pith. sign in

arxiv: 2606.01800 · v1 · pith:CUGZPUNUnew · submitted 2026-06-01 · 💻 cs.CL · cs.AI· cs.LG

Multilinguality of Large Language Models From a Structural Perspective

Haruki Sakajo , Yusuke Sakai , Hidetaka Kamigaito , Taro Watanabe This is my paper

Pith reviewed 2026-06-28 14:54 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords multilingualitylarge language modelsstructural analysislow-resource languagesrepresentational analysispost-trainingcross-lingual relationshipstoken representations
0
0 comments X

The pith

Low-resource languages are structurally more different from English in LLMs than high- and mid-resource languages, and language-specific post-training changes their structures while keeping inter-language relationships intact.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shifts focus from token-level representations to a structural view of how LLMs handle multiple languages. It applies representational structural analysis to show that languages with fewer resources in training data sit farther from English in this structural space. The analysis also tracks what happens after language-specific post-training and finds that individual language structures shift but the overall map of relationships between languages stays stable. A sympathetic reader would care because English dominates training data yet the models still acquire usable non-English capabilities; understanding the structural basis could clarify why some languages benefit more than others.

Core claim

Through representational structural analysis of LLMs, low-resource languages prove structurally more different from English than high- and mid-resource languages, while language-specific post-training alters the structures of the affected languages without disrupting the preserved relationships among languages.

What carries the argument

representational structural analysis applied to token representations across languages in LLMs

If this is right

  • High-resource languages maintain closer structural alignment to English even when English dominates the pre-training corpus.
  • Post-training on one language can be applied without collapsing the cross-lingual structural map that the model learned earlier.
  • Structural distance correlates with resource level, suggesting that low-resource languages may need targeted interventions to reduce their separation from the dominant language structure.
  • The preservation of inter-language relationships after post-training implies that the model retains a shared multilingual backbone even as individual languages are adjusted.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If structural distance predicts downstream performance gaps, then measuring it early could guide which languages need extra data or alignment steps.
  • The finding that post-training perturbs local structures while leaving global relationships fixed may extend to other adaptation techniques such as continued pre-training or instruction tuning.
  • Alternative structural probes that do not rely on representation geometry could be tested to see whether the resource-based pattern holds under different definitions of structure.

Load-bearing premise

The structural differences detected in the representations reflect genuine properties of the languages rather than artifacts produced by the chosen analysis method or tokenization choices.

What would settle it

Re-running the same structural comparison on the identical models but with an alternative distance measure, such as one derived from syntactic parse trees instead of representation geometry, and finding no greater distance for low-resource languages would undermine the central claim.

Figures

Figures reproduced from arXiv: 2606.01800 by Haruki Sakajo, Hidetaka Kamigaito, Taro Watanabe, Yusuke Sakai.

Figure 1
Figure 1. Figure 1: Representational structure distance and cosine [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our research. We construct tree structures that summarize inter-token relationships in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cosine similarity on pre-trained models. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Syntactic and typological distance. Brighter [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: TED between English and other languages. The error bars indicate the 95 % confidence interval. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: TED between Chinese and other languages. The error bars indicate the 95 % confidence interval. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Heatmaps of TED in a middle layer for each language pair. Brighter colors represent a smaller distance, [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: TED between English and other languages for larger models. The error bars indicate the 95 % confidence [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: TED for post-trained models. The error bars indicate the 95 % confidence interval [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Heatmaps of TED in a middle layer of pre- and post-trained models for each language pair. [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Cosine similarity on post-trained models. [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: L2 distance on pre-trained models 0 5 10 15 20 25 30 35 Layer 0 100 200 300 400 500 600 700 L2 Distance Qwen3 8B 0 5 10 15 20 25 30 35 Layer 0 200 400 600 800 L2 Distance Qwen3 Swallow 8B CPT English - German English - Japanese English - Chinese English - Indonesian English - Guarani English - Jinghpaw [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: L2 distance on post-trained models 14 [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Heatmaps of TED for each language pair. EnglishGerman Japanese Chinese Indonesian Aranese Guarani Jinghpaw English German Japanese Chinese Indonesian Aranese Guarani Jinghpaw 0.60 0.63 0.59 0.60 0.64 0.67 0.69 0.60 0.64 0.61 0.63 0.64 0.65 0.67 0.63 0.64 0.59 0.62 0.62 0.62 0.63 0.59 0.61 0.59 0.62 0.64 0.66 0.67 0.60 0.63 0.62 0.62 0.63 0.62 0.64 0.64 0.64 0.62 0.64 0.63 0.63 0.63 0.67 0.65 0.62 0.66 0.6… view at source ↗
Figure 15
Figure 15. Figure 15: Heatmaps of TED in pre- and post-trained models. [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Heatmaps of TED in larger models. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_16.png] view at source ↗
read the original abstract

Large language models (LLMs) have excelled in processing multiple languages through pre- and post-training on multilingual data, even though English dominates the training data. Prior work focusing on token representations has revealed how those LLMs process non-English text. Although these analyses have provided insightful findings, they fail to capture a structural view, which is an inherent property of language. In this study, we explore the multilinguality of LLMs through representational structural analysis. Our findings reveal that low-resource languages are structurally more different from English than high- and mid-resource languages, and that language-specific post-training alters their structures while preserving inter-language relationships.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper claims that representational structural analysis of LLMs reveals low-resource languages to be structurally more different from English than high- and mid-resource languages, and that language-specific post-training alters these structures while preserving inter-language relationships. The work positions this structural view as an advance over prior token-representation analyses.

Significance. If the structural findings prove robust to confounds, the work would supply a useful complement to token-level studies of multilingual LLMs, potentially guiding post-training strategies that respect language-specific geometry while maintaining cross-lingual alignments. The emphasis on resource-level structural gaps is a concrete, testable contribution.

major comments (3)
  1. [§4] §4 (Experimental Setup / Representational Structural Analysis): No controls are described for tokenization quality, subword fragmentation, or sequence-length matching across resource tiers. Because low-resource languages are tokenized more poorly by English-dominant vocabularies, any observed geometric differences could arise from embedding sparsity or length effects rather than syntax or semantics; this directly undermines the claim that the analysis isolates 'inherent structural properties.'
  2. [§5] §5 (Results): The statement that post-training 'alters their structures while preserving inter-language relationships' is presented without quantitative metrics (e.g., before/after correlation of pairwise distances or Procrustes alignment scores). Without these numbers it is impossible to judge whether the preservation is substantive or merely qualitative.
  3. [Abstract / §3] Abstract and §3: The manuscript supplies no error analysis, statistical tests comparing resource-level groups, or ablation on the distance metric itself. The central claim therefore rests on unvalidated differences whose magnitude and reliability cannot be assessed from the reported material.
minor comments (2)
  1. [Methods] Notation for the structural metric (e.g., what distance is computed on which hidden states) is introduced without a formal definition or reference to prior work on representational similarity analysis.
  2. [Figures] Figure captions do not state the exact number of languages per resource bucket or the model variants used, reducing reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments. We address each of the major concerns below and commit to revisions that will strengthen the empirical support for our claims.

read point-by-point responses

  1. Referee: [§4] §4 (Experimental Setup / Representational Structural Analysis): No controls are described for tokenization quality, subword fragmentation, or sequence-length matching across resource tiers. Because low-resource languages are tokenized more poorly by English-dominant vocabularies, any observed geometric differences could arise from embedding sparsity or length effects rather than syntax or semantics; this directly undermines the claim that the analysis isolates 'inherent structural properties.'

    Authors: We agree that the lack of controls for tokenization and sequence length represents a potential confound in our current analysis. The manuscript does not describe such controls, and we will revise §4 to include them. Specifically, we will add experiments that match sequence lengths across languages and control for subword fragmentation by reporting results on a subset of languages with similar tokenization efficiency. This will help confirm that the observed structural differences are not solely due to these factors. revision: yes

  2. Referee: [§5] §5 (Results): The statement that post-training 'alters their structures while preserving inter-language relationships' is presented without quantitative metrics (e.g., before/after correlation of pairwise distances or Procrustes alignment scores). Without these numbers it is impossible to judge whether the preservation is substantive or merely qualitative.

    Authors: The referee correctly identifies that our claim about preservation of inter-language relationships after post-training lacks quantitative backing in the current manuscript. We will add the suggested metrics in the revised §5, including before-and-after Procrustes alignment scores and correlations of pairwise distance matrices to provide a quantitative assessment of how much the relationships are preserved. revision: yes

  3. Referee: [Abstract / §3] Abstract and §3: The manuscript supplies no error analysis, statistical tests comparing resource-level groups, or ablation on the distance metric itself. The central claim therefore rests on unvalidated differences whose magnitude and reliability cannot be assessed from the reported material.

    Authors: We acknowledge the absence of statistical tests, error analysis, and metric ablations in the manuscript. To address this, we will incorporate these elements into the revised version: statistical significance tests for differences between resource tiers, error bars on the reported distances, and an ablation study varying the distance metric to show that the main findings hold across reasonable choices. revision: yes

Circularity Check

0 steps flagged

No circularity detected from available text

full rationale

The abstract presents empirical findings from representational structural analysis on LLM hidden states without any equations, fitted parameters, self-citations, or derivation steps that reduce to inputs by construction. No load-bearing claims rely on renaming, ansatz smuggling, or uniqueness theorems from prior author work. The central observation (low-resource languages structurally more different) is stated as an analysis outcome rather than a tautology or statistically forced prediction, making the paper self-contained against external benchmarks in the provided material.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no details on free parameters, axioms, or invented entities are provided.

pith-pipeline@v0.9.1-grok · 5639 in / 907 out tokens · 22842 ms · 2026-06-28T14:54:12.049377+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 27 canonical work pages

  1. [1]

    Proceedings of the 37th International Conference on Machine Learning , url =

    Xiong, Ruibin and Yang, Yunchang and He, Di and Zheng, Kai and Zheng, Shuxin and Xing, Chen and Zhang, Huishuai and Lan, Yanyan and Wang, Liwei and Liu, Tie-Yan , title =. Proceedings of the 37th International Conference on Machine Learning , url =. 2020 , publisher =

    2020

  2. [2]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention Is All You Need , booktitle =. 2017 , month = dec, series =

    2017

  3. [3]

    2021 , urldate =

    Elhage, Nelson and Nanda, Neel and Olsson, Catherine and Henighan, Tom and Joseph, Nicholas and Mann, Ben and Askell, Amanda and Bai, Yuntao and Chen, Anna and Conerly, Tom and DasSarma, Nova and Drain, Dawn and Ganguli, Deep and Hatfield-Dodds, Zac and Hernandez, Danny and Jones, Andy and Kernion, Jackson and Lovitt, Liane and Ndousse, Kamal and Amodei, ...

    2021

  4. [4]

    2025 , eprint=

    Diversity of Transformer Layers: One Aspect of Parameter Scaling Laws , author=. 2025 , eprint=

    2025

  5. [5]

    2026 , eprint=

    StructLens: A Structural Lens for Language Models via Maximum Spanning Trees , author=. 2026 , eprint=

    2026

  6. [6]

    What Language Do Non- E nglish-Centric Large Language Models Think in?

    Zhong, Chengzhi and Liu, Qianying and Cheng, Fei and Jiang, Junfeng and Wan, Zhen and Chu, Chenhui and Murawaki, Yugo and Kurohashi, Sadao. What Language Do Non- E nglish-Centric Large Language Models Think in?. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.1350

    work page doi:10.18653/v1/2025.findings-acl.1350 2025

  7. [7]

    Do Llamas Work in E nglish? On the Latent Language of Multilingual Transformers

    Wendler, Chris and Veselovsky, Veniamin and Monea, Giovanni and West, Robert. Do Llamas Work in E nglish? On the Latent Language of Multilingual Transformers. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.820

    work page doi:10.18653/v1/2024.acl-long.820 2024

  8. [8]

    2020 , month = Aug, urldate =

    Interpreting. 2020 , month = Aug, urldate =

    2020

  9. [9]

    Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages

    Brinkmann, Jannik and Wendler, Chris and Bartelt, Christian and Mueller, Aaron. Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1:...

    work page doi:10.18653/v1/2025.naacl-long.312 2025

  10. [10]

    The Thirteenth International Conference on Learning Representations , year=

    The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling , author=. The Thirteenth International Conference on Learning Representations , year=

  11. [11]

    and Tu, Zhuowen and Bergen, Benjamin K

    Chang, Tyler A. and Tu, Zhuowen and Bergen, Benjamin K. The Geometry of Multilingual Language Model Representations. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.9

    work page doi:10.18653/v1/2022.emnlp-main.9 2022

  12. [12]

    Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of

    Fran. Trans-Tokenization and Cross-lingual Vocabulary Transfers: Language Adaptation of. First Conference on Language Modeling , year=

  13. [13]

    Dictionaries to the Rescue: Cross-Lingual Vocabulary Transfer for Low-Resource Languages Using Bilingual Dictionaries

    Sakajo, Haruki and Ide, Yusuke and Vasselli, Justin and Sakai, Yusuke and Tian, Yingtao and Kamigaito, Hidetaka and Watanabe, Taro. Dictionaries to the Rescue: Cross-Lingual Vocabulary Transfer for Low-Resource Languages Using Bilingual Dictionaries. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.1333

    work page doi:10.18653/v1/2025.findings-acl.1333 2025

  14. [14]

    Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models

    Tang, Tianyi and Luo, Wenyang and Huang, Haoyang and Zhang, Dongdong and Wang, Xiaolei and Zhao, Xin and Wei, Furu and Wen, Ji-Rong. Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1...

    work page doi:10.18653/v1/2024.acl-long.309 2024

  15. [15]

    Tarjan, R. E. , year =. Finding optimum branchings , volume =. Networks , publisher =. doi:10.1002/net.3230070103 , number =

    work page doi:10.1002/net.3230070103

  16. [16]

    Simple Fast Algorithms for the Editing Distance Between Trees and Related Problems , volume =

    Zhang, Kaizhong and Shasha, Dennis , year =. Simple Fast Algorithms for the Editing Distance Between Trees and Related Problems , volume =. SIAM J. Comput. , doi =

  17. [17]

    Tree edit distance: Robust and memory-efficient , journal =

    Mateusz Pawlik and Nikolaus Augsten , keywords =. Tree edit distance: Robust and memory-efficient , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.is.2015.08.004 , url =

    work page doi:10.1016/j.is.2015.08.004 2016

  18. [18]

    2025 , eprint=

    Gemma 3 Technical Report , author=. 2025 , eprint=

    2025

  19. [19]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  20. [20]

    2025 , eprint=

    Olmo 3 , author=. 2025 , eprint=

    2025

  21. [21]

    Continual Pre-Training for Cross-Lingual

    Kazuki Fujii and Taishi Nakamura and Mengsay Loem and Hiroki Iida and Masanari Ohi and Kakeru Hattori and Hirai Shota and Sakae Mizuki and Rio Yokota and Naoaki Okazaki , booktitle=. Continual Pre-Training for Cross-Lingual. 2024 , url=

    2024

  22. [22]

    NLLB Team and Costa-juss \`a , Marta R. and Cross, James and C elebi, Onur and Elbayad, Maha and Heafield, Kenneth and Heffernan, Kevin and Kalbassi, Elahe and Lam, Janice and Licht, Daniel and Maillard, Jean and Sun, Anna and Wang, Skyler and Wenzek, Guillaume and Youngblood, Al and Akula, Bapi and Barrault, Loic and Gonzalez, Gabriel Mejia and Hansanti,...

    work page doi:10.1038/s41586-024-07335-x 2024

  23. [23]

    Languages Still Left Behind: Toward a Better Multilingual Machine Translation Benchmark

    Taguchi, Chihiro and Mai, Seng and Kurabe, Keita and Sakai, Yusuke and Agyei, Georgina and Eslami, Soudabeh and Chiang, David. Languages Still Left Behind: Toward a Better Multilingual Machine Translation Benchmark. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.1018

    work page doi:10.18653/v1/2025.emnlp-main.1018 2025

  24. [24]

    2020 , isbn =

    Kurabe, Keita , title =. 2020 , isbn =. doi:https://doi.org/10.15026/125363

    work page doi:10.15026/125363 2020

  25. [25]

    Kurabe, Keita , title =. 2020

    2020

  26. [26]

    doi:https://dx.doi.org/10.4225/72/59888e8ab2122 , year=

    Kachin folktales told in. doi:https://dx.doi.org/10.4225/72/59888e8ab2122 , year=

    work page doi:10.4225/72/59888e8ab2122

  27. [27]

    Kachin culture and history told in

    Kurabe, Keita , year=. Kachin culture and history told in

  28. [28]

    doi: 10.18653/v1/2020.emnlp-demos.6

    Wolf, Thomas and Debut, Lysandre and Sanh, Victor and Chaumond, Julien and Delangue, Clement and Moi, Anthony and Cistac, Pierric and Rault, Tim and Louf, Remi and Funtowicz, Morgan and Davison, Joe and Shleifer, Sam and von Platen, Patrick and Ma, Clara and Jernite, Yacine and Plu, Julien and Xu, Canwen and Le Scao, Teven and Gugger, Sylvain and Drame, M...

    work page doi:10.18653/v1/2020.emnlp-demos.6 2020

  29. [29]

    Bleu: a method for automatic evaluation of machine translation

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing. B leu: a Method for Automatic Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. 2002. doi:10.3115/1073083.1073135

    work page doi:10.3115/1073083.1073135 2002

  30. [30]

    (2018) ’ A Call for Clarity in Reporting BLEU Scores’ in Proceedings of the Third Conference on Machine Translation

    Post, Matt. A Call for Clarity in Reporting BLEU Scores. Proceedings of the Third Conference on Machine Translation: Research Papers. 2018. doi:10.18653/v1/W18-6319

    work page doi:10.18653/v1/w18-6319 2018

  31. [31]

    chr F : character n-gram F -score for automatic MT evaluation

    Popovi \'c , Maja. chr F : character n-gram F -score for automatic MT evaluation. Proceedings of the Tenth Workshop on Statistical Machine Translation. 2015. doi:10.18653/v1/W15-3049

    work page doi:10.18653/v1/w15-3049 2015

  32. [32]

    A Study of Translation Edit Rate with Targeted Human Annotation

    Snover, Matthew and Dorr, Bonnie and Schwartz, Rich and Micciulla, Linnea and Makhoul, John. A Study of Translation Edit Rate with Targeted Human Annotation. Proceedings of the 7th Conference of the Association for Machine Translation in the Americas: Technical Papers. 2006

    2006

  33. [33]

    and Khiu, Eric and Do

    Khan, Aditya and Shipton, Mason and Anugraha, David and Duan, Kaiyao and Hoang, Phuong H. and Khiu, Eric and Do. URIEL +: Enhancing Linguistic Inclusion and Usability in a Typological and Multilingual Knowledge Base. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  34. [34]

    Representational Isomorphism and Alignment of Multilingual Large Language Models

    Wu, Di and Lei, Yibin and Yates, Andrew and Monz, Christof. Representational Isomorphism and Alignment of Multilingual Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.823

    work page doi:10.18653/v1/2024.findings-emnlp.823 2024

  35. [35]

    2024 , eprint=

    GPT-4 Technical Report , author=. 2024 , eprint=

    2024

  36. [36]

    WECHSEL : Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models

    Minixhofer, Benjamin and Paischer, Fabian and Rekabsaz, Navid. WECHSEL : Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.293

    work page doi:10.18653/v1/2022.naacl-main.293 2022

  37. [37]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Zero-Shot Tokenizer Transfer , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  38. [38]

    An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference

    Yamaguchi, Atsuki and Villavicencio, Aline and Aletras, Nikolaos. An Empirical Study on Cross-lingual Vocabulary Adaptation for Efficient Language Model Inference. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.396

    work page doi:10.18653/v1/2024.findings-emnlp.396 2024

  39. [39]

    Computational Linguistics , volume =

    Yamaguchi, Atsuki and Villavicencio, Aline and Aletras, Nikolaos , title =. Computational Linguistics , volume =. 2026 , month =. doi:10.1162/COLI.a.581 , url =

    work page doi:10.1162/coli.a.581 2026

  40. [40]

    Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis

    Zhu, Wenhao and Liu, Hongyi and Dong, Qingxiu and Xu, Jingjing and Huang, Shujian and Kong, Lingpeng and Chen, Jiajun and Li, Lei. Multilingual Machine Translation with Large Language Models: Empirical Results and Analysis. Findings of the Association for Computational Linguistics: NAACL 2024. 2024. doi:10.18653/v1/2024.findings-naacl.176

    work page doi:10.18653/v1/2024.findings-naacl.176 2024

  41. [41]

    2024 , url=

    Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan , booktitle=. 2024 , url=

    2024

  42. [42]

    Linguistic Typology

    Song, Jae Jung. Linguistic Typology

  43. [43]

    Language, Usage and Cognition , publisher=

    Bybee, Joan , year=. Language, Usage and Cognition , publisher=

  44. [44]

    2013 , doi =

    WALS Online (v2020.4) , type =. 2013 , doi =

    2013

  45. [45]

    doi:10.14324/111.9781787352872 , publisher =

    A Grammar of Paraguayan Guarani , author =. doi:10.14324/111.9781787352872 , publisher =

    work page doi:10.14324/111.9781787352872

  46. [46]

    2015 , publisher=

    Gramatica basica der occitan aran. 2015 , publisher=

    2015

  47. [47]

    A Reproducibility Study on Quantifying Language Similarity: The Impact of Missing Values in the URIEL Knowledge Base

    Toossi, Hasti and Huai, Guo and Liu, Jinyu and Khiu, Eric and Do. A Reproducibility Study on Quantifying Language Similarity: The Impact of Missing Values in the URIEL Knowledge Base. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 4: Student Research Wo...

    work page doi:10.18653/v1/2024.naacl-srw.25 2024

  48. [48]

    Mazumder, Rahul and Hastie, Trevor and Tibshirani, Robert , title =. J. Mach. Learn. Res. , month = aug, pages =. 2010 , issue_date =

    2010

  49. [49]

    The Eleventh International Conference on Learning Representations , year=

    Language models are multilingual chain-of-thought reasoners , author=. The Eleventh International Conference on Learning Representations , year=

  50. [50]

    Language Models are Few-shot Multilingual Learners

    Winata, Genta Indra and Madotto, Andrea and Lin, Zhaojiang and Liu, Rosanne and Yosinski, Jason and Fung, Pascale. Language Models are Few-shot Multilingual Learners. Proceedings of the 1st Workshop on Multilingual Representation Learning. 2021. doi:10.18653/v1/2021.mrl-1.1

    work page doi:10.18653/v1/2021.mrl-1.1 2021

  51. [51]

    Extending Multilingual BERT to Low-Resource Languages

    Wang, Zihan and K, Karthikeyan and Mayhew, Stephen and Roth, Dan. Extending Multilingual BERT to Low-Resource Languages. Findings of the Association for Computational Linguistics: EMNLP 2020. 2020. doi:10.18653/v1/2020.findings-emnlp.240

    work page doi:10.18653/v1/2020.findings-emnlp.240 2020

  52. [52]

    Global MMLU : Understanding and Addressing Cultural and Linguistic Biases in Multilingual Evaluation

    Singh, Shivalika and Romanou, Angelika and Fourrier, Cl \'e mentine and Adelani, David Ifeoluwa and Ngui, Jian Gang and Vila-Suero, Daniel and Limkonchotiwat, Peerat and Marchisio, Kelly and Leong, Wei Qi and Susanto, Yosephine and Ng, Raymond and Longpre, Shayne and Ruder, Sebastian and Ko, Wei-Yin and Bosselut, Antoine and Oh, Alice and Martins, Andre a...

    work page doi:10.18653/v1/2025.acl-long.919 2025

  53. [53]

    MMLU - P ro X : A Multilingual Benchmark for Advanced Large Language Model Evaluation

    Xuan, Weihao and Yang, Rui and Qi, Heli and Zeng, Qingcheng and Xiao, Yunze and Feng, Aosong and Liu, Dairui and Xing, Yun and Wang, Junjue and Gao, Fan and Lu, Jinghui and Jiang, Yuang and Li, Huitao and Li, Xin and Yu, Kunyu and Dong, Ruihai and Gu, Shangding and Li, Yuekang and Xie, Xiaofei and Juefei-Xu, Felix and Khomh, Foutse and Yoshie, Osamu and C...

    work page doi:10.18653/v1/2025.emnlp-main.79 2025