Pith. sign in

REVIEW 3 major objections 7 minor 41 references

It's All About In-Context Learning! Teaching Extremely Low-Resource Languages to LLMs

T0 review · 3 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read For languages whose scripts are nearly absent from an LLM, zero-shot prompting with translations beats parameter-efficient fine-tuning.

desk verdict A broad, careful empirical study of ICL vs PEFT for very low-resource languages, but the word-level alignment results for the three rarest scripts lean on a test-set leakage the paper itself concedes; the clean sentence-level results still carry the main conclusion. read the letter →

arxiv 2508.19089 v1 pith:2XQESUJA submitted 2025-08-26 cs.CL

Yue Li , Zhixue Zhao , Carolina Scarton This is my paper
classification cs.CL
keywords in-contextlearningextremelylow-resourcelanguagesrarescriptsparameter-efficientfine-tuninglanguagealignmentmultilingualLLMstokenizationefficiencytopicclassification
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 sets out to determine whether LLMs can be taught extremely low-resource languages on the fly, without fine-tuning, by placing translations inside the prompt. Across 20 under-represented languages and three open multilingual LLMs, it finds a division: when both the language and its script are nearly absent from the model's tokenizer and pretraining data, parameter-efficient fine-tuning overfits early and barely improves accuracy, while zero-shot in-context learning with word- or sentence-level alignment produces large gains. For languages the model already partially represents, few-shot prompting and fine-tuning remain the better tools. The result gives practitioners a data-cheap path for unseen-script languages: invest in small parallel translation resources instead of annotation and fine-tuning compute.

What carries the argument

The load-bearing mechanism is prompt-level language alignment in zero-shot in-context learning: word-level alignment appends a per-word English gloss to every word of the target text, and sentence-level alignment prepends one or more unlabeled parallel target-English sentences selected by BM25. Two diagnostics carry the analysis: Information Parity (the ratio of target-language to English negative log-likelihood under the model) measures how much pretraining ability exists for the language, and token-to-byte ratio measures whether the tokenizer splits the script into raw bytes. These two numbers predict both when fine-tuning will collapse and when alignment will help.

What would settle it

Take the SIB-200 test inputs for Santali and Nko, have native speakers build a gold word-level dictionary, and rerun zero-shot ICL with alignment. If the reported 0.4–0.6 accuracy gains disappear with the verified dictionary, the effect is an artifact of the automatic dictionaries; if they persist, the alignment mechanism itself is confirmed.

Watch

Extended reading notes

Core claim

The central discovery is a boundary on how to adapt LLMs to languages they have essentially never seen. Using Information Parity and token-to-byte ratio to measure how well a model already handles a language, the paper shows that IA3-style fine-tuning fails precisely when a language and its script are severely under-represented: the tokenizer falls back to raw bytes and the model overfits at an early stage. In that regime, zero-shot ICL with alignment — appending a dictionary translation for each word, or one or more parallel target-English sentences retrieved by BM25 — yields substantial accuracy gains, often exceeding few-shot ICL and sometimes exceeding vocabulary extension through contin

Load-bearing premise

The large word-level alignment gains rest on dictionaries produced automatically by NLLB or fast_align, whose accuracy the authors could not check because no native speakers or gold translations were available.

Editorial extensions

If this is right

  • Practitioners should not fine-tune a model on a language whose script is nearly absent from the tokenizer and pretraining data; the expected gain is minimal or negative.
  • For such languages, creating a small set of in-domain parallel sentences or a word dictionary for zero-shot ICL is more effective than labelling data for fine-tuning.
  • Few-shot ICL and PEFT remain the right tools for low-resource languages the LLM already partially supports, roughly those with baseline accuracy above 0.45.
  • Zero-shot ICL with alignment can outperform vocabulary extension via continued pretraining on multilingual PLMs, without additional pretraining compute.
  • Alignment gains depend on semantic similarity: BM25-selected parallel examples help, while random examples often hurt.

Reading between the lines

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

  • Beyond the paper, the IP/TBR correlation suggests a cheap screening test: measure a candidate language's token-to-byte ratio before choosing between fine-tuning and prompting, since byte-level tokenization is a strong warning sign.
  • If the dictionary-quality caveat is resolved, word-level alignment could become a practical zero-resource recipe for any unseen script, because it needs only a wordlist rather than parallel sentences.
  • On longer and harder tasks such as reading comprehension, word-level glosses are not enough; the paper's own results imply sentence or passage alignment with more parallel data would be needed there.
  • The comparison with vocabulary extension is only indirect, so a direct head-to-head under equal compute would clarify whether prompting truly replaces continued pretraining or merely supplies a cheaper shortcut.
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

3 major / 7 minor

Summary. The paper asks whether LLMs can acquire extremely low-resource languages purely through in-context learning (ICL), with or without alignment signals, and how this compares with parameter-efficient fine-tuning (PEFT). It evaluates 20 languages from SIB-200 topic classification and 11 from BELEBELE reading comprehension using DeepSeek-7B, LLaMA-3.2-3B, and Gemma-2-2B. The authors report that IA3 fine-tuning fails for languages whose script and language are severely under-represented in the tokenizer and pretraining data, while zero-shot ICL with sentence- or word-level alignment yields large gains; for better-represented languages, few-shot ICL or PEFT is generally preferable. They also analyze token-to-byte ratio and information parity as explanatory factors. The paper argues that, for the weakest languages, zero-shot ICL with alignment can surpass vocabulary-extension-based continued pretraining on multilingual PLMs.

Significance. If the main claims hold, the paper provides practically useful guidance for adapting LLMs to very low-resource languages and is one of the first systematic evaluations across multiple LLMs and rare scripts. The study has genuine strengths: it uses public benchmarks, reports full per-language results in the appendix, conducts a large number of controlled experiments, and includes a candid acknowledgment in §4.4 footnote 10 that the fast_align-based word-level results may leak gold English translations. The clean sentence-level BM25 results already support the core ICL conclusion, so the paper is not built on a single fragile effect. However, the word-level alignment evidence for the three most extreme languages (nqo, sat, min) is contaminated by leakage, and the PEFT conclusion rests on a single adapter method, so the headline claims need to be re-scoped or re-supported.

major comments (3)
  1. [§4.2.2, Fig. 5, Table 5, Appendix B; cf. §4.4 fn. 10] The word-level dictionaries for nqo, sat, and min are constructed by training fast_align on SIB-200 training data and then aligning English and target-language words in the test set (Appendix B). This means that for each test sentence, the English glosses placed in the prompt are derived from the gold English translation of that exact sentence. Footnote 10 in §4.4 acknowledges this and excludes fast_align results from the zero- versus few-shot comparison, but Figure 5, Table 5, and the 'around 0.6 of accuracy improvement' claim in §4.2.2 still rely on these numbers (e.g., nqo 0.617/0.523/0.696; sat word-translation 0.754). These word-level results are therefore not valid evidence for the effectiveness of word-level alignment. Please recompute the word-level results for these languages using dictionaries built only from the training split or an independent lexicon, or remove the leaked nu
  2. [§3 and §4.1 (first bullet of §1)] The paper uses IA3 as the sole representative of PEFT and reports preliminary LoRA/DoRA results only as motivation. The abstract and the first main finding generalize to 'small-scaled fine-tuning' and 'PEFT'. Because the central contrast with prior work is that fine-tuning is ineffective for extremely under-represented scripts, the claim should be scoped to IA3, or additional results with at least one other adapter family (e.g., LoRA) on a subset of languages should be reported to show the effect is not adapter-specific.
  3. [§4.4, Table 4] The comparison between zero-shot ICL with alignment and vocabulary-extended XLM-R uses numbers from Adelani et al. (2024) rather than re-runs under the same evaluation protocol. The wording 'potentially surpassing' is hedged, but the table and surrounding text present the comparison as evidence. Differences in model architecture, size, fine-tuning method, and prompt template could explain part of the gap. Please state explicitly that this is a cross-paper reference, not a controlled comparison, or run a matched XLM-R experiment on the same languages and splits.
minor comments (7)
  1. [§4.2.2] Typo: 'world-level alignment' should be 'word-level alignment' in the sentence comparing word-level alignment and word-level translation.
  2. [Limitations] Typo: 'we were not able to accurately access the quality' should be 'assess the quality'.
  3. [Tables 3 and 4 captions] The statement 'Differences between baseline zero-shot ICL is statistical significant' is grammatically awkward and, more importantly, gives no test statistic, p-values, or multiple-comparison correction. With 20 languages × 3 models, correction is essential before claiming significance.
  4. [Table 5] The number of shots k for the few-shot ICL columns is not specified in the caption or table. BELEBELE Table 3 states 3-shot; SIB-200 Table 5 should do the same.
  5. [Table 2] For nqo, sat, and min the NLLB chrf++ columns are empty because NLLB does not support these languages; indicate 'not available' explicitly and state how dictionary quality for the fast_align dictionaries was assessed, since the Limitations say it could not be verified.
  6. [§4.2.1] The point-biserial correlation between tokenizer parity and benefit from additional parallel examples is described only as 'statistically significant'; please report the correlation coefficient, sample size, and p-value.
  7. [§4.4 / Fig. 6] The categories 'extremely low', 'low (acc<0.45)', and 'low (acc>0.45)' are defined post hoc and partly overlap; clarify the exact inclusion criteria and state whether these thresholds were pre-registered or are descriptive.

Circularity Check

1 steps flagged · score 6.0 of 10

Fast_align 'dictionary' for nqo/sat/min is built from the gold English side of the test set, so the word-level ICL gains for these languages are partially self-constructed; the central claim still has independent support from sentence-level alignment.

  1. self definitional [Section 3 (Word-level alignment), Appendix B (Dictionary), Section 4.4 footnote 10]
    "For languages not supported by NLLB (nqo, sat, and min), we train the word alignment tool fast_align (Dyer et al., 2013) to simulate a high-quality dictionary (See Appendix B). ... For the three languages that are not supported by NLLB, we train the word alignment tool fast_align (Dyer et al., 2013) with SIB-200 training data and then align the English words and target-language words in the test set. ... Results based on fast_align are excluded, as it potentially leaks gold standard English translations into the prompt."

    For nqo, sat, and min, the 'dictionary' used for word-level alignment is not a fixed resource: it is produced by aligning the English and target-language sides of the SIB-200 test set, i.e., using the gold English translation of each exact test sentence. The prompt then tells the model 'w_tgt means w_eng in English' for the very input it must classify. Thus the large word-level and word-translation gains reported in Table 5 and Figure 5 for these languages (e.g., nqo 0.617/0.523/0.696; sat word-translation 0.754) are constructed by exposing the model to the English content of the gold test sentence; they are not an independent measure of acquiring the language. The authors acknowledge this leak in footnote 10 and exclude those results from the zero- vs few-shot comparison, but Figure 5, Ta

full rationale

The paper is largely an empirical evaluation with no fitted equations or parametric derivation, and its self-citations (e.g., Kurz et al. 2024, Schlicht et al. 2025, Zhao and Aletras 2024) are not load-bearing; they are related-work references. The IP/TBR analyses are explanatory correlations rather than quantities that construct the outcome. However, the word-level alignment and word-translation results for nqo, sat, and min are partially circular by construction: the fast_align 'dictionary' is aligned on the test set's gold English side, so the prompt leaks the English content of the exact sentence being classified. The sentence-level BM25 results and the NLLB-based word-level results for other languages provide independent support for the general ICL conclusion, which prevents a higher score. But because Figure 5, Table 5, and the nqo vocabulary-extension comparison use the leaked word-level numbers, one or more of the paper's headline predictions reduce to the test-set input, giving a score of 6.

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

The central claims rest on four domain assumptions: IP is a valid proxy for capability, TBR measures tokenization quality, the automatically created dictionaries are accurate, and the benchmarks are representative. No free parameters or invented entities are introduced.

assumptions (4)
  • domain assumption Information Parity (IP) measured with the evaluated LLMs is a valid proxy for the model's pre-training competence in a language.
    Used in Section 3.1 to select target languages and in Section 4.1 to explain fine-tuning outcomes.
  • domain assumption Token-to-Byte Ratio (TBR) is a valid indicator of tokenization quality for rare scripts.
    Used in Section 4.1 to attribute fine-tuning failures to byte-level tokenization.
  • domain assumption The word-alignment dictionaries created by NLLB or fast_align on SIB-200 training data are sufficiently accurate to support the observed gains in zero-shot ICL.
    Used in Section 3 and 4.2.2; the paper itself flags inability to verify dictionary quality in Limitations.
  • domain assumption SIB-200 and BELEBELE are representative benchmarks for NLU in these languages.
    Used throughout; datasets are from prior work and cover only topic classification and reading comprehension.

how reviews work

0 comments
Cite this review

Pith. "Pith review of It's All About In-Context Learning! Teaching Extremely Low-Resource Languages to LLMs." pith.science (2026). https://pith.science/paper/2XQESUJA

@misc{pith2026250819089,
  author       = {Pith},
  title        = {Pith review of: It's All About In-Context Learning! Teaching Extremely Low-Resource Languages to LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XQESUJA}},
  note         = {Machine review of arXiv:2508.19089}
}
read the original abstract

Extremely low-resource languages, especially those written in rare scripts, as shown in Figure 1, remain largely unsupported by large language models (LLMs). This is due in part to compounding factors such as the lack of training data. This paper delivers the first comprehensive analysis of whether LLMs can acquire such languages purely via in-context learning (ICL), with or without auxiliary alignment signals, and how these methods compare to parameter-efficient fine-tuning (PEFT). We systematically evaluate 20 under-represented languages across three state-of-the-art multilingual LLMs. Our findings highlight the limitation of PEFT when both language and its script are extremely under-represented by the LLM. In contrast, zero-shot ICL with language alignment is impressively effective on extremely low-resource languages, while few-shot ICL or PEFT is more beneficial for languages relatively better represented by LLMs. For LLM practitioners working on extremely low-resource languages, we summarise guidelines grounded by our results on adapting LLMs to low-resource languages, e.g., avoiding fine-tuning a multilingual model on languages of unseen scripts.

Figures

Figures reproduced from arXiv: 2508.19089 by the authors.

Figure 1
Figure 1. Regional distribution of the languages stud [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Accuracy improvement from baseline zero-shot ICL (denoted as [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Correlation between information parity (IP), token-to-byte ratio (TBR) and accuracy score after fine-tuning. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Accuracy scores for DeepSeek in zero-shot [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Accuracy scores for LLaMA-3.2 over low￾resource languages in zero-shot ICL with word-level alignment (gray) or word-level translation (orange) set￾tings. Red denotes baseline zero-shot ICL. The zero-shot ICL performance with word-level alignment or translation on LLaMA…
Figure 6
Figure 6. Figure 6: Accuracy comparison among baseline (red), PEFT (green), best zero-shot ICL (blue) and best few￾shot ICL (black) on LLaMA-3.2. Languages are cate￾gorised into: (a) Both language and script are severely under-represented (names in red in Figure 3b, base￾line accuracy < 0…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 12 canonical work pages

  1. [1]

    Alabi, Yanke Mao, Haonan Gao, and En-Shiun Annie Lee

    David Ifeoluwa Adelani, Hannah Liu, Xiaoyu Shen, Nikita Vassilyev, Jesujoba O. Alabi, Yanke Mao, Haonan Gao, and En-Shiun Annie Lee. 2024. https://aclanthology.org/2024.eacl-long.14/ SIB -200: A simple, inclusive, and big evaluation dataset for topic classification in 200+ languages and dialects . In Proceedings of the 18th Conference of the European Chap...

    work page 2024

  2. [2]

    Abhinand Balachandran. 2023. Tamil-llama: A new tamil language model based on llama 2. arXiv preprint arXiv:2311.05845

    arXiv 2023

  3. [3]

    Lucas Bandarkar, Davis Liang, Benjamin Muller, Mikel Artetxe, Satya Narayan Shukla, Donald Husa, Naman Goyal, Abhinandan Krishnan, Luke Zettlemoyer, and Madian Khabsa. 2024. https://doi.org/10.18653/v1/2024.acl-long.44 The belebele benchmark: a parallel reading comprehension dataset in 122 language variants . In Proceedings of the 62nd Annual Meeting of t...

    doi:10.18653/v1/2024.acl-long.44 2024

  4. [4]

    Samuel Cahyawijaya, Holy Lovenia, and Pascale Fung. 2024. https://doi.org/10.18653/v1/2024.naacl-long.24 LLM s are few-shot in-context low-resource language learners . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 405--433, Me...

    doi:10.18653/v1/2024.naacl-long.24 2024

  5. [5]

    Marta R Costa-Juss \`a , James Cross, Onur C elebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, and 1 others. 2022. https://arxiv.org/abs/2207.04672 No language left behind: Scaling human-centered machine translation . arXiv preprint arXiv:2207.04672

    arXiv 2022

  6. [6]

    Yiming Cui, Ziqing Yang, and Xin Yao. 2023. Efficient and effective text encoding for chinese llama and alpaca. arXiv preprint arXiv:2304.08177

    arXiv 2023

  7. [7]

    Chris Dyer, Victor Chahuneau, and Noah A. Smith. 2013. https://aclanthology.org/N13-1073/ A simple, fast, and effective reparameterization of IBM model 2 . In Proceedings of the 2013 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies , pages 644--648, Atlanta, Georgia. Association for Co...

    work page 2013

  8. [8]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    arXiv 2024

Show all 41 references
  1. [9]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  2. [10]

    Suramya Jadhav, Abhay Shanbhag, Amogh Thakurdesai, Ridhima Sinare, and Raviraj Joshi. 2024. On limitations of llm as annotator for low resource languages. arXiv preprint arXiv:2411.17637

    2024 arXiv

  3. [11]

    Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. https://doi.org/10.18653/v1/2020.acl-main.560 The state and fate of linguistic diversity and inclusion in the NLP world . In Proceedings of the 58th Annual Meeting of the Association for Co...

    2020 doi

  4. [12]

    Taku Kudo and John Richardson. 2018. https://doi.org/10.18653/v1/D18-2012 S entence P iece: A simple and language independent subword tokenizer and detokenizer for neural text processing . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processin...

    2018 doi

  5. [13]

    Simon Kurz, Jian-Jia Chen, Lucie Flek, and Zhixue Zhao. 2024. Investigating language-specific calibration for pruning multilingual large language models. arXiv preprint arXiv:2408.14398

    2024

  6. [14]

    Celio Larcher, Marcos Piau, Paulo Finardi, Pedro Gengo, Piero Esposito, and Vinicius Carid \'a . 2023. Cabrita: closing the gap for foreign languages. arXiv preprint arXiv:2308.11878

    2023 arXiv

  7. [15]

    Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ili \'c , Daniel Hesslow, Roman Castagn \'e , Alexandra Sasha Luccioni, Fran c ois Yvon, Matthias Gall \'e , and 1 others. 2023. Bloom: A 176b-parameter open-access multilingual language model

    2023

  8. [16]

    Joseph Lev. 1949. The point biserial coefficient of correlation. The Annals of Mathematical Statistics, 20(1):125--126

    1949

  9. [17]

    Haokun Liu, Derek Tam, Muqeeth Mohammed, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022. https://openreview.net/forum?id=rBCvMG-JsPd Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning . In Advances in Neural Information Proc...

    2022

  10. [18]

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. 2024. Dora: weight-decomposed low-rank adaptation. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org

    2024

  11. [19]

    Ilya Loshchilov and Frank Hutter. 2019. https://openreview.net/forum?id=Bkg6RiCqY7 Decoupled weight decay regularization . In International Conference on Learning Representations

    2019

  12. [20]

    Hongyuan Lu, Haoran Yang, Haoyang Huang, Dongdong Zhang, Wai Lam, and Furu Wei. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.55 Chain-of-dictionary prompting elicits translation in large language models . In Proceedings of the 2024 Conference on Empirical Methods in Natur...

    2024 doi

  13. [21]

    Aleksandar Petrov, Emanuele La Malfa, Philip Torr, and Adel Bibi. 2023. https://openreview.net/forum?id=78yDLKi95p Language model tokenizers introduce unfairness between languages . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  14. [22]

    Evgeniia Razumovskaia, Ivan Vulic, and Anna Korhonen. 2024. https://doi.org/10.48550/arXiv.2403.01929 Analyzing and adapting large language models for few-shot multilingual nlu: Are we there yet? CoRR, abs/2403.01929

    2024 doi arXiv

  15. [23]

    Stephen Robertson, Hugo Zaragoza, and 1 others. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  16. [24]

    Ipek Baris Schlicht, Zhixue Zhao, Burcu Sayin, Lucie Flek, and Paolo Rosso. 2025. Do llms provide consistent answers to health-related questions across languages? In European Conference on Information Retrieval, pages 314--322. Springer

    2025

  17. [25]

    Uri Shaham, Jonathan Herzig, Roee Aharoni, Idan Szpektor, Reut Tsarfaty, and Matan Eyal. 2024. https://doi.org/10.18653/v1/2024.findings-acl.136 Multilingual instruction tuning with just a pinch of multilinguality . In Findings of the Association for Computational Linguistics:...

    2024 doi

  18. [26]

    Lingfeng Shen, Weiting Tan, Sihao Chen, Yunmo Chen, Jingyu Zhang, Haoran Xu, Boyuan Zheng, Philipp Koehn, and Daniel Khashabi. 2024. https://doi.org/10.18653/v1/2024.findings-acl.156 The language barrier: Dissecting safety challenges of LLM s in multilingual contexts . In Find...

    2024 doi

  19. [27]

    Eshaan Tanwar, Subhabrata Dutta, Manish Borthakur, and Tanmoy Chakraborty. 2023. https://doi.org/10.18653/v1/2023.acl-long.346 Multilingual LLM s are better cross-lingual in-context learners with alignment . In Proceedings of the 61st Annual Meeting of the Association for Comp...

    2023 doi

  20. [28]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2023 arXiv

  21. [29]

    Alexander Tsvetkov and Alon Kipnis. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.468 Information parity: Measuring and predicting the multilingual capabilities of language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7971--7...

    2024 doi

  22. [30]

    Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020. Neural machine translation with byte-level subwords. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9154--9160

    2020

  23. [31]

    Genta Winata, Shijie Wu, Mayank Kulkarni, Thamar Solorio, and Daniel Preotiuc-Pietro. 2022. https://doi.org/10.18653/v1/2022.aacl-main.59 Cross-lingual few-shot learning on unseen languages . In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association f...

    2022 doi

  24. [32]

    Atsuki Yamaguchi, Aline Villavicencio, and Nikolaos Aletras. 2024 a . https://doi.org/10.18653/v1/2024.findings-emnlp.396 An empirical study on cross-lingual vocabulary adaptation for efficient language model inference . In Findings of the Association for Computational Linguis...

    2024 doi

  25. [33]

    Atsuki Yamaguchi, Aline Villavicencio, and Nikolaos Aletras. 2024 b . How can we effectively expand the vocabulary of llms with 0.01 gb of target language text? arXiv preprint arXiv:2406.11477

    2024

  26. [34]

    Zheng Xin Yong, Hailey Schoelkopf, Niklas Muennighoff, Alham Fikri Aji, David Ifeoluwa Adelani, Khalid Almubarak, M Saiful Bari, Lintang Sutawika, Jungo Kasai, Ahmed Baruwa, Genta Winata, Stella Biderman, Edward Raff, Dragomir Radev, and Vassilina Nikoulina. 2023 a . https://d...

    2023 doi

  27. [35]

    Zheng Xin Yong, Ruochen Zhang, Jessica Forde, Skyler Wang, Arjun Subramonian, Holy Lovenia, Samuel Cahyawijaya, Genta Winata, Lintang Sutawika, Jan Christian Blaise Cruz, Yin Lin Tan, Long Phan, Long Phan, Rowena Garcia, Thamar Solorio, and Alham Fikri Aji. 2023 b . https://ac...

    2023

  28. [36]

    Chen Zhang, Xiao Liu, Jiuheng Lin, and Yansong Feng. 2024. https://doi.org/10.18653/v1/2024.findings-acl.519 Teaching large language models an unseen language on the fly . In Findings of the Association for Computational Linguistics: ACL 2024, pages 8783--8800, Bangkok, Thaila...

    2024 doi

  29. [37]

    Xiang Zhang, Senyu Li, Bradley Hauer, Ning Shi, and Grzegorz Kondrak. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.491 Don`t trust C hat GPT when your question is not in E nglish: A study of multilingual abilities and types of LLM s . In Proceedings of the 2023 Conference...

    2023 doi

  30. [38]

    Zhengyan Zhang, Yuxian Gu, Xu Han, Shengqi Chen, Chaojun Xiao, Zhenbo Sun, Yuan Yao, Fanchao Qi, Jian Guan, Pei Ke, and 1 others. 2021. Cpm-2: Large-scale cost-effective pre-trained language models. arXiv preprint arXiv:2106.10715

    2021 arXiv

  31. [39]

    Zhixue Zhao and Nikolaos Aletras. 2024. https://doi.org/10.18653/v1/2024.naacl-long.178 Comparing explanation faithfulness between multilingual and monolingual fine-tuned language models . In Proceedings of the 2024 Conference of the North American Chapter of the Association f...

    2024 doi

  32. [40]

    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...

  33. [41]

    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 5, 2026 · model on record in the stance chip above.