Pith. sign in

REVIEW 6 major objections 7 minor 40 references

CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation

T0 review · 6 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Two-stage method CCL-XCoT cuts hallucinations in low-resource-language QA by up to 62% by aligning representations with contrastive learning and reasoning in English before answering in the target language.

desk verdict Plausible recipe, but the headline gains rest on an evaluation that doesn't yet support them. read the letter →

arxiv 2507.14239 v1 pith:L2UBU3SU submitted 2025-07-17 cs.CL cs.AI

Weihua Zheng , Roy Ka-Wei Lee , Zhengyuan Liu , Kui Wu , AiTi Aw , Bowei Zou This is my paper
classification cs.CLcs.AI
keywords multilingualLLMhallucinationcross-lingualknowledgetransfercurriculum-basedcontrastivelearningchain-of-thoughtlow-resourcelanguagessemanticalignmentmitigationparameter-efficientfine-tuning
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 claims that hallucinations in low-resource languages, where multilingual LLMs fabricate answers because training data is scarce, can be sharply reduced without retrieval systems or model ensembles. The two-stage framework CCL-XCoT first aligns the semantic spaces of languages with curriculum-based contrastive learning combined with next-token prediction during continued pretraining, then applies cross-lingual Chain-of-Thought (XCoT) prompting at instruction time so the model reasons in English before producing its final answer in the target language. On a 500-question Singaporean cultural-knowledge test translated into Chinese, Malay, and Tamil, the method lifts hallucination-free rates from single-digit baselines (as low as 1% for Tamil) to between 55% and 74%, and the authors report up to a 62% reduction in hallucination rate. The paper also argues that cross-lingual transfer lives mainly in mid-level transformer layers, making targeted mid-layer fine-tuning a more effective and cheaper alternative to full-model tuning. If the results hold, this is a practical, parameter-efficient route to more factual generation in low-resource languages using task-specific data collected only in English.

What carries the argument

The argument runs on two coupled mechanisms. The first is a curriculum-based contrastive learning objective added to next-token prediction during continued pretraining: bilingual sentence pairs are pulled together in embedding space by a sequence-level contrastive loss, then bilingual paragraph pairs are aligned through mean-pooled embeddings, building a shared cross-lingual semantic space without giving up language modeling. The second is the XCoT prompting pipeline, a three-step template used at instruction tuning and inference: reason in English, produce a concise English answer, then restate the answer in the target low-resource language, so factual knowledge stored in English is reused. The paper also identifies mid-level transformer layers (9–19 in Gemma-7B) as the region where cross-lingual transfer is most effective, explaining the alignment mechanism and suggesting a cheaper fine-tuning recipe.

What would settle it

Rerun the comparison on a larger, independently constructed test set scored blind by multiple annotators with reported agreement and confidence intervals, and additionally grade each CCL-XCoT answer with the English reasoning steps removed so only the final target-language response counts; if the 55–74% hallucination-free rates do not survive blind scoring, or the advantage over baselines collapses when outputs are length-normalized, the central claim would be falsified.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that low-resource hallucination is less a problem of model scale or training-data volume than a problem of cross-lingual representation alignment and reasoning scaffolding. After continued pretraining with a curriculum that aligns sentence pairs and then paragraph pairs through a sequence-level contrastive loss, kept alongside next-token prediction to avoid catastrophic forgetting, and after instruction tuning with XCoT prompts that decompose a low-resource question into English reasoning, an English answer, and a target-language answer, both Gemma-7B and LLaMA-3.1-8B reach hallucination-free rates between 55% and 74% on the Chinese, Malay, and Tamil test sets. That contrasts with 1–18% for direct instruction tuning on the same English-only data. A layer-wise experiment locates the transfer in the middle transformer layers: tuning only layers 9–19 of Gemma-7B outperforms full-model fine-tuning, while fine-tuning low-level layers degrades performance, which the paper interprets as evidence that mid-level layers bridge surface linguistic form and higher-order reasoning.

Load-bearing premise

The headline numbers rest on a single 500-question test set that was manually selected and manually scored, with no reported measure of rater agreement or statistical significance, and because XCoT outputs are longer and more structured than baseline outputs, human scorers may be crediting completeness of form rather than correctness of fact.

Editorial extensions

If this is right

  • A tuned 8B model can beat a 70B instruct model on this domain-specific multilingual test, so model scale alone is not the fix for low-resource hallucination.
  • Hallucination mitigation works without retrieval infrastructure, which matters for low-resource languages where retrieval coverage is weak.
  • Fine-tuning only mid-level layers can outperform full-model fine-tuning for cross-lingual transfer, reducing the computational cost of multilingual adaptation.
  • Factual knowledge injected in English transfers into typologically distant languages, and consistency improves even between non-English pairs (e.g., Chinese–Tamil), suggesting English acts as a pivot language.
  • The contrastive pretraining component also lifts cross-lingual NLU accuracy (up to +20% on Cross-MMLU for Chinese), giving the alignment step value beyond hallucination reduction.

Reading between the lines

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

  • The paper does not report blind or length-controlled evaluation, and its Limitations section concedes that fine-grained factual errors such as wrong years, road names, or designer names persist; a natural check is to score only the final target-language sentence after stripping the English reasoning steps, with scorers unaware of which method produced each answer and with baseline outputs matched
  • If mid-level layers are the real site of cross-lingual alignment, a testable extension is to probe which subspaces within those layers encode language-neutral facts, and whether the same mid-level region mediates transfer for other language pairs and other domains.
  • The sentence-to-paragraph curriculum suggests a data-efficiency question: measuring hallucination-free rates as a function of parallel-corpus size and answer length would show how much alignment data is actually needed before the gains saturate.
  • The method's main external requirement is sentence-aligned parallel data for each target pair, which the paper had to collect itself for Malay and Tamil; releasing that data or showing the curriculum works with weaker alignment signals would determine how far the recipe extends to other low-resource languages.
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

6 major / 7 minor

Summary. The paper proposes CCL-XCoT, a two-stage fine-tuning framework for multilingual large language models. Stage 1 performs continued pretraining with a curriculum of sentence-level then paragraph-level contrastive learning, jointly optimized with next-token prediction, to align cross-lingual semantic spaces. Stage 2 applies cross-lingual chain-of-thought (XCoT) instruction fine-tuning, in which the model reasons in English before generating the final answer in the target low-resource language. The method is evaluated on Gemma-7B and LLaMA-3.1-8B for Singaporean cultural QA in Chinese, Malay, and Tamil using a manually selected and scored 500-question test set derived from CRAFT. The main claim is that CCL-XCoT reduces hallucination rates by up to 62%; the paper also reports improved cross-lingual semantic alignment, consistency, and accuracy on SEA-Eval subtasks, plus a layer-wise analysis identifying mid-level transformer layers as the most effective for cross-lingual transfer.

Significance. If the headline results are correct, the paper offers a practical, retrieval-free route to improving factual generation in low-resource languages, with a plausible two-stage mechanism and a clean curriculum design. The manuscript has several strengths: it evaluates on two base models and three target languages including a truly low-resource language (Tamil); Figure 2 provides quantitative evidence that contrastive learning changes the geometry of multilingual representations; the semantic alignment and consistency analyses (Figures 3 and 4) go beyond a single metric; and the layer-wise study, once fully specified, could guide efficient adaptation. The central contribution is clearly stated and the method is simple enough to reproduce. However, the headline 'up to 62%' reduction rests on a manual evaluation with no inter-annotator agreement, significance testing, or error bars, and the experimental protocol omits the nearest existing baseline (the published XCoT method). These issues must be resolved before the results can be considered established. The paper does not provide code, data, or detailed descriptions of the self-collected corpora, which further limits reproducibility.

major comments (6)
  1. [Section 4.2, Table 1; Abstract] The headline claim that CCL-XCoT reduces hallucination rates by up to 62% is carried entirely by Table 1, which reports hallucination-free rates from a 500-question manual evaluation with no inter-annotator agreement, no error bars, and no significance tests. At n=500, the standard error of a proportion is roughly 2 percentage points, so the differences between CL-XCoT and CCL-XCoT (e.g., 68% vs 71% for Chinese, 62% vs 66% for Malay, 55% vs 60% for Tamil on Gemma) are within the noise and do not establish the specific contribution of paragraph-level contrastive learning. Furthermore, the rubric's first criterion ('accurately and fully answers the question') risks rewarding the longer, more structured XCoT outputs for their form rather than for factual content, because XCoT responses include an English reasoning trace plus a target-language answer, unlike the shorter SFT baselines. The authors should report inter-annotator agreement, provide confidence intervals or significance tests for the key comparisons, and use a scoring procedure that evaluates factual claims independently of output length and structure.
  2. [Section 4.2, Table 1; Related Work] The existing cross-lingual CoT method of Chai et al. (2024), which is cited in the references as 'xcot', is not included as a baseline in Table 1. The proposed XCoT strategy is essentially the same idea—reasoning in a high-resource language before answering in a low-resource language—so without a comparison to the published version, the reader cannot determine whether the gains come from the contrastive pretraining stage or simply from re-implementing a known prompting strategy. The authors should add the official XCoT method (or a faithful reproduction) as a row in Table 1 and discuss the incremental contribution of CCL-XCoT over it.
  3. [Section 4.2, RAG baseline] The RAG baseline ('Gemma-7B+SFT+RAG') is not reproducible because the retrieval corpus is unspecified. The text states that Sentence-BERT retrieves 'top-2 relevant paragraphs as in-context prompts' but does not say from which corpus. If the corpus contains the CRAFT test answers, the RAG numbers are inflated; if it is a small or mismatched set, they are depressed. This matters because the RAG comparison is used to argue that retrieval-based methods are fragile in low-resource languages. The authors should specify the retrieval corpus, its size, its domain, and the exact retrieval setup, and ideally verify that the test answers are not present in the retrieved contexts.
  4. [Section 4.5, Figure 5] The layer-wise analysis is underspecified. It is unclear whether the selective fine-tuning applies to the continued-pretraining stage, the instruction-tuning stage, or both, and whether the hallucination-free rates in Figure 5 correspond to the full CCL-XCoT pipeline or to an ablated version. The main configuration in Section 4.1 uses layers 9–19 for continued pretraining and all layers for instruction fine-tuning, yet Section 4.5 concludes that mid-level layers are best; the relationship between the two settings needs to be stated explicitly. As written, the claim in contribution (iv) that mid-level layers are key to cross-lingual knowledge transfer is not auditable.
  5. [Appendix A.2, Table 3] The comparison between LLaMA-3.1-8B+CCL-XCoT and LLaMA-3.1-70B-Instruct conflates model scale with training. The 70B model appears to be used in a zero-shot setting (it is an instruct model), while the 8B model is fine-tuned on a large domain-specific instruction set that includes 3,000 CoT examples per language pair. The claim that 'merely increasing model size and training data is insufficient for enhancing cross-lingual knowledge generalization' is not supported by this comparison, because the 70B model was not trained on the same data. The authors should either fine-tune the 70B model with the same data or clearly label the 70B row as a zero-shot reference point and temper the conclusion accordingly.
  6. [Section A.3, Table 4] The self-collected EN-MS and EN-TA parallel corpora (100K sentence pairs and 10K paragraph pairs each) are not described in terms of source, domain, quality control, or overlap with the test sets. The semantic alignment test sets also use 'a subset sampled from our self-collected data' with no specification of the sampling procedure. The contrastive-learning component in Stage 1 is the central novelty of the method, and its success depends on the quality and non-leakage of these corpora. The authors should provide a detailed data collection description, report basic statistics (e.g., source texts, language identification, filtering criteria), and state how they verified that the contrastive training data and the alignment test data do not overlap with the 500-question hallucination test set.
minor comments (7)
  1. [Abstract; Section 4.2] The abstract states that CCL-XCoT 'reduces hallucination rates by up to 62%' without defining how this number is computed. Table 1 reports hallucination-free rates, and the relative reduction varies by language (e.g., LLaMA-3.1 Chinese moves from 4% to 74%, a ~73% relative reduction). Please state the formula and the exact language/condition for the 62% figure.
  2. [References; Section 2] The Chai et al. (2024) reference is incomplete in the bibliography, and the in-text citation appears in the 'Translation-based Methods' paragraph, which mischaracterizes the XCoT method as translation-based. Please correct the citation and describe Chai et al.'s work accurately in the related work.
  3. [Section 4.2] The sentence 'Hallucination-free responses increase by 38% (Gemma) and 36.6% (LLaMA-3.1) on average' is ambiguous: these appear to be percentage-point increases (e.g., from 8.7% to 47% for Gemma), not relative increases. Please clarify whether all reported gains are percentage points or relative percentages.
  4. [Section 4.2] The sentence 'We achieve average increase of hallucination-free rates across both models, reaching 59% (Chinese), 46% (Malay), 54% (Tamil) for Gemma and 65% (Chinese), 41% (Malay), 49% (Tamil) for LLaMA-3.1' is confusing: these numbers appear to be the CL-XCoT column values, not average increases. Please rewrite for clarity.
  5. [Section 3.2, Eq. (3)] In Eq. (3), the denominator in the second fraction uses 'TP' which appears to be a typo; it should likely be 'T' (or the sum should run over j=1..T). Please correct.
  6. [Figure 1] The figure contains a Chinese question and an English reasoning trace; the fonts are small and the layout is cramped. Consider enlarging the text or splitting the figure for readability.
  7. [Section 4.4, Figures 3 and 4] In Figure 4, the y-axis labels appear duplicated; please check the axis range and ensure the heatmap is legible. In Figure 3, the bar values are hard to read; consider increasing label size.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: headline results are held-out human-scored evaluations; no fitted parameter is repackaged as a prediction.

full rationale

The paper's core claim, a reduction in hallucination rates by up to 62%, is an empirical measurement on a manually scored held-out test set of 500 questions selected from CRAFT and translated by the authors. The reported hallucination-free rates in Table 1 are human judgments of model outputs; they are not computed from any fitted parameter or from the training objective by construction. The contrastive loss in Eq. (3) maximizes cosine similarity between aligned sentence pairs, and Section 4.3 reports cosine similarity on a separate 5,000-pair test split, which is a standard held-out generalization check rather than a circular prediction. No uniqueness theorem from the authors' prior work is invoked to force the method choice, and the citations to CRAFT and SEA-Eval are data resources rather than load-bearing self-citations. Author overlap on these datasets does not make the measured improvements definitional. The absence of inter-annotator agreement and significance testing is a validity and reproducibility limitation, not a circular-derivation flaw. Accordingly, no circular step is identified.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical rather than derived, so the ledger captures the assumptions needed for the experiments to support it. The main burdens are the XCoT transfer hypothesis, the quality and non-leakage of the internal parallel data, the validity of GPT-4o-generated CoT annotations, and the reliability of the manual test set. No new physical or ontological entities are introduced.

free parameters (3)
  • contrastive temperature tau
    Eq. 3 uses an InfoNCE temperature that controls alignment sharpness; its value is hand-chosen and never reported, so the central training configuration is incomplete.
  • LoRA rank, alpha, and learning rate
    Section 4.1 says LoRA is used but gives no ranks, alphas, learning rates, batch sizes, or step counts, leaving the exact fine-tuning recipe unspecified.
  • continued-pretraining layer ranges = 9-19 (Gemma), 11-22 (LLaMA)
    Section 4.1 chooses these layer ranges by hand, and Section 4.5 later reports that mid-level tuning performs best; this is a design choice rather than a fitted constant, but it is used to produce the main results.
assumptions (5)
  • domain assumption InfoNCE contrastive loss with bilingual sentence pairs as positives aligns semantic spaces in a decoder-only LM
    Section 3.2, Eq. 3: the paper assumes this standard contrastive objective transfers from encoder models to causal decoders at sequence level, which is plausible but not proven by the paper.
  • domain assumption Reasoning in English, then answering in the target language improves factual accuracy in low-resource languages
    Section 3.3: XCoT's three-step design assumes English reasoning traces carry non-language-specific knowledge that survives into the target-language answer.
  • ad hoc to paper The self-collected EN-MS and EN-TA parallel corpora are high quality and non-leaking
    Section A.3: 100K sentence pairs and 10K paragraph pairs for Malay and Tamil are 'collected internally by our team' and not released; the paper provides no validation or leakage checks.
  • domain assumption ChatGPT-4o-generated CoT annotations manually verified are a valid training signal for factual QA
    Section A.3: 3,000 CoT QA pairs per language were generated via 3-shot prompting and manual verification is claimed, but no verification statistics are reported.
  • ad hoc to paper The 500 manually selected CRAFT test QA pairs with translated Tamil are representative and unambiguous for hallucination measurement
    Section 4.1: the test selection protocol and translation validation details are not specified, and no inter-annotator agreement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation." pith.science (2026). https://pith.science/paper/L2UBU3SU

@misc{pith2026250714239,
  author       = {Pith},
  title        = {Pith review of: CCL-XCoT: An Efficient Cross-Lingual Knowledge Transfer Method for Mitigating Hallucination Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L2UBU3SU}},
  note         = {Machine review of arXiv:2507.14239}
}
read the original abstract

Multilingual Large Language Models(MLLMs) demonstrate strong generalization across languages, yet they remain prone to hallucinations, especially in low-resource languages, due to training data imbalances. These hallucinations, which include inaccurate or fabricated outputs, are particularly problematic in domain-specific generation tasks (Chataigner et al., 2024). To address this challenge, we propose CCL-XCoT(Curriculum-based Contrastive Learning-based Cross-lingual Chain-of-Thought), a two-stage fine-tuning framework for mitigating hallucination in MLLMs. Our approach first enhances cross-lingual semantic alignment through curriculum-based contrastive learning combined with next-token prediction during continued pre-training. Building on this foundation, we then introduce a cross-lingual Chain-of-Thought (XCoT) prompting strategy during instruction fine-tuning, which guides the model to reason in a high-resource language before generating answers in the target low-resource language. Experimental results show that CCL-XCoT reduces hallucination rates by up to 62% and substantially improves factual knowledge transfer across language pairs, without relying on external retrieval or multi-model ensembles.

Figures

Figures reproduced from arXiv: 2507.14239 by the authors.

Figure 1
Figure 1. Overview of our proposed framework. The model is pre-trained with a combination of next-token [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Semantic space alignment across pre-trained [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Accuracy comparison on Cross-MMLU and Cross-LogiQA across four languages using Gemma-7B and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Answer consistency heatmap across language pairs for Gemma-7B and Gemma-7B-CCL. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Hallucination-free rates by language across [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 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]

    Orlando Ayala and Patrice Bechard. 2024. https://doi.org/10.18653/v1/2024.naacl-industry.19 Reducing hallucination in structured outputs via retrieval-augmented generation . In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 6: Industry Track), page 228–...

    doi:10.18653/v1/2024.naacl-industry.19 2024

  4. [4]

    xcot: Cross-lingual instruction tuning for cross-lingual chain-of-thought reasoning, 2024

    Linzheng Chai, Jian Yang, Tao Sun, Hongcheng Guo, Jiaheng Liu, Bing Wang, Xiannian Liang, Jiaqi Bai, Tongliang Li, Qiyao Peng, et al. xcot: Cross-lingual instruction tuning for cross-lingual chain-of-thought reasoning, 2024. URL https://arxiv. org/abs/2401, 7037

    work page 2024

  5. [5]

    Cléa Chataigner, Afaf Taïk, and Golnoosh Farnadi. 2024. https://arxiv.org/abs/2410.18270 Multilingual hallucination gaps in large language models . Preprint, arXiv:2410.18270

    arXiv 2024

  6. [6]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116

    arXiv 2019

  7. [7]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171--4186

    2019

  8. [9]

    Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023 b . Chain-of-verification reduces hallucination in large language models. arXiv preprint arXiv:2309.11495

    arXiv 2023

Show all 40 references
  1. [10]

    Suchin Gururangan, Ana Marasovi \'c , Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964

    2020 arXiv

  2. [11]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  3. [12]

    Peng Hu, Sizhe Liu, Changjiang Gao, Xin Huang, Xue Han, Junlan Feng, Chao Deng, and Shujian Huang. 2025. https://arxiv.org/abs/2406.16655 Large language models are cross-lingual knowledge-free reasoners . Preprint, arXiv:2406.16655

    2025 arXiv

  4. [13]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. https://arxiv.org/abs/2208.03299 Atlas: Few-shot learning with retrieval augmented language models . Preprint, a...

    2022 arXiv

  5. [14]

    Nihal Jain, Dejiao Zhang, Wasi Uddin Ahmad, Zijian Wang, Feng Nan, Xiaopeng Li, Ming Tan, Ramesh Nallapati, Baishakhi Ray, Parminder Bhatia, et al. 2022. Contraclm: Contrastive learning for causal language model. arXiv preprint arXiv:2210.01185

    2022 arXiv

  6. [15]

    Ziwei Ji, Tiezheng Yu, Yan Xu, Nayeon Lee, Etsuko Ishii, and Pascale Fung. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.123 Towards mitigating LLM hallucination via self reflection . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1827-...

    2023 doi

  7. [16]

    Zixuan Ke and Bing Liu. 2022. Continual learning of natural language processing tasks: A survey. arXiv preprint arXiv:2211.12701

    2022 arXiv

  8. [17]

    Deren Lei, Yaxi Li, Mengya Hu, Mingyu Wang, Vincent Yun, Emily Ching, and Eslam Kamal. 2023. https://arxiv.org/abs/2310.03951 Chain of natural language inference for reducing large language model ungrounded hallucinations . Preprint, arXiv:2310.03951

    2023 arXiv

  9. [18]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461

    2019 arXiv

  10. [19]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. https://arxiv.org/abs/2005.11401 Retrieval-augmented generation for knowledge-int...

    2021 arXiv

  11. [20]

    Johnny Li, Saksham Consul, Eda Zhou, James Wong, Naila Farooqui, Yuxin Ye, Nithyashree Manohar, Zhuxiaona Wei, Tian Wu, Ben Echols, Sharon Zhou, and Gregory Diamos. 2024. https://arxiv.org/abs/2406.17642 Banishing llm hallucinations requires rethinking generalization . Preprin...

    2024 arXiv

  12. [21]

    Zihao Li, Yucheng Shi, Zirui Liu, Fan Yang, Ali Payani, Ninghao Liu, and Mengnan Du. 2025. Language ranker: A metric for quantifying llm performance across high and low-resource languages. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 28186--28194

    2025

  13. [22]

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

    2024 arXiv

  14. [23]

    Yinhan Liu. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  15. [24]

    Yixin Liu and Pengfei Liu. 2021. Simcls: A simple framework for contrastive learning of abstractive summarization. arXiv preprint arXiv:2106.01890

    2021 arXiv

  16. [25]

    Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109--165. Elsevier

    1989

  17. [26]

    Cheng Niu, Yuanhao Wu, Juno Zhu, Siliang Xu, Kashun Shum, Randy Zhong, Juntong Song, and Tong Zhang. 2024. https://arxiv.org/abs/2401.00396 Ragtruth: A hallucination corpus for developing trustworthy retrieval-augmented language models . Preprint, arXiv:2401.00396

    2024 arXiv

  18. [27]

    Costa-jussà

    James Cross Onur Çelebi 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 Prangthip Hansanti John Hoffman Semarley Jarrett Kaush...

    2022

  19. [28]

    Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training

    2018

  20. [29]

    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

  21. [30]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2023. https://arxiv.org/abs/1910.10683 Exploring the limits of transfer learning with a unified text-to-text transformer . Preprint, arXiv:1910.10683

    2023 arXiv

  22. [31]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. https://doi.org/10.1162/tacl_a_00605 In-context retrieval-augmented language models . Transactions of the Association for Computational Linguistics, 11:1316--1331

    2023 doi

  23. [32]

    Roger Ratcliff. 1990. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97(2):285

    1990

  24. [33]

    Nils Reimers and Iryna Gurevych. 2020. https://arxiv.org/abs/2004.09813 Making monolingual sentence embeddings multilingual using knowledge distillation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Association for Computational L...

    2020 arXiv

  25. [34]

    Lisa Schut, Yarin Gal, and Sebastian Farquhar. 2025. https://arxiv.org/abs/2502.15603 Do multilingual llms think in english? Preprint, arXiv:2502.15603

    2025 arXiv

  26. [35]

    Juntong Song, Xingguang Wang, Juno Zhu, Yuanhao Wu, Xuxin Cheng, Randy Zhong, and Cheng Niu. 2024. https://doi.org/10.18653/v1/2024.emnlp-industry.113 RAG - HAT : A hallucination-aware tuning pipeline for LLM in retrieval-augmented generation . In Proceedings of the 2024 Confe...

    2024 doi

  27. [36]

    Yixuan Su, Tian Lan, Yan Wang, Dani Yogatama, Lingpeng Kong, and Nigel Collier. 2022. A contrastive framework for neural text generation. Advances in Neural Information Processing Systems, 35:21548--21561

    2022

  28. [37]

    Zhongxiang Sun, Xiaoxue Zang, Kai Zheng, Yang Song, Jun Xu, Xiao Zhang, Weijie Yu, Yang Song, and Han Li. 2025. https://arxiv.org/abs/2410.11414 Redeep: Detecting hallucination in retrieval-augmented generation via mechanistic interpretability . Preprint, arXiv:2410.11414

    2025

  29. [38]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca

    2023

  30. [39]

    Bin Wang, Geyu Lin, Zhengyuan Liu, Chengwei Wei, and Nancy F. Chen. 2024. https://arxiv.org/abs/2405.03138 Craft: Extracting and tuning cultural instructions from the wild . Preprint, arXiv:2405.03138

    2024 arXiv

  31. [40]

    Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy F Chen. 2023. Seaeval for multilingual foundation models: From cross-lingual alignment to cultural reasoning. arXiv preprint arXiv:2309.04766

    2023 arXiv

  32. [41]

    Jiaheng Wei, Yuanshun Yao, Jean-Francois Ton, Hongyi Guo, Andrew Estornell, and Yang Liu. 2024. https://arxiv.org/abs/2402.10412 Measuring and reducing llm hallucination without gold-standard answers . Preprint, arXiv:2402.10412

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.