, C.,
Zhang, H., Bing, L.: SeaLLMs - large language models for Southeast Asia. In:
Cao, Y., Feng, Y., Xiong, D. (eds.) Proceedings of ACL. pp. 294–304 (Aug 2024).
https://doi.org/10.18653/v1/2024.acl- demos.28, https://aclanthology.
org/2024.acl-demos.28/
46. Nyandwi, J.D.D., Song, Y., Khanuja, S., Neubig, G.: Grounding multilingual multi-
modal LLMs with cultural knowledge. In: Christodoulopoulos, C., Chakraborty, T.,
Rose, C., Peng, V. (eds.) Proceedings of the 2025 Conference on Empirical Methods
in Natural Language Processing. pp. 24187–24231. Association for Computational
Linguistics, Suzhou, China (Nov 2025). https://doi.org/10.18653/v1/2025.
emnlp-main.1232, https://aclanthology.org/2025.emnlp-main.1232/
47. Nyborg, J.,

, T.,
Rose, C., Peng, V. (eds.) Proceedings of the 2025 Conference on Empirical Methods
in Natural Language Processing. pp. 24187–24231. Association for Computational
Linguistics, Suzhou, China (Nov 2025). https://doi.org/10.18653/v1/2025.
emnlp-main.1232, https://aclanthology.org/2025.emnlp-main.1232/
47. Nyborg, J., Pelletier, C., Lefèvre, S., Assent, I.: Timematch: Unsupervised cross-
region adaptation by temporal shift estimation. ISPRS Journal of Photogram-
metry and Remote Sensing 188, 301–313 (2022). https://doi.org/https:
//doi.org/10.1016/j.isprsjprs.2022.04.018, https://www.sciencedirect.
com/science/article/pii/S0924271622001216
48. Patel, H.L., Agarwal, A., Das, A., Kumar, B., Panda, S., Pattnayak, P., Rafi,
T.H., Kumar, T., Chae, D.K.: Sweeval: Do llms really swear? a safety benchmark
for testing limits for enterprise use. In: Proceedings of the 2025 Conference of the
Nations of the Americas Chapter of the Association for Computational Linguistics:
Human Language Technologies (Volume 3: Industry Track). pp. 558–582 (2025)
49. Singapore, A.: Sea-lion (southeast asian languages in one network): A family of
large language models for southeast asia. https://github.com/aisingapore/
sealion (2024)
50. Steiner, A.e.a.: Paligemma 2: A family of versatile vlms for transfer. arXiv preprint
arXiv:2412.03555 (2024)
51. Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk,
J., Dai, A.M., Hauth, A., Millican, K., Silver, D., Johnson, M., Antonoglou, I.,
Schrittwieser, J., Glaese, A., Chen, J., Pitler, E., Lillicrap, T., Lazaridou, A., Firat,
O., Molloy, J., Isard, M., Barham, P.R., Hennigan, T., Lee, B., Viola, F., Reynolds,
M., Xu, Y., Doherty, R., Collins, E., Meyer, C., Rutherford, E., Moreira, E., Ayoub,
K., Goel, M., Krawczyk, J., Du, C., Chi, E., Cheng, H.T., Ni, E., Shah, P., Kane,

-- 23 of 35 --

24 Cahyawijaya et al.
P., et al.: Gemini: A family of highly capable multimodal models (2023), https:
//arxiv.org/abs/2312.11805
52.

s,
M., Xu, Y., Doherty, R., Collins, E., Meyer, C., Rutherford, E., Moreira, E., Ayoub,
K., Goel, M., Krawczyk, J., Du, C., Chi, E., Cheng, H.T., Ni, E., Shah, P., Kane,

-- 23 of 35 --

24 Cahyawijaya et al.
P., et al.: Gemini: A family of highly capable multimodal models (2023), https:
//arxiv.org/abs/2312.11805
52. Team, G., Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S.,
Matejovicova, T., Ramé, A., Rivière, M., Rouillard, L., Mesnard, T., Cideron, G.,
bastien Grill, J., Ramos, S., Yvinec, E., Casbon, M., Pot, E., Penchev, I., Liu, G.,
Visin, F., Kenealy, K., Beyer, L., Zhai, X., Tsitsulin, A., Busa-Fekete, R., Feng, A.,
Sachdeva, N., Coleman, B., Gao, Y., Mustafa, B., Barr, I., Parisotto, E., Tian, D.,
Eyal, M., Cherry, C., Peter, J.T., Sinopalnikov, D., Bhupatiraju, S., Agarwal, R.,
Kazemi, M., Malkin, D., Kumar, R., Vilar, D., Brusilovsky, I., Luo, J., Steiner, A.,
Friesen, A., Sharma, A., Sharma, A., Gilady, A.M., Goedeckemeyer, A., Saade, A.,
Kolesnikov, A., Bendebury, A., Abdagic, A., Vadi, A., György, A., Pinto, A.S., Das,
A., Bapna, A., Miech, A., Yang, A., Paterson, A., Shenoy, A., Chakrabarti, A., Piot,
B., Wu, B., Shahriari, B., Petrini, B., Chen, C., Lan, C.L., Choquette-Choo, C.A.,
Carey, C., Brick, C., Deutsch, D., Eisenbud, D., Cattle, D., Cheng, D., Paparas,
D., Sreepathihalli, D.S., Reid, D., Tran, D., Zelle, D., Noland, E., Huizenga, E.,
Kharitonov, E., Liu, F., Amirkhanyan, G., Cameron, G., Hashemi, H., Klimczak-
Plucińska, H., Singh, H., Mehta, H., Lehri, H.T., Hazimeh, H., Ballantyne, I.,
Szpektor, I., Nardini, I., Pouget-Abadie, J., Chan, J., Stanton, J., Wieting, J., Lai,
J., Orbay, J., Fernandez, J., Newlan, J., yeong Ji, J., Singh, J., Black, K., Yu,
K., Hui, K., Vodrahalli, K., Greff, K., Qiu, L., Valentine, M., Coelho, M., Ritter,
M., Hoffman, M., Watson, M., Chaturvedi, M., Moynihan, M., Ma, M., Babar, N.,
Byrd, N., Roy, N., Momchev, N., Chauhan, N., Bunyan, O., Botarda, P., Caron, P.,
Rubenstein, P.K.,

z, J., Newlan, J., yeong Ji, J., Singh, J., Black, K., Yu,
K., Hui, K., Vodrahalli, K., Greff, K., Qiu, L., Valentine, M., Coelho, M., Ritter,
M., Hoffman, M., Watson, M., Chaturvedi, M., Moynihan, M., Ma, M., Babar, N.,
Byrd, N., Roy, N., Momchev, N., Chauhan, N., Bunyan, O., Botarda, P., Caron, P.,
Rubenstein, P.K., Culliton, P., Schmid, P., Sessa, P.G., Xu, P., Stanczyk, P., Tafti,
P., Shivanna, R., Wu, R., Pan, R., Rokni, R., Willoughby, R., Vallu, R., Mullins,
R., Jerome, S., Smoot, S., Girgin, S., Iqbal, S., Reddy, S., Sheth, S., Põder, S.,
Bhatnagar, S., Eiger, S., Zhang, S., Liu, T., Yacovone, T., Liechty, T., Kalra, U.,
Evci, U., Misra, V., Roseberry, V., Feinberg, V., Kolesnikov, V., Han, W., Kwon,
W., Chen, X., Chow, Y., Zhu, Y., Wei, Z., Egyed, Z., Cotruta, V., Giang, M., Kirk,
P., Rao, A., Black, K., Babar, N., Lo, J., Moreira, E., Martins, L.G., Sanseviero,
O., Gonzalez, L., Gleicher, Z., Warkentin, T., Mirrokni, V., Senter, E., Collins, E.,
Barral, J., Ghahramani, Z.: Gemma 3 technical report. arXiv:2503.19786 (2025),
https://arxiv.org/abs/2503.19786
53. Team, N., Costa-jussà, M.R., Cross, J., Çelebi, O., Elbayad, M., Heafield, K., Hef-
fernan, K., Kalbassi, E., Lam, J., Licht, D., Maillard, J., Sun, A., Wang, S., Wenzek,
G., Youngblood, A., Akula, B., Barrault, L., Gonzalez, G.M., Hansanti, P., Hoff-
man, J., Jarrett, S., Sadagopan, K.R., Rowe, D., Spruit, S., Tran, C., Andrews, P.,
Ayan, N.F., Bhosale, S., Edunov, S., Fan, A., Gao, C., Goswami, V., Guzmán, F.,
Koehn, P., Mourachko, A., Ropers, C., Saleem, S., Schwenk, H., Wang, J.: No lan-
guage left behind: Scaling human-centered machine translation. arXiv:2207.04672
(2022), https://arxiv.org/abs/2207.04672
54. Urailertprasert, N., Limkonchotiwat, P., Suwajanakorn, S., Nutanong, S.: SEA-
VQA: Southeast Asian cultural context dataset for visual question answering. In:
Gu, J., Fu, T.J.R., Hudson, D., Celikyilmaz, A., Wang, W. (eds.) Proceedings of
the 3rd Workshop on Advances in Language and Vision

arxiv.org/abs/2207.04672
54. Urailertprasert, N., Limkonchotiwat, P., Suwajanakorn, S., Nutanong, S.: SEA-
VQA: Southeast Asian cultural context dataset for visual question answering. In:
Gu, J., Fu, T.J.R., Hudson, D., Celikyilmaz, A., Wang, W. (eds.) Proceedings of
the 3rd Workshop on Advances in Language and Vision Research (ALVR). pp. 173–
185. Association for Computational Linguistics, Bangkok, Thailand (Aug 2024).
https://doi.org/10.18653/v1/2024.alvr-1.15, https://aclanthology.org/
2024.alvr-1.15/
55. Verma, S., Khanuja, M.S.U.R., Kumar, V., Murthy, R., Sen, J.: Milu: A multi-task
indic language understanding benchmark. arXiv preprint arXiv:2411.02538 (2025)

-- 24 of 35 --

Anthropogenic Regional Adaptation 25
56. Wang, H., Yao, Y., Liu, J., Zhang, X., Zhao, Y., Li, S., Liu, Z., Zhang, X.,
Zeng, Y.: Unsupervised cross-regional and cross-year adaptation by climate in-
dicator discrepancy for crop classification. Journal of Remote Sensing 5 (Jan
2025). https://doi.org/10.34133/remotesensing.0439, http://dx.doi.org/
10.34133/remotesensing.0439
57. Wang, J., Adelani, D.I., et al.: Afrimte and africomet: Enhancing comet to embrace
under-resourced african languages pp. 5997–6023 (2024)
58. Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J.,
Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J.,
Lin, J.: Qwen2-vl: Enhancing vision-language model’s perception of the world at
any resolution. arXiv preprint arXiv:2409.12191 (2024)
59. Wang, Y., Zang, Y., Li, H., Jin, C., Wang, J.: Unified reward model for multimodal
understanding and generation (2026), https://arxiv.org/abs/2503.05236
60. Winata, G.I., Aji, A.F., Cahyawijaya, S., Mahendra, R., Koto, F., Romadhony,
A., Kurniawan, K., Moeljadi, D., Prasojo, R.E., Fung, P., Baldwin, T., Lau, J.H.,
Sennrich, R., Ruder, S.: NusaX: Multilingual parallel sentiment dataset for 10
Indonesian local languages. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of

Winata, G.I., Aji, A.F., Cahyawijaya, S., Mahendra, R., Koto, F., Romadhony,
A., Kurniawan, K., Moeljadi, D., Prasojo, R.E., Fung, P., Baldwin, T., Lau, J.H.,
Sennrich, R., Ruder, S.: NusaX: Multilingual parallel sentiment dataset for 10
Indonesian local languages. In: Vlachos, A., Augenstein, I. (eds.) Proceedings of the
17th Conference of the European Chapter of the Association for Computational
Linguistics. pp. 815–834. Association for Computational Linguistics, Dubrovnik,
Croatia (May 2023). https://doi.org/10.18653/v1/2023.eacl-main.57, https:
//aclanthology.org/2023.eacl-main.57/
61. Winata, G.I., Hudi, F., Irawan, P.A., Anugraha, D., Putri, R.A., Yutong, W.,
Nohejl, A., Prathama, U.A., Ousidhoum, N., Amriani, A., et al.: Worldcuisines:
A massive-scale benchmark for multilingual and multicultural visual question an-
swering on global cuisines. In: Proceedings of the 2025 Conference of the Nations
of the Americas Chapter of the Association for Computational Linguistics: Human
Language Technologies (Volume 1: Long Papers). pp. 3242–3264 (2025)
62. Winata, G.I., et al.: Worldcuisines: A benchmark dataset for multilingual and
multicultural image classification. arXiv preprint arXiv:2405.14133 (2024)
63. Wu, X., Hao, Y., Sun, K., Chen, Y., Zhu, F., Zhao, R., Li, H.: Human preference
score v2: A solid benchmark for evaluating human preferences of text-to-image
synthesis (2023), https://arxiv.org/abs/2306.09341
64. Xu, J., Huang, Y., Cheng, J., Yang, Y., Xu, J., Wang, Y., Duan, W., Yang, S.,
Jin, Q., Li, S., Teng, J., Yang, Z., Zheng, W., Liu, X., Zhang, D., Ding, M.,
Zhang, X., Gu, X., Huang, S., Huang, M., Tang, J., Dong, Y.: Visionreward:
Fine-grained multi-dimensional human preference learning for image and video
generation (2026), https://arxiv.org/abs/2412.21059
65. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere-
ward: Learning and evaluating human preferences for text-to-image generation.
Advances in Neural Information

-grained multi-dimensional human preference learning for image and video
generation (2026), https://arxiv.org/abs/2412.21059
65. Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere-
ward: Learning and evaluating human preferences for text-to-image generation.
Advances in Neural Information Processing Systems 36, 15903–15935 (2023)
66. Yue, X., Song, Y., Asai, A., Kim, S., de Dieu Nyandwi, J., Khanuja, S., Kan-
tharuban, A., Sutawika, L., Ramamoorthy, S., Neubig, G.: Pangea: A fully open
multilingual multimodal llm for 39 languages. arXiv preprint arXiv:2410.16153
(2024), https://arxiv.org/abs/2410.16153
67. Zhang, Y., Li, M., Long, D., Zhang, X., Lin, H., Yang, B., Xie, P., Yang, A., Liu,
D., Lin, J., Huang, F., Zhou, J.: Qwen3 embedding: Advancing text embedding
and reranking through foundation models (2025), https://arxiv.org/abs/2506.
05176

-- 25 of 35 --

Appendix
A Assessment for SEA languages Translation Quality
Table 6: Human evaluation of English→5 SEA languages translation quality using a
five-point scale (mean ± standard deviation)
Language Model Grammar Naturalness
Burmese (mya)
NLLB 4.6 ± 0.8 4.1 ± 1.1
Gemma-3-27b 4.1 ± 1.2 3.6 ± 1.2
Gemini-2.5-flash 4.8 ± 0.6 4.6 ± 0.8
Gemini-2.5-pro 4.9 ± 0.4 4.9 ± 0.5
Thai (tha)
NLLB 3.3 ± 1.1 3.0 ± 1.2
Gemma-3-27b 4.2 ± 0.9 4.1 ± 0.9
Gemini-2.5-flash 4.8 ± 0.5 4.7 ± 0.5
Gemini-2.5-pro 4.7 ± 0.5 4.6 ± 0.5
Filipino (fil)
NLLB 4.4 ± 0.6 4.3 ± 0.9
Gemma-3-27b 4.8 ± 0.6 4.6 ± 0.8
Gemini-2.5-flash 4.7 ± 0.4 4.4 ± 0.7
Gemini-2.5-pro 4.8 ± 0.4 4.5 ± 0.6
Indonesian (ind)
NLLB 4.3 ± 0.9 3.8 ± 1.1
Gemma-3-27b 4.6 ± 0.6 4.2 ± 0.9
Gemini-2.5-flash 4.5 ± 0.6 4.1 ± 0.9
Gemini-2.5-pro 4.6 ± 0.5 4.3 ± 0.8
Vietnamese (vie)
NLLB 4.7 ± 0.5 4.4 ± 0.8
Gemma-3-27b 5 ± 0.1 4.9 ± 0.3
Gemini-2.5-flash 5 ± 0.1 4.9 ± 0.2
Gemini-2.5-pro 5 ± 0.1 5 ± 0.1
We conducted a systematic evaluation on 100 randomly selected human-
generated English captions from the SEA-VL [15] dataset for Burmese (mya),
Thai (tha), Filipino

± 0.5 4.3 ± 0.8
Vietnamese (vie)
NLLB 4.7 ± 0.5 4.4 ± 0.8
Gemma-3-27b 5 ± 0.1 4.9 ± 0.3
Gemini-2.5-flash 5 ± 0.1 4.9 ± 0.2
Gemini-2.5-pro 5 ± 0.1 5 ± 0.1
We conducted a systematic evaluation on 100 randomly selected human-
generated English captions from the SEA-VL [15] dataset for Burmese (mya),
Thai (tha), Filipino (fil), Indonesian (ind), and Vietnamese (vie). The evaluation
employed a five-point scale (1–5) assessing grammatical accuracy and natural-
ness, using NLLB [53], Gemma-3-27b [52], Gemini-2.5-flash, and Gemini-2.5-
pro [51] models. Table 6 presents the results of the translation quality evalu-
ation in five Southeast Asian languages using a five-point scale (mean ± stan-
dard deviation). Overall, the Gemini-2.5 models consistently achieve the highest
scores for Burmese (mya) and Thai (tha), with Gemini-2.5-Pro obtaining the
best performance for Burmese and Gemini-2.5-Flash performing slightly better
for Thai. For Filipino (fil), Indonesian (ind), and Vietnamese (vie), Gemma-
3-27B achieves competitive or top performance, often matching or exceeding

-- 26 of 35 --

Anthropogenic Regional Adaptation 27
the Gemini models in grammatical accuracy while maintaining high naturalness
scores. Based on these findings, Gemma-3-27b was used for Indonesian (ind),
Vietnamese (vie), Standard Malay (zsm), Filipino (fil), and Chinese (zho), where
it demonstrated strong and stable performance. Gemini-2.5-flash was used for
Thai (tha), Burmese (mya), Lao (lao), Khmer (khm), and Tamil (tam), as the
Gemini family showed superior robustness and higher translation quality for
these languages.
B Reward Model Ablation
We evaluated four reward models: HPSv2 [65], HPSv2 [63], VisionReward-Image
[64], and UnifiedReward [59]. In order to pick a reward model to use for data cura-
tion as described in section 5.1, we measured how much the reward models agreed
with human evaluations from SEA-VL [16]. For each image In, we normalized the
scores via min-max normalization and then

65], HPSv2 [63], VisionReward-Image
[64], and UnifiedReward [59]. In order to pick a reward model to use for data cura-
tion as described in section 5.1, we measured how much the reward models agreed
with human evaluations from SEA-VL [16]. For each image In, we normalized the
scores via min-max normalization and then took the average across categories to
get a human annotator score sn
human ∈ [0, 1]. Notably, the absolute score range of
ImageReward-style models varies across datasets, making direct normalization
and comparison with human ground-truth scores unreliable. Therefore, instead
of comparing absolute scores, we reformulated the evaluation as a pairwise pref-
erence task. Specifically, We randomly sampled 500 pairs of images Ia and Ib
such that sa
human > sb
human. For each reward model, we then compute the rate at
which it gives Ia a higher score than Ib i.e. sa
rm > sb
rm. Results are summarized in
table 7. Although VisionReward-Image achieved the highest pairwise agreement
with human judgments, we ultimately selected UnifiedReward by considering
both computational efficiency and predictive performance. While its accuracy
is slightly lower, UnifiedReward offers a more favorable trade-off overall, as it
supports efficient batched inference through vLLM [34], substantially improving
throughput for large-scale data filtering.
Table 7: Rate at which sa
rm > sb
rm for each reward model under consideration.
Reward Model Rate where sa
rm > sb
rm
ImageReward 	0.394
HPSv2 	0.384
VisionReward-Image 	0.466
UnifiedReward 	0.442
C Annotated Human Evaluation Test Sets
We conduct human evaluation on two datasets: SEA-VL VQA and SEA AYA.

-- 27 of 35 --

28 Cahyawijaya et al.
C.1 SEA-VL VQA
SEA-VL VQA contains ∼1.1k visual questions spanning 9 official languages in
Southeast Asia: Malay (zsm), Vietnamese (vie), Thai (tha), Indonesian (ind),
Filipino (fil), Tamil (tam), Khmer (khm), Chinese (cmn), and Burmese (mya).
The culturally grounded images were originally

35 --

28 Cahyawijaya et al.
C.1 SEA-VL VQA
SEA-VL VQA contains ∼1.1k visual questions spanning 9 official languages in
Southeast Asia: Malay (zsm), Vietnamese (vie), Thai (tha), Indonesian (ind),
Filipino (fil), Tamil (tam), Khmer (khm), Chinese (cmn), and Burmese (mya).
The culturally grounded images were originally collected via crowdsourcing by [16].
For each image, a native speaker authored a visual question in the target lan-
guage, which was subsequently reviewed by two additional native speakers, with
revisions made when needed to ensure quality.
To promote consistency and cultural relevance, we provided annotators with
detailed instructions for writing high-quality visual questions. Specifically, ques-
tions were required to: (1) be directly grounded in the visual content of the
image; (2) go beyond low-level perception and instead target culturally salient
elements, such as traditional attire, local cuisine, religious practices, architectural
styles, or social activities; (3) be clear, grammatically correct, and contextually
appropriate; and (4) be answerable solely from the image.
We also specified several categories of questions to avoid. These included: (1)
overly generic questions (e.g., "What is this?" or "What color is the shirt?"); (2)
purely factual questions lacking cultural grounding (e.g., "How many people are
in the image?"); (3) questions that cannot be answered from the image alone
(e.g., "What is her name?" unless explicitly shown); (4) questions that fail to
reflect the Southeast Asian context without clear justification; and (5) questions
that can be answered without the visual context.
C.2 SEA AYA
We construct SEA AYA by manually translating 135 visual questions from the
Aya Vision Benchmark [35] into 6 official Southeast Asian languages: Thai (tha),
Malay (zsm), Filipino (fil), Tamil (tam), Chinese (cmn), and Burmese (mya).
Combined with the English (eng) and Indonesian (ind) subsets already avail-
able in Aya Vision Benchmark, the

y manually translating 135 visual questions from the
Aya Vision Benchmark [35] into 6 official Southeast Asian languages: Thai (tha),
Malay (zsm), Filipino (fil), Tamil (tam), Chinese (cmn), and Burmese (mya).
Combined with the English (eng) and Indonesian (ind) subsets already avail-
able in Aya Vision Benchmark, the resulting benchmark contains ∼1.2k instances
covering 9 vision-language tasks: image captioning, chart and figure understand-
ing, image difference detection, general VQA, OCR, document understanding,
text transcription, visual reasoning, and screenshot-to-code generation. For each
example, annotators were first given a rough machine-translated draft as a start-
ing point. They were then asked to revise it to ensure that the final translation
is fluent, culturally appropriate, and semantically faithful to the English source.
For quality and consistency, our annotation guidelines emphasized several
principles: (1) preserve the meaning, tone, and intent of the original English
prompt; (2) produce text that is natural and fluent for native speakers of the
target language; (3) use grammar, vocabulary, and syntax that conform to stan-
dard usage in the target language; (4) match the level of formality of the source;
(5) respect language-specific cultural and pragmatic norms, including idiomatic
usage and politeness conventions; and (6) appropriately translate, retain, or nor-
malize named entities and technical terms when necessary.

-- 28 of 35 --

Anthropogenic Regional Adaptation 29
We also requested annotators to avoid common translation errors, includ-
ing semantic drift, omission of essential information, overly literal adherence to
machine-translated phrasing, literal translation of idiomatic expressions, over-
translation or under-translation, awkward or unidiomatic wording, and inconsis-
tent terminology, particularly across multi-sentence examples.
D Comparison of SEA-Gemma-3 and Other VLMs
We provide the full comparison per language on CVQA and

hine-translated phrasing, literal translation of idiomatic expressions, over-
translation or under-translation, awkward or unidiomatic wording, and inconsis-
tent terminology, particularly across multi-sentence examples.
D Comparison of SEA-Gemma-3 and Other VLMs
We provide the full comparison per language on CVQA and per region on SEA-
VQA in Table 8 and Table 9, respectively. The merged global-regional model
SEA-Gemma-3 10% achieves strongest performance on SEA regions and other
regions on CVQA, producing a solid performance even hgher compared to other
smaller and similar sized models like Llama-3.2 Vision [22], Pixtral [4], Gemma-
3 [52], and even outperform larger models such as Aya-Vision (32B) [19], Qwen-3
VL (35B) [8], Qwen-2.5-VL (72B) [9], and Molmo (72B) [20]. Similarly on SEA-
VQA, SEA-Gemma-3 10% also achieves the best average regional score, although
on several regions it is still outperformed by large models especially Aya-Vision
(32B) [19] which showcase strong performance on multiple SEA regions but falls
short on Indonesia, Thailand, and Laos.
E Results Breakdown
We provide the results breakdown in Table 10, Table 12, Table 11, and Table 15
for SEA-VLM; Table 13 for SEA-ImageGen, and Table 14 fo SEA-VL Embed-
ding.
F Sample Outputs of SEA-VLM and SEA-ImageGen
We provide samples generated by our best SEA-VLM and SEA-ImageGen com-
pared to the baselines on Figure 5 and Figure 6, respectively. As shown in Fig-
ure 6, the SEA-SDXL 25% model produces the most accurate representation
among others, while retaining the image naturalness and overall quality of the
original SDXL. For example, while the baselines also capture the component of
nasi lemask such as nasi putih (white rice) and telur (egg), the SEA-SDXL 25%
model incorporates more detailed factors including timun (cucumber), kacang
(peanut), ikan bilis (anchovy). Similarly for cultural and tradition categories,
SEA-SDXL 25% captures a more accurate illustration of playing angklung and
Hung king

nasi lemask such as nasi putih (white rice) and telur (egg), the SEA-SDXL 25%
model incorporates more detailed factors including timun (cucumber), kacang
(peanut), ikan bilis (anchovy). Similarly for cultural and tradition categories,
SEA-SDXL 25% captures a more accurate illustration of playing angklung and
Hung king festival in Phu Tho, Vietnam.
G List of Globalization Index by Region by Year
We provide the detailed per year per region globalization index which we used
for deciding the globalization factor α in GG-EZ in Table 16.

-- 29 of 35 --

30 Cahyawijaya et al.
Table 8: Comparison of our best model (SEA-Gemma-3 10%) againsts other competi-
tor models and the regionally-adapted Gemma-3 (SEA-Gemma-3) model on CVQA.
Llama-3.2 Pixtral Gemma-3 SEA-Gemma-3 SEA-Gemma-3 Aya-Vision Qwen-3 VL Qwen-2.5-VL Molmo
Language 	Vision (11B) (12B) (27B) 	(27B) 	10% (27B) 	(32B) 	(35B) 	(72B) 	(72B)
Indonesian 	56.31 	62.86 	66.50 	21.60 	66.26 	67.88 	66.26 	66.50 	63.83
Filipino 	51.72 	64.53 	66.01 	24.14 	70.94 	66.34 	64.04 	65.02 	64.53
Malay 	56.19 	61.90 	67.62 	29.84 	71.43 	72.38 	70.16 	62.50 	69.84
Chinese 	62.80 	68.93 	72.66 	40.73 	75.14 	77.48 	78.28 	80.98 	79.29
Tamil 	58.41 	51.87 	73.36 	40.65 	74.30 	61.68 	61.68 	58.88 	58.41
Javanese 	47.81 	51.18 	62.96 	22.22 	63.64 	55.56 	54.55 	55.22 	54.88
Minangkabau 	51.79 	51.39 	64.54 	23.90 	67.73 	61.75 	54.98 	51.79 	58.17
Sundanese 	44.00 	49.00 	59.00 	12.50 	63.50 	52.53 	55.50 	56.50 	52.00
SEA Avg. 	53.63 	57.71 	66.58 	26.95 	69.12 	64.45 	63.18 	62.17 	62.62
Amharic 	39.32 	32.91 	62.82 	26.07 	62.39 	29.18 	42.74 	36.48 	45.30
Bengali 	55.59 	48.25 	72.03 	33.57 	75.87 	64.31 	64.34 	61.27 	68.88
Breton 	34.57 	35.80 	42.96 	27.65 	45.68 	39.36 	34.81 	37.78 	35.06
Bulgarian 	49.06 	22.91 	62.26 	34.23 	64.42 	56.49 	54.45 	61.99 	57.68
Arabic 	49.26 	43.35 	62.07 	32.02 	65.02 	68.47 	57.64 	61.08 	58.62
Hindi 	68.16 	30.85 	80.60 	42.79 	81.09 	78.11 	75.62 	75.12 	78.11
Igbo 	44.00 	41.50 	43.50

1.27 	68.88
Breton 	34.57 	35.80 	42.96 	27.65 	45.68 	39.36 	34.81 	37.78 	35.06
Bulgarian 	49.06 	22.91 	62.26 	34.23 	64.42 	56.49 	54.45 	61.99 	57.68
Arabic 	49.26 	43.35 	62.07 	32.02 	65.02 	68.47 	57.64 	61.08 	58.62
Hindi 	68.16 	30.85 	80.60 	42.79 	81.09 	78.11 	75.62 	75.12 	78.11
Igbo 	44.00 	41.50 	43.50 	25.50 	47.00 	38.00 	34.50 	36.55 	41.50
Irish 	53.99 	57.67 	64.11 	31.29 	65.95 	56.13 	55.52 	57.98 	57.06
Japanese 	50.74 	49.26 	53.20 	29.06 	53.69 	59.11 	58.62 	58.62 	57.14
Kinyarwanda 	35.32 	34.47 	53.62 	28.51 	53.19 	40.43 	32.77 	38.30 	40.43
Korean 	59.66 	73.45 	76.90 	42.76 	77.93 	80.00 	77.24 	77.59 	74.14
Marathi 	48.02 	31.19 	74.75 	34.16 	77.23 	66.17 	63.86 	61.39 	68.81
Mongolian 	39.42 	39.74 	50.32 	23.40 	50.64 	36.01 	44.23 	39.10 	47.76
Norwegian 	54.52 	64.21 	66.89 	38.80 	69.57 	66.22 	61.20 	68.56 	69.90
Oromo 	34.11 	35.51 	45.79 	24.77 	47.20 	36.45 	38.32 	35.05 	42.06
Portuguese 	57.75 	73.59 	80.63 	41.20 	80.63 	78.01 	76.06 	76.76 	77.46
Romanian 	58.94 	67.88 	75.17 	40.73 	76.49 	74.17 	70.20 	75.83 	70.20
Russian 	66.50 	37.00 	80.50 	45.50 	81.50 	80.00 	76.50 	79.00 	84.00
Sinhala 	48.00 	28.44 	65.33 	29.78 	65.78 	39.56 	43.56 	45.50 	45.78
Spanish 	57.02 	69.13 	69.34 	37.61 	69.92 	73.25 	68.20 	71.44 	71.55
Swahili 	53.11 	60.07 	76.19 	36.26 	80.95 	66.18 	60.07 	55.31 	67.77
Telugu 	55.50 	32.50 	71.50 	32.00 	74.50 	57.79 	61.00 	58.50 	57.00
Urdu 	55.98 	48.02 	73.39 	36.93 	75.69 	66.69 	66.99 	64.92 	69.50
Other Avg. 	50.81 	45.99 	65.39 	33.68 	67.06 	58.70 	57.32 	58.01 	60.25
Global Avg. 	51.53 	49.01 	65.69 	31.94 	67.59 	60.18 	58.83 	59.08 	60.86
Table 9: Comparison of our best model (SEA-Gemma-3 10%) againsts other competi-
tor models and the regionally-adapted Gemma-3 (SEA-Gemma-3) model on SEA-VQA.
Gemma-3 SEA-Gemma-3 SEA-Gemma-3 Aya-Vision Qwen-3 VL
Region (27B) (27B) 10% (27B) (32B) (35B)
Cambodia 50.7 40.5 52.3 56.25 50.99
Indonesia 43.5 49.3 69.0 66.09 66.09
Laos 0.0 50.0 68.1

ur best model (SEA-Gemma-3 10%) againsts other competi-
tor models and the regionally-adapted Gemma-3 (SEA-Gemma-3) model on SEA-VQA.
Gemma-3 SEA-Gemma-3 SEA-Gemma-3 Aya-Vision Qwen-3 VL
Region (27B) (27B) 10% (27B) (32B) (35B)
Cambodia 50.7 40.5 52.3 56.25 50.99
Indonesia 43.5 49.3 69.0 66.09 66.09
Laos 0.0 50.0 68.1 45.83 63.89
Malaysia 37.5 44.4 62.4 69.31 60.32
Philippines 63.4 46.6 58.8 63.40 56.21
Singapore 56.3 46.9 68.8 71.88 65.63
Thailand 51.6 39.1 52.2 49.46 52.72
Vietnam 18.9 18.5 57.5 63.26 53.04
SEA Avg. 40.2 41.9 61.1 60.68 58.61

-- 30 of 35 --

Anthropogenic Regional Adaptation 31
Table 10: SEA-VLM Zero-shot performance on SEA-VQA. The result is broken down
per country. “Camb.” = Cambodia, “Indo.” = Indonesia, “Mala.” = Malaysia, ‘Phil.’ =
Phillipines, “Sing.”=Singapore, “Thai.” = Thailand, “Viet.” = Vietname
Avg. 	Country
Model Score Camb. Indo. Laos Mala. Phil. Sing. Thai. Viet.
Google Gemma-3 41.0 50.7 43.5 0.0 37.5 63.4 56.3 51.6 18.9
SEA-Gemma-3 5% 61.2 50.3 68.8 66.7 62.4 59.5 65.6 51.1 57.8
SEA-Gemma-3 10% 61.7 52.3 69.0 68.1 62.4 58.8 68.8 52.2 57.5
SEA-Gemma-3 50% 59.5 52.0 67.3 62.5 57.1 56.9 53.1 52.7 55.0
SEA-Gemma-3 70% 54.0 51.3 67.3 63.9 56.6 58.2 53.1 51.6 20.5
SEA-Gemma-3 (w/o merge) 41.9 40.5 49.3 50.0 44.4 46.6 46.9 39.1 18.5
Table 11: SEA-VLM Zero-shot performance on World Cuisine. (NC) = No Context,
(C) = Context, (Adv.) = Adversarial
Overall SEA 	By Prompt Type
Model 	Score Avg. Task 1 (NC) Task 2 Task 1 (C) Task 1 (Adv.)
Google Gemma-3 	59.8 60.1 64.8 48.3 78.5 47.5
SEA-Gemma-3 5% 	60.0 60.3 64.5 48.1 79.2 48.1
SEA-Gemma-3 10% 	60.0 60.2 64.5 48.0 79.1 48.1
SEA-Gemma-3 50% 	51.6 51.4 53.7 42.0 67.0 43.6
SEA-Gemma-3 70% 	51.9 52.6 53.5 43.5 67.1 43.4
SEA-Gemma-3 (w/o merge) 48.5 48.6 49.3 43.5 61.9 39.4
Table 12: SEA-VLM Zero-shot performance on CVQA.
Model Overall SEA Prompt Type
Score Avg. English Local Lang.
Gemma-3 67.2 67.83 67.96 66.44
SEA-Gemma-3 5% 68.72 69.45 69.39 68.05
SEA-Gemma-3 10% 68.76 69.45 69.42

43.6
SEA-Gemma-3 70% 	51.9 52.6 53.5 43.5 67.1 43.4
SEA-Gemma-3 (w/o merge) 48.5 48.6 49.3 43.5 61.9 39.4
Table 12: SEA-VLM Zero-shot performance on CVQA.
Model Overall SEA Prompt Type
Score Avg. English Local Lang.
Gemma-3 67.2 67.83 67.96 66.44
SEA-Gemma-3 5% 68.72 69.45 69.39 68.05
SEA-Gemma-3 10% 68.76 69.45 69.42 68.09
SEA-Gemma-3 50% 61.84 62.62 62.9 60.79
SEA-Gemma-3 70% 60.58 61.27 61.51 59.66
SEA-Gemma-3 (w/o merge) 35.58 36.16 38.0 33.16
Table 13: Detailed per aspect breakdown (Attribute, Entity, Other, and Relation) of
the DPGBench evaluation.
Attribute 	Entity Other Relation
Model color other shape size texture part state whole count text non-spatial spatial
StabilityAI SDXL 80.88 79.49 69.87 67.36 79.98 78.52 74.23 82.71 56.00 76.00 81.13 86.73
SEA-SDXL 25% 82.34 81.49 75.11 63.22 80.82 82.81 73.39 83.45 59.50 70.00 77.36 86.93
SEA-SDXL 50% 82.39 80.49 73.80 66.53 82.52 80.08 74.13 82.67 58.50 66.00 76.73 87.80
SEA-SDXL 75% 81.59 79.49 72.49 67.36 81.19 79.49 73.71 83.09 61.50 70.00 81.76 86.23
SEA-SDXL 100% 82.14 81.26 77.73 64.46 81.73 79.69 74.45 83.20 62.00 78.00 84.91 87.06

-- 31 of 35 --

32 Cahyawijaya et al.
Table 14: Detailed per country zero-shot performance of the SEAVQA evaluation.
Model 	SEAVQA Zero-Shot Performance
Overall Cambodia Indonesia Laos Malaysia Phillipines Singapore Thailand Vietnam
Google SigLIP2 25.81 28.9 25.0 22.2 24.3 28.8 21.9 25.5 25.6
SEA-SigLIP2 25% 26.36 27.6 25.4 29.2 29.1 28.8 28.1 26.1 24.0
SEA-SigLIP2 50% 28.06 31.3 27.0 22.2 29.6 27.5 21.9 31.0 27.2
SEA-SigLIP2 75% 29.66 28.6 30.1 19.4 32.3 34.0 25.0 29.9 28.8
SEA-SigLIP2 100% 28.96 28.3 28.6 29.2 26.5 32.7 31.3 30.4 29.1
Table 15: SEA-VLM Zero-shot performance on SEA-VQA. The result is broken down
per country. “Khm” = Cambodia, “Lao” = Laos, “Ind” = Indonesia, “Mys” = Malaysia,
‘Phl’ = Phillipines, “Sgp”=Singapore, “Th.” = Thailand, “Vie” = Vietnam. “CG” de-
notes CulturalGround, “WC” denotes WorldCuisine, while “OE” and “MC” denote open-
ended and

VLM Zero-shot performance on SEA-VQA. The result is broken down
per country. “Khm” = Cambodia, “Lao” = Laos, “Ind” = Indonesia, “Mys” = Malaysia,
‘Phl’ = Phillipines, “Sgp”=Singapore, “Th.” = Thailand, “Vie” = Vietnam. “CG” de-
notes CulturalGround, “WC” denotes WorldCuisine, while “OE” and “MC” denote open-
ended and multiple-choice, respectively.
Avg. Country
Model Score Khm Ind Lao Mys Phl Sgp Tha Vie
SEA-Mammoth 50k 13.0 9.9 18.8 11.1 11.1 11.8 18.8 14.67 2.88
SEA-Mammoth 250k 39.6 13.8 52.9 52.8 42.9 47.1 46.9 44.0 20.8
SEA-Mammoth 250k + CG OE 41.9 40.5 49.3 50.0 44.4 46.6 46.9 39.1 18.5
SEA-Mammoth 250k + CG MC 21.6 26.6 20.9 26.4 16.4 24.8 28.1 28.3 14.1
SEA-Mammoth 250k + WC 30.4 10.2 39.9 34.7 37.6 34.0 50.0 32.6 16.9

-- 32 of 35 --

Anthropogenic Regional Adaptation 33
Fig. 5: Generated responses using different model architypes. From left-to-right: global
model (Gemma-3), our regional model (SEA-Gemma-3), our merged model (SEA-
Gemma-3 10%) along with the prompts. Our model produces the most correct image
among others, while retaining the image naturalness and overall quality of the original
Gemma-3.

-- 33 of 35 --

34 Cahyawijaya et al.
Fig. 6: Generated image using different model architypes. From left-to-right: global
model (SDXL), our regional model (SEA-SDXL), our merged model (SEA-SDXL 25%),
and reference of natural images.

-- 34 of 35 --

Anthropogenic Regional Adaptation 35
Table 16: Globalization index by region by yea from 1993 to 2023. “EAP” = East
Asia and Pacific , ‘ECA” = Europe and Central Asia,“LAC” = Latin America and
Caribbean, “MENA” Middle East and North Africa, “NA” = North America, “SA” =
South Asia, “SEA” = Southeast Asian, “SSA” = Sub-Saharan Africa.
Year EAP ECA LAC MENA NA SA SEA SSA World
1993 49.32 58.82 51.20 57.00 68.59 28.05 36.04 34.37 48.86
1994 51.07 59.15 51.00 56.85 68.58 28.25 36.40 34.26 49.15
1995 51.40 59.72 51.67 56.62 68.66 28.39 37.34 34.13 49.42
1996 52.02 60.23 51.67 56.74 69.43 28.37 37.89 33.99

= Southeast Asian, “SSA” = Sub-Saharan Africa.
Year EAP ECA LAC MENA NA SA SEA SSA World
1993 49.32 58.82 51.20 57.00 68.59 28.05 36.04 34.37 48.86
1994 51.07 59.15 51.00 56.85 68.58 28.25 36.40 34.26 49.15
1995 51.40 59.72 51.67 56.62 68.66 28.39 37.34 34.13 49.42
1996 52.02 60.23 51.67 56.74 69.43 28.37 37.89 33.99 49.64
1997 51.68 60.43 51.75 56.79 69.48 28.35 37.76 33.43 49.53
1998 51.63 60.71 51.86 56.89 69.83 28.33 37.70 33.55 49.66
1999 52.00 61.00 52.20 56.94 70.64 28.26 38.59 33.45 49.85
2000 52.48 61.65 53.33 57.03 71.09 28.94 39.23 33.63 50.38
2001 52.61 62.06 53.88 57.43 71.74 28.96 39.12 33.80 50.70
2002 52.63 62.83 54.53 57.66 72.50 30.16 40.06 33.92 51.15
2003 52.75 63.55 55.32 58.04 73.02 30.84 39.76 34.82 51.79
2004 53.41 64.48 55.88 58.42 73.52 31.06 40.63 34.95 52.33
2005 53.93 65.47 56.19 58.48 74.17 32.29 41.61 35.24 52.86
2006 54.32 66.19 56.86 59.34 74.39 32.87 41.96 35.70 53.46
2007 55.16 67.67 58.04 59.78 75.24 33.99 42.90 36.40 54.46
2008 55.55 68.76 58.41 60.78 75.57 36.58 43.59 36.72 55.17
2009 55.70 68.42 58.59 61.54 75.41 37.98 44.27 37.16 55.38
2010 55.94 68.92 58.37 62.05 75.87 39.49 43.19 37.37 55.68
2011 56.40 70.04 58.50 62.51 76.43 40.59 44.17 38.87 56.53
2012 56.55 70.52 58.75 63.24 76.64 42.27 44.50 38.92 56.88
2013 56.50 70.93 59.06 63.55 77.25 42.87 45.33 39.14 57.16
2014 56.77 71.07 59.00 63.44 77.80 43.90 46.15 39.54 57.36
2015 57.13 70.32 58.76 62.50 77.86 44.75 45.79 38.99 56.98
2016 57.09 70.13 58.97 62.44 78.44 44.38 45.58 37.91 56.70
2017 57.04 70.88 59.00 60.91 78.81 44.14 45.12 37.71 56.75
2018 57.34 71.10 58.85 60.79 78.82 43.25 46.21 37.15 56.65
2019 56.36 70.84 58.51 60.92 78.95 40.53 44.29 36.54 56.12
2020 54.54 69.16 56.97 59.02 77.72 38.80 41.33 34.97 54.45
2021 53.94 69.60 57.23 59.66 78.02 37.97 39.87 34.86 54.54
2022 54.96 70.78 58.30 60.00 79.04 39.20 42.11 35.36 55.42
2023 55.74 71.03 58.91 60.13 79.36 40.06 43.40 35.46 55.79

-- 35 of 35 --