Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Cross-Lingual Optimization for Language Transfer in Large Language Models

T0 review · 4 major / 5 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 CLO, a modified DPO loss that contrasts English and translated target-language responses in the same batch, improves target-language instruction following and preserves English better than standard SFT.

desk verdict Useful empirical recipe for cross-lingual transfer, but the main loss is just DPO on translated pairs and the attention-only mask is never controlled in the baselines. read the letter →

arxiv 2505.14297 v1 pith:MEWYLTWH submitted 2025-05-20 cs.CL

Jungseob Lee , Seongtae Hong , Hyeonseok Moon , Heuiseok Lim This is my paper
classification cs.CL
keywords languageenglishlanguagescross-lingualdatamodelsperformancelarge
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

Many large language models are built from English-heavy data. As a result, when given a question in another language, they either fail to understand it or answer in English. The usual remedy is to fine-tune the model on translated instructions, but this often degrades English ability and helps little when translated data is scarce.

The authors propose Cross-Lingual Optimization (CLO). They start with English instruction-response pairs and translate both the instruction and the response into a target language using a translation model. Then they construct preference pairs: for an English prompt, the English response is marked as good and the translated response as bad; for a translated prompt, the translated response is good and the English response is bad. A preference loss, in the style of Direct Preference Optimization (DPO), trains the model to prefer the response that matches the input language. The model is also trained with a standard language-modeling loss on the target-language responses only.

CLO is tested on five base models (Llama-2, Llama-3, Mistral, Qwen) and six languages ranging from high-resource (Chinese, German) to low-resource (Swahili, Yoruba). The method typically beats standard fine-tuning on instruction-following win rates, reading comprehension (BELEBELE), and reasoning (MMMLU), while also maintaining or improving English scores. The largest gains appear in low-resource languages: for Swahili, CLO trained on about 3,200 examples performs around the level of standard fine-tuning trained on 6,400 examples.

Extended reading notes

Core claim

The central claim is that Cross-Lingual Optimization (CLO), a DPO-based loss with cross-lingual preference pairs and attention-only fine-tuning, 'consistently outperforms SFT in both acquiring target language proficiency and maintaining English performance,' with the strongest gains in low-resource languages where CLO with 3,200 samples surpasses SFT with 6,400 samples (Abstract, Section 6, Table 1). If true, CLO offers a data- and compute-efficient alternative to standard supervised fine-tuning for language transfer.

Load-bearing premise

The paper's first key hypothesis (Section 3): 'given an input query in a non-English target language, suppressing English responses while strengthening responses in the target language enables the model to leverage its existing English knowledge to generate outputs in the target language.' This assumes the model's knowledge is language-independent and can be expressed in any target language once the output-language bias is corrected. If false, CLO would increase language-matching but not real capability, and the BELEBELE/MMMLU gains would vanish. The evidence for transferability is indirect: a single Swahili example and benchmark improvements.

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 / 5 minor

Summary. The paper proposes Cross-Lingual Optimization (CLO), a fine-tuning method for transferring English-centric LLMs to target languages. CLO builds a cross-lingual dataset by translating English instruction data into the target language and defines a loss that combines a target-language Negative Log-Likelihood term with DPO-style preference pairs that suppress English responses to target-language prompts and target-language responses to English prompts. The method also restricts parameter updates to attention layers. Experiments on five LLMs and six languages (Chinese, German, Korean, Indonesian, Swahili, Yoruba) report that CLO achieves higher AlpacaEval win rates against an SFT baseline than SFT+DPO, with additional gains on BELEBELE and MMMLU, and shows better data efficiency in low-resource languages.

Significance. If the results are substantiated, CLO would provide a practical, data-efficient alternative to supervised fine-tuning for language transfer, using only translated English instruction data while preserving English performance. The paper has notable strengths: a broad empirical matrix (five base models across six languages), evaluation on external benchmarks (BELEBELE, MMMLU) that go beyond the custom AlpacaEval setup, and ablation studies that probe the parameter mask and the NLL loss. The main weaknesses are that the causal role of the cross-lingual objective is not isolated from the attention-only mask and that the mathematical novelty of the loss relative to standard DPO is not established. The overall contribution is potentially useful but is currently overstated.

major comments (4)
  1. [Section 3.2/4.1, Tables 1 and 4] The reported gains of CLO over SFT and SFT+DPO cannot be attributed to the cross-lingual loss because the baselines are not trained with the same attention-only parameter mask. The paper only states that 'all baselines and models' use identical hyperparameters (Section 4.1), which does not cover the parameter mask; standard SFT and DPO implementations update all parameters. Table 4 shows the mask is not neutral: in Swahili, attention-only CLO wins only 29.69% against full-parameter CLO. Without an SFT baseline and an SFT+DPO baseline using the same mask, the consistent win rates in Table 1 could be due to the regularizing effect of parameter-efficient fine-tuning rather than the cross-lingual preference objective. This confound is load-bearing for the paper's central claim that the CLO objective is responsible for the gains.
  2. [Appendix A, Eqs. (5)-(7), (10)-(12)] The paper claims that L_CL differs from standard DPO by 'utilizing cross-lingual data pairs within the same batch.' However, the loss is an additive average of per-example DPO log-sigmoid terms for English and target preference pairs, with no cross-sample or batch-level coupling. This is mathematically identical to applying the standard DPO loss to the union of the two preference datasets with the same reference model and β. The only remaining differences are the target-only NLL term and the attention mask, not the preference loss itself. The statement in Appendix A that standard DPO 'might not utilize the relationships between accepted and rejected responses across languages' is therefore unsupported by the loss formulation.
  3. [Table 1] The claim that CLO 'consistently outperforms SFT' is not statistically supported in many cells. For example, Mistral-7B shows target win rates of 50.8±1.76 (German), 50.5±1.76 (Korean), 51.1±1.76 (Indonesian), and 51.1±1.52 (Yoruba), which are within one standard deviation of 50%. Similarly, Llama-2-7B has deltas of +0.4 to +1.3 in Korean, Indonesian, and Swahili. With the reported standard deviations, these differences are not meaningful. The paper should provide significance tests or confidence intervals for the win-rate comparisons, or soften the 'consistent' claim to a directional trend.
  4. [Section 7.1 and Table 4] The opening statement that attention-only training 'maintains or even improves' performance is contradicted by the table itself, which shows attention-only CLO loses to full-parameter CLO in Swahili by 29.69% win versus 69.57% lose. While the next sentence acknowledges this drop, the blanket claim should be revised to state the limitation, especially because Swahili is one of the headline low-resource languages in the abstract.
minor comments (5)
  1. [Abstract and Section 6] The abstract states 'CLO with only 3,200 samples surpasses SFT with 6,400 samples,' while Section 6 describes the 1,600-sample CLO as 'comparable' to the 6,400-sample SFT and the 3,200-sample CLO as 'superior.' Please align the wording for consistency.
  2. [Section 5, BELEBELE] There is a typo: 'BELEBLE' should be 'BELEBELE'.
  3. [Appendix D] The description of the answer-extraction procedure ('the evaluation model included 10 English and 10 target language test samples per target language in its training') is unclear; specify which model is trained and how exactly the test samples are excluded from scoring.
  4. [References] Shaham et al. (2024a) and (2024b) appear to be the same paper and should be cited once. The reference 'Zeping and Sophia (2024)' is non-archival ('Under review'); since it motivates the attention-only design, a citable version or additional evidence would be helpful.
  5. [Table 1 caption] The caption 'evaluated against their SFT baselines' is ambiguous; clarify that the win rate is computed against the corresponding SFT-trained model.

Circularity Check

2 steps flagged · score 4.0 of 10

The core loss reduces to standard DPO on cross-lingual pairs, and the main AlpacaEval metric rewards the exact behavior CLO is trained to enforce; external benchmarks temper the circularity.

  1. renaming known result [Appendix A, Eqs. (10)-(12); Section 3.2]
    "In the standard DPO approach with cross-lingual data augmentation, the loss function is applied independently to each language's data, and the model does not utilize cross-lingual data pairs within the same batch. ... This is fundamentally different from just applying DPO on cross-lingual data augmentation."

    The claimed distinction is contradicted by the paper's own equations. L_CL in Eq. (10) is an average of log sigma(z_en^{(i)}) + log sigma(z_l^{(i)}), where z_en^{(i)} is the DPO log-ratio for the pair (x_en, y_en, y_l) and z_l^{(i)} is the DPO log-ratio for (x_l, y_l, y_en). Averaging these terms is exactly the standard DPO loss on the combined cross-lingual augmented dataset; there is no coupling between different samples i or any joint batch-level term. Thus the 'cross-lingual loss' is a relabeling of standard DPO applied to translated preference pairs, not a new objective that 'differs from the standard DPO loss.'

  2. other [Appendix C (AlpacaEval setup); Section 3.2, Eq. (7)]
    "Additionally, since the purpose is to measure multilinguality, if a model responds in a language different from the instruction's language, this should be reflected in the evaluation."

    The CLO loss in Eq. (7) is specifically designed to increase log pi(y_l|x_l) and decrease log pi(y_en|x_l), i.e. to suppress English responses to target-language inputs. The modified AlpacaEval evaluator is explicitly instructed to penalize responses in a language different from the instruction. Therefore the target-language win-rate advantage reported in Table 1 is partly enforced by aligning the metric with the training objective, rather than measuring transferred knowledge independently. The BELEBELE and MMMLU results are external and not circular, which limits the severity.

full rationale

The paper does not fit parameters to its target benchmarks, and it does not rely on a load-bearing self-citation chain or an imported uniqueness theorem. The genuinely circular component is the claimed novelty of the loss: L_CL in Eq. (10) is, up to a constant, the standard DPO loss on the two cross-lingual preference pairs (x_en, y_en, y_l) and (x_l, y_l, y_en). The Appendix A claim that CLO is 'fundamentally different' from standard DPO on cross-lingual data is not supported by the equations, since each sample's z_en and z_l are independent DPO log-ratios and the average over the batch is just DPO on the augmented dataset. A second, partial issue is that the main AlpacaEval metric was modified to penalize language mismatch, which is exactly the behavior CLO's Eq. (7) is trained to produce; this makes part of the headline win-rate and data-efficiency results on AlpacaEval favorable by construction. However, the paper is not purely circular: all models are also evaluated on BELEBELE and MMMLU, which do not encode the CLO objective, and those external benchmarks independently show gains. The overall circularity is therefore partial and concentrated in the loss-formulation claim and the AlpacaEval evaluation design, not in every reported result.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on one main design parameter (λ), a standard DPO coefficient, and several domain assumptions about language-independent knowledge, translation quality, and the sufficiency of attention-only fine-tuning. No new physical or conceptual entities are postulated.

free parameters (2)
  • λ (trade-off parameter) = 0.5
    Fixed at 0.5 for all models and languages; chosen by hand rather than tuned. Balances the supervised NLL loss and the cross-lingual preference loss, directly controlling the strength of language-correspondence training.
  • β (DPO KL coefficient) = 0.1
    Standard DPO hyperparameter controlling KL-regularization strength; set to 0.1. Not fitted to the target result.
assumptions (4)
  • domain assumption The model's internal knowledge is language-independent and can be expressed in the target language once output-language bias is corrected.
    Section 3, first key hypothesis. If false, CLO only teaches language-switching without knowledge transfer; benchmark gains would not materialize.
  • domain assumption Translated target-language data (via M2M100) preserves the instruction-response semantic content well enough for preference learning.
    Section 3.2. The DPO pairs are valid only if y_l is a faithful translation of y_en. The paper acknowledges translation errors but argues that both baselines share the same data.
  • domain assumption Training only attention layers is sufficient for cross-lingual transfer.
    Section 3.2, based on Zeping and Sophia (2024), an ACL Rolling Review submission under review. The paper's own ablation (Table 4) shows attention-only is much worse than full training for Swahili, so this does not hold universally.
  • domain assumption The GPT-4o evaluation with language-specific prompts and language-mismatch penalty is a valid measure of target-language instruction following.
    Appendix C. This is a constructed metric; external benchmarks provide additional support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Lingual Optimization for Language Transfer in Large Language Models." pith.science (2026). https://pith.science/paper/MEWYLTWH

@misc{pith2026250514297,
  author       = {Pith},
  title        = {Pith review of: Cross-Lingual Optimization for Language Transfer in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEWYLTWH}},
  note         = {Machine review of arXiv:2505.14297}
}
read the original abstract

Adapting large language models to other languages typically employs supervised fine-tuning (SFT) as a standard approach. However, it often suffers from an overemphasis on English performance, a phenomenon that is especially pronounced in data-constrained environments. To overcome these challenges, we propose \textbf{Cross-Lingual Optimization (CLO)} that efficiently transfers an English-centric LLM to a target language while preserving its English capabilities. CLO utilizes publicly available English SFT data and a translation model to enable cross-lingual transfer. We conduct experiments using five models on six languages, each possessing varying levels of resource. Our results show that CLO consistently outperforms SFT in both acquiring target language proficiency and maintaining English performance. Remarkably, in low-resource languages, CLO with only 3,200 samples surpasses SFT with 6,400 samples, demonstrating that CLO can achieve better performance with less data. Furthermore, we find that SFT is particularly sensitive to data quantity in medium and low-resource languages, whereas CLO remains robust. Our comprehensive analysis emphasizes the limitations of SFT and incorporates additional training strategies in CLO to enhance efficiency.

Figures

Figures reproduced from arXiv: 2505.14297 by the authors.

Figure 1
Figure 1. Example responses to a Swahili query gen [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of cross-lingual dataset preparation and optimization method. The process begins with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of win rates between CLO and SFT on Llama-2-7B models trained with varying amounts [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of average MMMLU performance by category for CLO and SFT models of Llama-2 and [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Comparison of win rates between CLO and SFT Llama-3 models, trained with varying amounts of [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 9 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Lucas Bandarkar, Davis Liang, Benjamin Muller, Mikel Artetxe, Satya Narayan Shukla, Donald Husa, Naman Goyal, Abhinandan Krishnan, Luke Zettlemoyer, and Madian Khabsa. 2023. The belebele benchmark: a parallel reading comprehension dataset in 122 language variants. arXiv preprint arXiv:2308.16884

    arXiv 2023

  5. [5]

    Kuang-Ming Chen and Hung-yi Lee. 2024. Instructioncp: A fast approach to transfer large language models into target language. arXiv preprint arXiv:2405.20175

    arXiv 2024

  6. [6]

    Nadezhda Chirkova and Vassilina Nikoulina. 2024. Zero-shot cross-lingual transfer in instruction tuning of large language model. arXiv preprint arXiv:2402.14778

    work page Pith review arXiv 2024

  7. [7]

    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

  8. [8]

    Wietse De Vries, Martijn Bartelds, Malvina Nissim, and Martijn Wieling. 2021. Adapting monolingual models: Data can be scarce when language similarity is high. arXiv preprint arXiv:2105.02855

    work page Pith review arXiv 2021

Show all 36 references
  1. [9]

    Konstantin Dobler and Gerard De Melo. 2023. Focus: Effective embedding initialization for monolingual specialization of multilingual models. arXiv preprint arXiv:2305.14481

    2023 arXiv

  2. [10]

    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

    2024 arXiv

  3. [11]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306

    2024 arXiv

  4. [12]

    Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, et al. 2021. Beyond english-centric multilingual machine translation. Journal of Machine Learning Research, 22(107):1--48

    2021

  5. [13]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang S...

    2024 doi

  6. [14]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR)

    2021

  7. [15]

    Jiwoo Hong, Noah Lee, and James Thorne. 2024. Orpo: Monolithic preference optimization without reference model. arXiv preprint arXiv:2403.07691, 2(4):5

    2024 arXiv

  8. [16]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

    2023 arXiv

  9. [17]

    o pf, Yannic Kilcher, Dimitri von R \

    Andreas K \"o pf, Yannic Kilcher, Dimitri von R \"u tte, Sotiris Anagnostidis, Zhi Rui Tam, Keith Stevens, Abdullah Barhoum, Duc Nguyen, Oliver Stanley, Rich \'a rd Nagyfi, et al. 2024. Openassistant conversations-democratizing large language model alignment. Advances in Neura...

    2024

  10. [18]

    Jungseob Lee, Hyeonseok Moon, Seungjun Lee, Chanjun Park, Sugyeong Eo, Hyunwoong Ko, Jaehyung Seo, Seungyoon Lee, and Heui-Seok Lim. 2024. Length-aware byte pair encoding for mitigating over-segmentation in korean machine translation. In Findings of the Association for Computa...

    2024

  11. [19]

    SeungJun Lee, Taemin Lee, Jeongwoo Lee, Yoonna Jang, and Heuiseok Lim. 2023. Kullm: Learning to construct korean instruction-following large language models. In Annual Conference on Human and Language Technology, pages 196--202. Human and Language Technology

    2023

  12. [20]

    Hashimoto

    Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval

    2023

  13. [21]

    Yudong Li, Yuhao Feng, Wen Zhou, Zhe Zhao, Linlin Shen, Cheng Hou, and Xianxu Hou. 2024. Dynamic data sampler for cross-language transfer learning in large language models. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p...

    2024

  14. [22]

    Kelly Marchisio, Wei-Yin Ko, Alexandre B \'e rard, Th \'e o Dehaze, and Sebastian Ruder. 2024. Understanding and mitigating language confusion in llms. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6653--6677

    2024

  15. [23]

    Benjamin Minixhofer, Fabian Paischer, and Navid Rekabsaz. 2021. Wechsel: Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models. arXiv preprint arXiv:2112.06598

    2021 arXiv

  16. [24]

    Masayasu Muraoka, Bishwaranjan Bhattacharjee, Michele Merler, Graeme Blackwood, Yulong Li, and Yang Zhao. 2023. Cross-lingual transfer of large language model by visually-derived supervision toward low-resource languages. In Proceedings of the 31st ACM International Conference...

    2023

  17. [25]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

    2024

  18. [26]

    Uri Shaham, Jonathan Herzig, Roee Aharoni, Idan Szpektor, Reut Tsarfaty, and Matan Eyal. 2024 a . Multilingual instruction tuning with just a pinch of multilinguality. arXiv preprint arXiv:2401.01854

    2024 arXiv

  19. [27]

    Uri Shaham, Jonathan Herzig, Roee Aharoni, Idan Szpektor, Reut Tsarfaty, and Matan Eyal. 2024 b . 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 Linguisti...

    2024 doi

  20. [28]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi \`e re, Mihir Sanjay Kale, Juliette Love, et al. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295

    2024 arXiv

  21. [30]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023 b . Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  22. [31]

    Ke Tran. 2020. From english to foreign languages: Transferring pre-trained language models. arXiv preprint arXiv:2002.07306

    2020 arXiv

  23. [32]

    Ethan Waisberg, Joshua Ong, Mouayad Masalkhi, Sharif Amit Kamran, Nasif Zaman, Prithul Sarker, Andrew G Lee, and Alireza Tavakkoli. 2023. Gpt-4: a new era of artificial intelligence in medicine. Irish Journal of Medical Science (1971-), 192(6):3197--3200

    2023

  24. [33]

    Tengyang Xie, Dylan J Foster, Akshay Krishnamurthy, Corby Rosset, Ahmed Awadallah, and Alexander Rakhlin. 2024. Exploratory preference optimization: Harnessing implicit q*-approximation for sample-efficient rlhf. arXiv preprint arXiv:2405.21046

    2024 arXiv

  25. [34]

    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. [35]

    Yu Zeping and Ananiadou Sophia. 2024. https://openreview.net/forum?id=HaaPCU2LvE Neuron-level knowledge attribution in large language models . In Submitted to ACL Rolling Review - June 2024. Under review

    2024

  27. [36]

    Jun Zhao, Zhihao Zhang, Qi Zhang, Tao Gui, and Xuanjing Huang. 2024 a . Llama beyond english: An empirical study on language capability transfer. arXiv preprint arXiv:2401.01055

    2024 arXiv

  28. [37]

    Yiran Zhao, Wenxuan Zhang, Huiming Wang, Kenji Kawaguchi, and Lidong Bing. 2024 b . Adamergex: Cross-lingual transfer with large language models via adaptive adapter merging. arXiv preprint arXiv:2402.18913

    2024 arXiv

Pith tools

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