Pith. sign in

REVIEW 4 major objections 6 minor 32 references

CALM: Unleashing the Cross-Lingual Self-Aligning Ability of Language Model Question Answering

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read By sampling answers to the same question in several languages and treating the majority answer as the target, a language model can align its own multilingual knowledge without ground-truth labels.

desk verdict A simple, plausible cross-lingual self-training recipe with real gains, but the claim to beat ground-truth DPO is uncontrolled and needs a fix. read the letter →

arxiv 2501.18457 v2 pith:QC6NVL4Z submitted 2025-01-30 cs.CL

classification cs.CL
keywords cross-lingualconsistencydirectpreferenceoptimizationself-consistencymultilingualmajorityvotingchain-of-thoughtpromptinglabel-freealignmentquestionansweringretrieval-augmentedgeneration
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

Large language models often answer the same question differently depending on the language it is asked in, even when the question does not depend on culture. This paper argues that this inconsistency is itself a usable training signal: sample several chain-of-thought answers in different languages, take the majority-voted answer as the preferred response, treat divergent answers as negative responses, and fine-tune the model with direct preference optimization. The authors report average accuracy gains of +3.76 percentage points on MEDQA and +5.55 percentage points on X-CSQA across three base models, together with higher cross-lingual consistency, and gains that appear even in languages and datasets not used during training. If the claim is right, a model can improve its multilingual question answering using only its own outputs, with no human annotation.

What carries the argument

The central mechanism is multilingual majority voting over chain-of-thought answers followed by preference-pair construction and the direct preference optimization (DPO) objective $$L_{\text{DPO}}(\pi_\$\theta$;\pi_{\text{ref}}) = \mathbb{E}_{(q,p_w,p_l)\sim\mathcal{D}}\left[\log\$\sigma$\left(\$\beta$\log\frac{\pi_\$\theta$(p_w|q)}{\pi_{\text{ref}}(p_w|q)} - \$\beta$\log\frac{\pi_\$\theta$(p_l|q)}{\pi_{\text{ref}}(p_l|q)}\right)\right]$$ which pushes the model toward the voted answer $p_w$ and away from each inconsistent answer $p_l$. The voted answer is translated into the language of each negative sample so that every preference pair is expressed in one language. This machinery turns the model's own cross-lingual consistency into a label-free training reward.

What would settle it

Take a multilingual question set with known answers, run CALM's majority voting, and compare models trained only on pairs where the majority answer is correct against models trained on all pairs: if the accuracy gain disappears or reverses when majority answers are factually wrong, the claim that incorrect-but-consistent answers improve training is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that cross-lingual self-consistency can be converted into a preference signal for direct preference optimization. CALM samples multiple chain-of-thought responses to each question in English and two other languages, selects the most-voted answer as the positive sample, uses every inconsistent answer as a negative sample, and fine-tunes the model on the resulting pairs. The paper reports that this training raises accuracy beyond both the base model and a supervised fine-tuning baseline, that the gains grow as more languages join the vote, and that the effect transfers to unseen languages and to a different dataset. Since the winning answer can come from any language, the method exploits knowledge stored in non-English languages rather than forcing everything into English.

Load-bearing premise

The method relies on the premise that the answer a model gives most often across languages is a better training target than the answers it gives less often, even when that majority answer is factually wrong.

Editorial extensions

If this is right

  • Multilingual question answering can be improved without any annotated answers, using only the model's own responses, which lowers the cost of aligning models for low-resource languages.
  • Adding more languages to the voting pool increases both accuracy and cross-lingual consistency, so including diverse languages in training should be beneficial even if those languages are not the target of deployment.
  • The gains extend to languages never seen in CALM training, suggesting that cross-lingual alignment of training data produces a more language-agnostic internal knowledge representation.
  • The method also works when the base model is retrieval-augmented, meaning external documents can be aligned across languages in the same label-free way.
  • If the gains over ground-truth DPO hold, then a confident, self-consistent answer is a better training signal than a factually correct label alone for improving reasoning.

Reading between the lines

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

  • The authors leave implicit that the same self-consistency trick could be applied to output dimensions other than language, such as sampling temperature, domain, or persona, treating self-consistency as a general label-free reward.
  • A testable extension would be to weight DPO pairs by the majority-vote margin or by a confidence estimate, so that low-margin questions do not contribute noisy preference signal; the paper's ablations do not separate margin-based quality.
  • Since the authors note their translations come from an external API, a fully translation-free variant that samples from multilingual prompts natively available to the model could isolate the self-alignment effect from translation error.
  • An implication the authors do not draw is that the chosen-but-incorrect answers might be reinforcing a transferable reasoning style rather than factual correctness, which could be tested by comparing fine-tuned models on questions where the majority answer is wrong.
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 / 6 minor

Summary. The paper proposes CALM, a three-stage method for cross-lingual knowledge alignment. For each question, CALM samples multiple Chain-of-Thought outputs in up to three languages, selects the most frequent answer via majority voting as a positive sample, pairs it with inconsistent answers as negatives, and applies direct preference optimization (DPO) with LoRA. The authors evaluate on MEDQA and X-CSQA with Llama3-8B, Mistral-7B, and Self-RAG, reporting accuracy, consistency, and AC3 gains over the base models and an SFT variant trained on the same self-generated positives. They also report results in a retrieval-augmented setting, claim that CALM outperforms ground-truth DPO and SFT, and argue that increasing the number of training languages improves both accuracy and consistency.

Significance. If the central claims hold, CALM is a practical, annotation-free method for improving multilingual consistency and accuracy of open-weight LLMs on knowledge-grounded QA, with a simple recipe and a public code link. The consistent directional gains across three base models and two datasets, together with the positive scaling trend when more languages are added, make the core idea worth taking seriously. The paper is also transparent about its hyperparameters and training configuration, which supports reproducibility. The main caveat is that the evidence for the more ambitious claims—that noisy self-consistency labels are sufficient and that CALM beats ground-truth supervision—currently rests on uncontrolled or under-reported comparisons.

major comments (4)
  1. [§4.2, Table 2 vs Table 1] The conclusion that 'CALM outperforms ground truth DPO and SFT' is not supported by the reported numbers. For Mistral on X-CSQA English, Table 2 lists SFT-w/GT = 73.0 and DPO-w/GT = 73.2, while Table 1 lists CALM = 68.1; the per-language comparison therefore directly contradicts the text. If the claim is intended as an average over languages, that average is not stated or tested. In addition, the GT baselines are trained only on the subset of pairs whose positive label equals ground truth, so they receive fewer examples than CALM; the comparison conflates label quality with data quantity. Please add a controlled comparison that matches the number of training examples (e.g., subsample CALM pairs to the GT subset size, or upsample the GT subset) and report per-language numbers alongside any aggregate claim.
  2. [Tables 1–3, 9, 10] No variance or significance testing is reported for any of the central accuracy or consistency results. The headline gains are driven by averages over languages and models, and several per-language differences are small (e.g., Llama3 MEDQA EN 60.1 to 63.5; Mistral MEDQA ZH 36.4 to 38.5), so it is impossible to assess whether the improvements are stable. Please report standard deviations or bootstrap confidence intervals over multiple sampling seeds and, where meaningful, paired significance tests over the test questions.
  3. [§4.3, Table 3] The cross-dataset generalizability claim is not verifiable as printed. The text says both transfer settings 'surpass the baseline,' but Table 3 contains no baseline rows or columns, and it is unclear which numbers correspond to MEDQA-to-X-CSQA versus X-CSQA-to-MEDQA. When compared with the SFT baselines in Table 1, some values are far below (e.g., Llama3 X-CSQA EN 60.9 in Table 3 versus 73.8 for +SFT in Table 1; Mistral X-CSQA EN 52.9 versus 67.7). Please clarify the experimental protocol and add explicit baseline values for each transfer direction.
  4. [§4.1, Table 4] The paper's defense of noisy self-consistent labels rests on one qualitative example. Section 4.1 acknowledges that majority-voted answers can be factually incorrect, but the only evidence that such positives 'demonstrate better context awareness' is the single item in Table 4. Because the training signal is entirely self-generated (Section 2.2), this is a key risk of the method: DPO could reinforce a confidently wrong consensus. Please add a quantitative analysis, for example comparing the accuracy of chosen versus rejected samples before and after training, or an ablation that trains CALM only on majority-voted pairs that happen to be correct, to isolate the contribution of noisy positives.
minor comments (6)
  1. [Figure 2] The figure contains garbled multilingual labels (e.g., 'Expliciation' instead of 'Explanation', and mixed Chinese/French fragments) that should be cleaned for publication.
  2. [Table 5] The scheduler name 'consine' should be 'cosine'; please check the table for other typos.
  3. [§2.2] The definition of the chosen set S is unclear: 'the set S = {(rik, yik)}k of the most voted explanation-answer pair' should be rewritten as a set of all sampled pairs whose answer equals the majority vote.
  4. [Table 7 caption] 'English tasks up the largest portion' should read 'takes up'; the caption should be edited.
  5. [§3.1] In the consistency metric formula, 'Cs 6' should be written as the binomial coefficient C(6,s) or explained in prose.
  6. [Table 10] The text attributes improved consistency to adding training languages, but Table 10 only reports accuracy; please include consistency or AC3 values for the incremental setting.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation of the accuracy gains: evaluation uses held-out ground-truth labels; the consistency metric tracks the training objective and the ground-truth comparison is confounded by data quantity, but these are validity issues, not circular steps.

full rationale

There is no significant circularity in the claimed derivation chain. CALM's training signal is self-generated through multilingual majority voting, but the paper's headline accuracy results are scored against held-out ground-truth labels on MEDQA and X-CSQA, so the accuracy gains are externally falsifiable and are not recycled from the training labels. The 'Consistency' metric is indeed closely aligned with the DPO objective: the chosen output is the majority-voted answer and the rejected outputs are inconsistent with it, so consistency gains partially measure whether the model learned its own preference objective; this weakens the independent evidentiary value of the consistency numbers but is not an equation-level reduction of a prediction to an input. The authors' own Limitations section concedes that 'the accuracy after majority-voting is still low', which is a real threat to the noisy-label defense. In addition, the comparison in Table 2 is uncontrolled: ground-truth SFT and DPO keep only preference pairs where the positive label matches ground truth ('we only keep the portion of DPO and SFT data that are factually correct'), meaning those baselines are trained on fewer examples than CALM; the Conclusion's statement that 'CALM outperforms ground truth DPO and SFT' is therefore not established by the reported experiments. These are experimental-validity problems, not circularity. The self-consistency assumption is supported by an external citation to Wang et al. (2023), and no load-bearing argument reduces to a self-citation by the present authors. Overall, the core accuracy result is self-contained against external benchmarks, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on standard ML machinery (DPO, LoRA, self-consistency) and on several domain assumptions about multilingual LLM behavior. The most consequential is that a majority-voted answer from the model's own outputs is a good training target even when it is factually wrong; the paper gives qualitative support but no statistical evidence. There are no invented entities.

free parameters (4)
  • m = 3 = 3
    Number of CoT samples generated per question-language; chosen by hand in Section 2.1 and Appendix A.
  • Sampling temperature and top_p = T=1, top_p=0.9
    Decoding hyperparameters reported in Table 6; these affect the diversity of the sampled answers used for voting.
  • DPO beta = 0.1
    Preference optimization strength in DPO objective, set in Table 5.
  • LoRA rank and alpha = r=8, alpha=16
    LoRA configuration for DPO and SFT training, reported in Table 5; these constrain the parameter update.
assumptions (5)
  • standard math The DPO objective is a valid way to optimize preference pairs for LLMs.
    The paper relies on DPO from Rafailov et al. (2024) as the training objective; treated as proven background.
  • domain assumption The answer with the most votes across multilingual samples reflects the model's highest confidence and is more likely to be correct.
    Adopted from self-consistency literature (Wang et al. 2023) and applied across languages; justified by Figure 3, which shows voting accuracy above monolingual accuracy.
  • domain assumption Culturally-independent questions should receive consistent answers across languages.
    This is the conceptual motivation of the paper (Introduction, Figure 1); if false, the goal of cross-lingual consistency would not imply better knowledge alignment.
  • ad hoc to paper Majority-voted answers, even when factually incorrect, contain better task understanding and can serve as effective training targets.
    Stated in Section 4.1 and the Conclusion, supported only by one qualitative example in Table 4; this is the load-bearing assumption of the method.
  • domain assumption Google Translate preserves enough meaning of the questions for training and evaluation to be valid.
    The method translates datasets with Google Translate API (Limitations section); poor translation of medical terminology is acknowledged as a potential limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CALM: Unleashing the Cross-Lingual Self-Aligning Ability of Language Model Question Answering." pith.science (2026). https://pith.science/paper/QC6NVL4Z

@misc{pith2026250118457,
  author       = {Pith},
  title        = {Pith review of: CALM: Unleashing the Cross-Lingual Self-Aligning Ability of Language Model Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QC6NVL4Z}},
  note         = {Machine review of arXiv:2501.18457}
}
read the original abstract

Large Language Models (LLMs) are pretrained on extensive multilingual corpora to acquire both language-specific cultural knowledge and general knowledge. Ideally, while LLMs should provide consistent responses to culture-independent questions across languages, we observe significant performance disparities. To address this, we explore the Cross-Lingual Self-Aligning ability of Language Models (CALM) to align knowledge across languages. Specifically, for a given question, we sample multiple responses across different languages and select the most self-consistent response as the target, leaving the remaining responses as negative examples. We then employ direct preference optimization (DPO) to align the model's knowledge across different languages. Evaluations on the MEDQA and X-CSQA datasets demonstrate CALM's effectiveness in enhancing cross-lingual knowledge question answering, both in zero-shot and retrieval-augmented settings. We also found that increasing the number of languages involved in CALM training leads to higher accuracy and consistency. We offer a qualitative analysis of how cross-lingual consistency can enhance knowledge alignment and explore the method's generalizability.

Figures

Figures reproduced from arXiv: 2501.18457 by the authors.

Figure 1
Figure 1. Knowledge is not well-aligned across lan [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An example of the three stages in our proposed method assuming a question input originally in English. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Visualization of mono-lingual (EN, ZH-CN, FR) percentage accuracy against the multilingual majority [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 7 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  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]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://arxiv.org/abs/2310.11511 Self-rag: Learning to retrieve, generate, and critique through self-reflection . Preprint, arXiv:2310.11511

  4. [4]

    Nuo Chen, Zinan Zheng, Ning Wu, Ming Gong, Yangqiu Song, Dongmei Zhang, and Jia Li. 2023. https://arxiv.org/abs/2310.20246 Breaking language barriers in multilingual mathematical reasoning: Insights and observations . Preprint, arXiv:2310.20246

  5. [5]

    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

  6. [6]

    Changjiang Gao, Hongda Hu, Peng Hu, Jiajun Chen, Jixing Li, and Shujian Huang. 2024. https://arxiv.org/abs/2404.04659 Multilingual pretraining and instruction tuning improve cross-lingual knowledge alignment, but only shallowly . Preprint, arXiv:2404.04659

  7. [7]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . Preprint, arXiv:2106.09685

  8. [8]

    Jiaxin Huang, Shixiang Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.67 Large language models can self-improve . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1051--1068, Singapore. Association for Computational Linguistics

Show all 32 references
  1. [9]

    Kung-Hsiang Huang, Mingyang Zhou, Hou Pong Chan, Yi Fung, Zhenhailong Wang, Lingyu Zhang, Shih-Fu Chang, and Heng Ji. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.41 Do LVLM s understand charts? analyzing and correcting factual errors in chart captioning . In Finding...

  2. [10]

    Yue Huang, Chenrui Fan, Yuan Li, Siyuan Wu, Tianyi Zhou, Xiangliang Zhang, and Lichao Sun. 2024 b . 1+ 1> 2: Can large language models serve as cross-lingual knowledge aggregators? arXiv preprint arXiv:2406.14721

  3. [11]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas ...

  4. [12]

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2020. https://arxiv.org/abs/2009.13081 What disease does this patient have? a large-scale open domain question answering dataset from medical exams . Preprint, arXiv:2009.13081

  5. [13]

    Anubha Kabra, Sanketh Rangreji, Yash Mathur, Aman Madaan, Emmy Liu, and Graham Neubig. 2023. https://arxiv.org/abs/2311.09553 Program-aided reasoners (better) know what they know . Preprint, arXiv:2311.09553

  6. [14]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2024. Large language models are zero-shot reasoners. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA. Curran Associates Inc

  7. [15]

    Bill Yuchen Lin, Seyeon Lee, Xiaoyang Qiao, and Xiang Ren. 2021. https://doi.org/10.18653/v1/2021.acl-long.102 Common sense beyond E nglish: Evaluating and improving multilingual language models for commonsense reasoning . In Proceedings of the 59th Annual Meeting of the Assoc...

  8. [16]

    Geyu Lin, Bin Wang, Zhengyuan Liu, and Nancy F. Chen. 2024. https://arxiv.org/abs/2404.11932 Crossin: An efficient instruction tuning approach for cross-lingual knowledge alignment . Preprint, arXiv:2404.11932

  9. [17]

    Jiateng Liu, Lin Ai, Zizhou Liu, Payam Karisani, Zheng Hui, Yi Fung, Preslav Nakov, Julia Hirschberg, and Heng Ji. 2025. https://aclanthology.org/2025.coling-main.376/ P ropa I nsight: Toward deeper understanding of propaganda in terms of techniques, appeals, and intent . In P...

  10. [18]

    Minh-Thang Luong and Christopher Manning. 2015. https://aclanthology.org/2015.iwslt-evaluation.11 S tanford neural machine translation systems for spoken language domains . In Proceedings of the 12th International Workshop on Spoken Language Translation: Evaluation Campaign, p...

  11. [19]

    Jirui Qi, Raquel Fernández, and Arianna Bisazza. 2023. https://arxiv.org/abs/2310.10378 Cross-lingual consistency of factual knowledge in multilingual language models . Preprint, arXiv:2310.10378

  12. [20]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. https://arxiv.org/abs/2305.18290 Direct preference optimization: Your language model is secretly a reward model . Preprint, arXiv:2305.18290

  13. [21]

    Shuaijie She, Wei Zou, Shujian Huang, Wenhao Zhu, Xiang Liu, Xiang Geng, and Jiajun Chen. 2024. https://arxiv.org/abs/2401.06838 Mapo: Advancing multilingual reasoning through multilingual alignment-as-preference optimization . Preprint, arXiv:2401.06838

  14. [22]

    Chenkai Sun, Jinning Li, Yi Fung, Hou Chan, Tarek Abdelzaher, ChengXiang Zhai, and Heng Ji. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.4 Decoding the silent majority: Inducing belief augmented social graph with large language model for response forecasting . In Proceedi...

  15. [23]

    Xiangru Tang, Chunyuan Deng, Hanminwang Hanminwang, Haoran Wang, Yilun Zhao, Wenqi Shi, Yi Fung, Wangchunshu Zhou, Jiannan Cao, Heng Ji, Arman Cohan, and Mark Gerstein. 2024. https://doi.org/10.18653/v1/2024.emnlp-demo.49 MIMIR : A customizable agent tuning platform for enhanc...

  16. [24]

    Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy F. Chen. 2024. https://arxiv.org/abs/2309.04766 Seaeval for multilingual foundation models: From cross-lingual alignment to cultural reasoning . Preprint, arXiv:2309.04766

  17. [25]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. https://arxiv.org/abs/2203.11171 Self-consistency improves chain of thought reasoning in language models . Preprint, arXiv:2203.11171

  18. [26]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2024. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processi...

  19. [27]

    Fung, Cheng Qian, Jeonghwan Kim, Dilek Hakkani-Tur, and Heng Ji

    Shujin Wu, Yi R. Fung, Cheng Qian, Jeonghwan Kim, Dilek Hakkani-Tur, and Heng Ji. 2025 a . https://aclanthology.org/2025.coling-main.511/ Aligning LLM s with individual preferences via interaction . In Proceedings of the 31st International Conference on Computational Linguisti...

  20. [28]

    (May) Fung, Paul Pu Liang, and Heng Ji

    Shujin Wu, Cheng Qian, Yi R. (May) Fung, Paul Pu Liang, and Heng Ji. 2025 b . Plata: Proactive learning with teacher assistance for weak-to-strong generalization

  21. [29]

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2024. https://arxiv.org/abs/2306.13063 Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms . Preprint, arXiv:2306.13063

  22. [30]

    Yuemei Xu, Ling Hu, Jiayi Zhao, Zihan Qiu, Yuqi Ye, and Hanwen Gu. 2024. https://arxiv.org/abs/2404.00929 A survey on multilingual large language models: Corpora, alignment, and bias . Preprint, arXiv:2404.00929

  23. [31]

    Jifan Yu, Xiaozhi Wang, Shangqing Tu, Shulin Cao, Daniel Zhang-Li, Xin Lv, Hao Peng, Zijun Yao, Xiaohan Zhang, Hanming Li, Chunyang Li, Zheyuan Zhang, Yushi Bai, Yantao Liu, Amy Xin, Nianyi Lin, Kaifeng Yun, Linlu Gong, Jianhui Chen, Zhili Wu, Yunjia Qi, Weikai Li, Yong Guan, ...

  24. [32]

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. 2024. http://arxiv.org/abs/2403.13372 Llamafactory: Unified efficient fine-tuning of 100+ language models . In Proceedings of the 62nd Annual Meeting of the Association for Compu...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.