Pith. sign in

REVIEW 4 major objections 7 minor 40 references

Cross-Lingual Transfer of Cultural Knowledge: An Asymmetric Phenomenon

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that during continual pretraining, low-resource languages transfer cultural knowledge to English far more readily than English transfers it back into those languages.

desk verdict A genuinely controlled and transparent study of cross-lingual cultural knowledge transfer, with a plausible frequency-based explanation, but the key asymmetry claim rests on a gap that may not isolate co-occurrence as cleanly as advertised. read the letter →

arxiv 2506.01675 v1 pith:JSCIEXHG submitted 2025-06-02 cs.CL

Chen Zhang , Zhiyuan Liao , Yansong Feng This is my paper
classification cs.CL
keywords cross-lingualtransferculturalknowledgecontinualpretraininglow-resourcelanguagesfrequencyhypothesislanguageadaptationmultilingualmodelsdensity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper studies how cultural knowledge moves between languages when a language model is adapted to a new language, and it finds an asymmetry: for low-resource languages, knowledge flows mainly from the non-English language into English, with little flowing back. The authors build a transparent setup in which a 0.5-billion-parameter model is pretrained on English Wikipedia and then continually pretrained on Korean, Chinese, Tibetan, or Mongolian, with pairs of parallel sentences either concatenated as bridges or kept apart. The accuracy gap between these two settings measures cross-lingual transfer, and it is systematically larger for Tibetan and Mongolian into English than from English into those languages. They propose that corpus frequency drives this: cultural facts that appear often in a language's pretraining data are more likely to cross over, and their corpus analysis supports the idea for both directions.

What carries the argument

The load-bearing apparatus is a two-setting continual pretraining comparison. A base model is pretrained from scratch on English Wikipedia with non-Latin characters stripped out, then continually pretrained on a non-English corpus; in the with-bridges setting each pair of parallel sentences is concatenated into a single document, while in the without-bridges setting the same sentences are shuffled apart so they never co-occur. The gap in cloze-question accuracy between the two settings estimates how much knowledge actually crosses languages, isolated from improved language proficiency. The accompanying cultural density metric, estimated by retrieving documents and judging entailment of each cultural item, connects this transfer gap to corpus frequency.

What would settle it

Find several cultural knowledge items that appear very frequently in a low-resource corpus yet fail to transfer to English in the bridged setting, or items that transfer despite near-zero frequency; either would undermine the frequency explanation. A direct controlled version would train the same model on corpora that differ only in how often one fact appears and test whether the transfer gap tracks that frequency.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-lingual transfer of cultural knowledge during language adaptation is asymmetric for low-resource languages: continual pretraining on Tibetan or Mongolian transfers a detectable amount of knowledge about those cultures into the model's English answers, while the same training transfers very little English-only cultural knowledge into the low-resource language. For high-resource languages, Korean and Chinese, transfer is bidirectional and of similar magnitude. The authors attribute the asymmetry to a frequency effect: cultural knowledge that occurs more often in a pretraining corpus transfers more easily across languages. They support this with a cultural density measure, the count of knowledge-bearing documents normalized by corpus size, showing that low-resource cultures are an order of magnitude denser in their own-language corpus than in English, and that successfully transferred knowledge items appear above average frequency in the source corpus.

Load-bearing premise

The whole measurement rests on the premise that shuffling parallel sentences apart truly prevents the model from learning any cross-lingual alignment, so that the accuracy gap between settings is purely cross-lingual transfer.

Editorial extensions

If this is right

  • If the frequency hypothesis holds, adding more culturally relevant text in a low-resource language to its corpus should increase transfer of that culture into English.
  • For high-resource languages, the paper's results imply that English and the target language can mutually reinforce each other's cultural knowledge during continual pretraining.
  • For low-resource languages, English-to-minority-language cultural knowledge is not delivered by language adaptation alone, so culturally aware models for those communities will need other mechanisms or data.
  • The bridge/no-bridge gap offers a reusable diagnostic: any observed improvement in a target language after multilingual training can be checked for whether it comes from transfer or from proficiency gains.
  • The Anglophonic-culture control experiment indicates that the asymmetry is not an artifact of non-English cultures alone: high-frequency English knowledge transfers into non-English languages once bridges are provided.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If frequency is the driving factor, the same asymmetry should appear for non-cultural factual knowledge whenever source and target corpora differ in frequency, which could be tested with constructed random facts.
  • The direction of transfer may compound existing representational biases: knowledge that is rare in English will remain rare in English-centric models even after language adaptation, so low-resource communities bear the cost of making their culture legible to English rather than the reverse.
  • One testable extension is to hold a fixed cultural fact and vary its frequency in the non-English corpus while keeping the model and bridge setup constant; the transfer gap should scale monotonically with frequency if the hypothesis is causal rather than correlational.
  • The use of 0.5B models leaves open whether larger models show stronger or weaker asymmetry; if larger models align languages better, the low-resource-to-English direction might become even more dominant.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper studies how cultural knowledge transfers across languages during continual pretraining. The authors pretrain a 0.5B Qwen-2.5 model from scratch on Latin-script-filtered English Wikipedia, then continually pretrain on Korean, Chinese, Tibetan, and Mongolian using two settings: one concatenates parallel sentence pairs ('with bridges') to encourage cross-lingual transfer, and the other treats the same sentences as independent shuffled documents ('without bridges'). They evaluate on bilingual cloze-style cultural questions and interpret the accuracy gap between the two settings as the amount of cross-lingual transfer. They report bidirectional transfer for Korean and Chinese, but asymmetric transfer for Tibetan and Mongolian, where knowledge flows more strongly from the low-resource language to English than in the reverse direction. They propose a frequency-based hypothesis—cultural knowledge that appears more often in the pretraining corpus transfers more easily—and support it with retrieval-based estimates of cultural density in the English and non-English corpora.

Significance. If the findings hold, the paper provides a transparent, controlled framework for studying cross-lingual knowledge transfer, together with a new evaluation resource for the understudied Tibetan and Mongolian cultures. The frequency hypothesis is a crisp, falsifiable claim that connects corpus statistics to transfer outcomes, and the authors make code and data publicly available. At the same time, the empirical support is currently limited by the identification strategy behind the bridge/no-bridge contrast and by the absence of uncertainty quantification, so the strength of the conclusions exceeds what the experiments can currently justify.

major comments (4)
  1. [Section 2, Figure 2, Figure 3 (second row)] The without-bridges condition is not a pure 'no-transfer' baseline: it still trains on the English half of every parallel sentence pair as an independent document. Therefore, accuracy on English-version questions in the without-bridges runs can improve from direct exposure to the English translations, and the with-bridges/without-bridges gap in panels 1b-4b measures only the marginal effect of co-occurrence, not the total flow of knowledge from the non-English language to English. The abstract's claim that low-resource languages 'primarily transfer knowledge to English' is read off this gap, so it is conditional on the assumption that co-occurrence is the only operative mechanism. A condition that removes the English half of the parallel corpus, or a separate analysis of the without-bridges English-version accuracy relative to a monolingual baseline, is needed to support the 'flow into English' claim. In addition, the without-bridges setting still trains on both languages, so implicit cross-lingual alignment through shared tokenization or optimization is not fully excluded; the paper should justify or empirically test that co-occurrence is the only meaningful difference.
  2. [Section 2, Figure 2] The bridged documents are concatenations of parallel sentence pairs, making them approximately twice as long as the unbridged documents, which split each pair into separate documents. Because training uses fixed-size token batches and a fixed number of steps, the two settings differ in document-length distribution as well as in the presence of cross-lingual bridges. The accuracy gap between settings therefore cannot be attributed solely to cross-lingual alignment; a matched-length control (e.g., concatenating non-parallel sentences or using length-matched documents) is required to isolate the effect of parallel co-occurrence.
  3. [Figure 3, Section 3] No error bars, confidence intervals, or significance tests are reported, and no multiple-seed variation is described. The differences that drive the asymmetry claim are often on the order of one to three accuracy points (e.g., panels 3a and 4a show the with-bridge and without-bridge curves nearly overlapping), and the qualitative labels 'notable' versus 'less evident' are assigned by visual inspection. A statistical comparison of the gaps between directions and across languages is needed before the central asymmetry can be considered established; at minimum, the authors should report the variance across random seeds or a bootstrap confidence interval for each gap.
  4. [Section 4, Table 1, Appendix C.4] The frequency hypothesis is validated on the same bridge/no-bridge contrast that defines 'successfully transferred' instances, and the culture-level density comparison is dominated by corpus-size differences: normalizing by document count makes the low-resource non-English densities large partly because those corpora are much smaller than the English corpus. The retrieval pipeline for Tibetan and Mongolian relies on BM25 plus NLLB translation and Llama-based entailment, which may be noisy for low-resource languages. The instance-level counts (9.0 vs 4.2 and 4.7 vs 2.2) are described as 'significantly higher' but no statistical test is provided. A predictive or out-of-sample test of the frequency hypothesis (e.g., whether within-language frequency ordering predicts transfer item by item) would be considerably more convincing than the current correlational analysis.
minor comments (7)
  1. [Appendix B vs Table 3] The text says 552 questions were collected for Mongols, but Table 3 lists 276 for Mongolian; please reconcile the counts.
  2. [Appendix C.1] The typo 'Deepseed' should be 'DeepSpeed'.
  3. [Appendix C.3] The phrase 'high-pressure languages' should be 'high-resource languages'.
  4. [Appendix D] The word 'untrivial' should be 'nontrivial'.
  5. [References] The reference 'and National Ethnic Affairs Commission. 2009' appears to have a stray leading 'and'.
  6. [Figure 3] The y-axis ranges differ across panels; a common scale or explicit annotations of the gap sizes would make cross-panel comparisons easier to assess.
  7. [Section 4] The statement that transferred knowledge instances occur 'significantly higher' than average is not backed by a test statistic; please report the relevant distributions or a bootstrap interval.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: transfer measure and frequency hypothesis are operationally distinct; remaining concerns are experimental-validity issues, not definitional reductions.

full rationale

The paper's derivation chain is empirical rather than definitional. Cross-lingual transfer is operationalized as the accuracy gap between with-bridge and without-bridge continual pretraining (Section 2), while the frequency hypothesis (Section 4) is a separate correlational claim tested against corpus densities and instance-level occurrence counts. No equation defines transfer in terms of frequency, and no fitted parameter is renamed as a prediction. The only self-citation (Zhang et al. 2024, MC2) supplies pretraining corpora for Tibetan and Mongolian and is not invoked to justify the paper's conclusions or to exclude alternatives. The with-bridge versus without-bridge contrast does carry potential confounds (e.g., document length differences and direct exposure to English parallel sentences in the no-bridge condition), and the low-resource asymmetry may partly reflect these design choices; however, those are threats to construct validity, not circular reductions where an output is identical to an input by construction. The instance-level 'successfully transferred' criterion and the frequency comparisons are logically independent: selection depends on accuracy patterns, while frequency is measured from the corpora. Thus no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework relies on several domain assumptions about the experimental setup and measurement pipeline. No new theoretical entities are introduced; the contributions are empirical and methodological.

free parameters (3)
  • success_transfer_criterion = consistent correct on last 3 checkpoints
    Hand-chosen threshold in Appendix C.4 for labeling instances as transferred; it affects the instance-level frequency comparison.
  • retrieval_top_k = 50
    Number of documents retrieved per question in Section 4 setup; affects density estimates.
  • cultural_density_normalization = count / total documents
    Normalizing by document count rather than tokens is a design choice that could affect cross-linguistic density comparisons.
assumptions (4)
  • domain assumption The performance gap between with-bridges and without-bridges continual pretraining isolates cross-lingual transfer effects.
    Section 2 states the gap serves as an indicator; assumes no other factors differ between the two settings.
  • domain assumption Removing non-Latin characters from English and non-Latin characters from target corpora prevents lexical overlap and cross-lingual transfer through shared tokens.
    Section 2 relies on this to control lexical overlap.
  • domain assumption Retrieval and entailment with Llama-3.1-70B accurately estimate whether a document contains the cultural knowledge.
    Section 4 and Appendix C.3; retrieval errors could bias density estimates.
  • domain assumption Findings from a 0.5B model generalize to larger LLMs.
    Acknowledged in Limitations; the study uses 0.5B models only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Lingual Transfer of Cultural Knowledge: An Asymmetric Phenomenon." pith.science (2026). https://pith.science/paper/JSCIEXHG

@misc{pith2026250601675,
  author       = {Pith},
  title        = {Pith review of: Cross-Lingual Transfer of Cultural Knowledge: An Asymmetric Phenomenon},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JSCIEXHG}},
  note         = {Machine review of arXiv:2506.01675}
}
read the original abstract

Despite substantial research efforts evaluating how well large language models~(LLMs) handle global cultural diversity, the mechanisms behind their cultural knowledge acquisition, particularly in multilingual settings, remain unclear. We study this question by investigating how cultural knowledge transfers across languages during language adaptation of LLMs. We introduce an interpretable framework for studying this transfer, ensuring training data transparency and controlling transfer effects. Through a study of four non-Anglophonic cultures, we observe bidirectional cultural transfer between English and other high-resource languages, while low-resource languages primarily transfer knowledge to English with limited reverse flow. To explain this asymmetric phenomenon, we propose a frequency-based hypothesis: cultural knowledge appearing more frequently in the pretraining data transfers more easily, which is supported by empirical analysis of the training corpora.

Figures

Figures reproduced from arXiv: 2506.01675 by the authors.

Figure 1
Figure 1. LLMs might exhibit different patterns of cross [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our framework for studying the cross-lingual transfer of cultural knowledge. We use the transfer between [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Accuracy on the non-Anglophonic cultural questions with different continual pretraining (CT) steps under [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 9 canonical work pages

  1. [1]

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report. arXiv preprint arXiv:2305.10403

    arXiv 2023

  2. [2]

    Terra Blevins and Luke Zettlemoyer. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.233 Language contamination helps explains the cross-lingual capabilities of E nglish pretrained models . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3563--3574, Abu Dhabi, United Arab Emirates. Association for Computat...

    doi:10.18653/v1/2022.emnlp-main.233 2022

  3. [3]

    Hoyeon Chang, Jinho Park, Seonghyeon Ye, Sohee Yang, Youngkyung Seo, Du-Seong Chang, and Minjoon Seo. 2024. https://openreview.net/forum?id=TYdzj1EvBP How do large language models acquire factual knowledge during pretraining? In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page 2024

  4. [4]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. https://arxiv.org/abs/2402.03216 Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . Preprint, arXiv:2402.03216

    arXiv 2024

  5. [5]

    John Dang, Shivalika Singh, Daniel D'souza, Arash Ahmadian, Alejandro Salamanca, Madeline Smith, Aidan Peppin, Sungjin Hong, Manoj Govindassamy, Terrence Zhao, et al. 2024. Aya expanse: Combining research breakthroughs for a new multilingual frontier. arXiv preprint arXiv:2412.04261

    arXiv 2024

  6. [6]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv 2024

  7. [7]

    Philipp Dufter and Hinrich Sch \"u tze. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.358 Identifying elements essential for BERT `s multilinguality . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4423--4437, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.emnlp-main.358 2020

  8. [8]

    Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lopez de Lacalle, and Mikel Artetxe. 2024. Bertaqa: How much do language models know about local culture? arXiv preprint arXiv:2406.07302

    work page Pith review arXiv 2024

Show all 40 references
  1. [9]

    Yi Fung, Ruining Zhao, Jae Doo, Chenkai Sun, and Heng Ji. 2024. Massively multi-cultural knowledge acquisition & lm benchmarking. arXiv preprint arXiv:2402.09369

    2024 arXiv

  2. [10]

    Daniel Hershcovich, Stella Frank, Heather Lent, Miryam de Lhoneux, Mostafa Abdou, Stephanie Brandl, Emanuele Bugliarello, Laura Cabello Piqueras, Ilias Chalkidis, Ruixiang Cui, Constanza Fierro, Katerina Margatina, Phillip Rust, and Anders S gaard. 2022. https://doi.org/10.186...

    2022 doi

  3. [11]

    Peng Hu, Sizhe Liu, Changjiang Gao, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. 2024. Large language models are cross-lingual knowledge-free reasoners. arXiv preprint arXiv:2406.16655

    2024 arXiv

  4. [12]

    Kaiyu Huang, Fengran Mo, Hongliang Li, You Li, Yuanchi Zhang, Weijian Yi, Yulong Mao, Jinchen Liu, Yuzhuang Xu, Jinan Xu, et al. 2024. A survey on large language models with multilingualism: Recent advances and new frontiers. arXiv preprint arXiv:2405.10936

    2024 arXiv

  5. [13]

    Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning, pages 15696--15707. PMLR

    2023

  6. [14]

    Nora Kassner, Philipp Dufter, and Hinrich Sch \"u tze. 2021. https://doi.org/10.18653/v1/2021.eacl-main.284 Multilingual LAMA : Investigating knowledge in multilingual pretrained language models . In Proceedings of the 16th Conference of the European Chapter of the Association...

    2021 doi

  7. [15]

    Eunsu Kim, Juyoung Suk, Philhoon Oh, Haneul Yoo, James Thorne, and Alice Oh. 2024. https://aclanthology.org/2024.lrec-main.296/ CLI c K : A benchmark dataset of cultural and linguistic intelligence in K orean . In Proceedings of the 2024 Joint International Conference on Compu...

    2024

  8. [16]

    Huihan Li, Arnav Goel, Keyu He, and Xiang Ren. 2025. https://openreview.net/forum?id=XrsOu4KgDE Attributing culture-conditioned generations to pretraining corpora . In The Thirteenth International Conference on Learning Representations

    2025

  9. [17]

    Huihan Li, Liwei Jiang, Nouha Dziri, Xiang Ren, and Yejin Choi. 2024. https://openreview.net/forum?id=DbsLm2KAqP CULTURE - GEN : Revealing global cultural perception in language models through natural language prompting . In First Conference on Language Modeling

    2024

  10. [18]

    Peiqin Lin, Andr \'e FT Martins, and Hinrich Sch \"u tze. 2024. A recipe of parallel corpora exploitation for multilingual large language models. arXiv preprint arXiv:2407.00436

    2024 arXiv

  11. [19]

    Chen Cecilia Liu, Iryna Gurevych, and Anna Korhonen. 2024. Culturally aware and adapted nlp: A taxonomy and a survey of the state of the art. arXiv preprint arXiv:2406.03930

    2024 arXiv

  12. [20]

    Chi-Liang Liu, Tsung-Yuan Hsu, Yung-Sung Chuang, and Hung-Yi Lee. 2020. A study of cross-lingual ability and language-specific information in multilingual bert. arXiv preprint arXiv:2004.09205

    2020 arXiv

  13. [21]

    and National Ethnic Affairs Commission. 2009. Ethnic Minorities in China (Revised Edition). 5. People's Publishing House

    2009

  14. [22]

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

    2024

  15. [23]

    Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Augenstein. 2024. Survey of cultural awareness in language models: Text and beyond. arXiv preprint arXiv:2411.00860

    2024 arXiv

  16. [24]

    Siddhesh Pawar, Junyeong Park, Jiho Jin, Arnav Arora, Junho Myung, Srishti Yadav, Faiz Ghifari Haznitrama, Inhwa Song, Alice Oh, and Isabelle Augenstein. 2025. https://doi.org/10.1162/COLI.a.14 Survey of cultural awareness in language models: Text and beyond . Computational Li...

    2025 doi

  17. [25]

    Fred Philippy, Siwen Guo, and Shohreh Haddadan. 2023. https://doi.org/10.18653/v1/2023.acl-long.323 Towards a common understanding of contributing factors for cross-lingual transfer in multilingual language models: A review . In Proceedings of the 61st Annual Meeting of the As...

    2023 doi

  18. [26]

    Jirui Qi, Raquel Fern \'a ndez, and Arianna Bisazza. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.658 Cross-lingual consistency of factual knowledge in multilingual language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing...

    2023 doi

  19. [27]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  20. [28]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  21. [29]

    Weiyan Shi, Ryan Li, Yutong Zhang, Caleb Ziems, Raya Horesh, Rog \'e rio Abreu de Paula, Diyi Yang, et al. 2024. Culturebank: An online community-driven knowledge base towards culturally aware language technologies. arXiv preprint arXiv:2404.15238

    2024 arXiv

  22. [30]

    Xinyi Wang, Antonis Antoniades, Yanai Elazar, Alfonso Amayuelas, Alon Albalak, Kexun Zhang, and William Yang Wang. 2025. https://openreview.net/forum?id=IQxBDLmVpT Generalization v.s. memorization: Tracing language models capabilities back to pretraining data . In The Thirteen...

    2025

  23. [31]

    Xiangpeng Wei, Haoran Wei, Huan Lin, Tianhao Li, Pei Zhang, Xingzhang Ren, Mei Li, Yu Wan, Zhiwei Cao, Binbin Xie, et al. 2023. Polylm: An open source polyglot large language model. arXiv preprint arXiv:2307.06018

    2023 arXiv

  24. [32]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. https://arxiv.org/abs/2309.07597 C-pack: Packaged resources to advance general chinese embedding . Preprint, arXiv:2309.07597

    2023 arXiv

  25. [33]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  26. [34]

    Jiacheng Ye, Xijia Tao, and Lingpeng Kong. 2023. Language versatilists vs. specialists: An empirical revisiting on multilingual transfer ability. arXiv preprint arXiv:2306.06688

    2023 arXiv

  27. [35]

    Da Yin, Hritik Bansal, Masoud Monajatipoor, Liunian Harold Li, and Kai-Wei Chang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.132 G eo MLAMA : Geo-diverse commonsense probing on multilingual pre-trained language models . In Proceedings of the 2022 Conference on Empirical...

    2022 doi

  28. [36]

    Fei Yuan, Yinquan Lu, Wenhao Zhu, Lingpeng Kong, Lei Li, Yu Qiao, and Jingjing Xu. 2023. https://doi.org/10.18653/v1/2023.findings-acl.731 L ego- MT : Learning detachable models for massively multilingual machine translation . In Findings of the Association for Computational L...

    2023 doi

  29. [37]

    Chen Zhang, Mingxu Tao, Quzhe Huang, Jiuheng Lin, Zhibin Chen, and Yansong Feng. 2024. https://doi.org/10.18653/v1/2024.acl-long.479 MC ^2 : Towards transparent and culturally-aware NLP for minority languages in C hina . In Proceedings of the 62nd Annual Meeting of the Associa...

    2024 doi

  30. [38]

    Xin Zhao, Naoki Yoshinaga, and Daisuke Oba. 2024. Tracing the roots of facts in multilingual language models: Independent, shared, and transferred knowledge. arXiv preprint arXiv:2403.05189

    2024 arXiv

  31. [39]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.