Pith. sign in

REVIEW 4 major objections 7 minor 36 references

When Scale Meets Diversity: Evaluating Language Models on Fine-Grained Multilingual Claim Verification

T0 review · 4 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 A 270M-parameter encoder model, XLM-R, beats every tested 7B–12B LLM on fine-grained multilingual claim verification, reaching 57.7% macro-F1 versus the best LLM's 16.9%.

desk verdict Honest benchmark on X-Fact with a testable weakness: unvalidated LLM label parsing may inflate the headline gap, and one evidence-integration claim is contradicted by the paper's own table; still deserves a serious referee. read the letter →

arxiv 2507.20700 v1 pith:D3UIFHFH submitted 2025-07-28 cs.CL

Hanna Shcharbakova , Tatiana Anikina , Natalia Skachkova , Josef van Genabith This is my paper
classification cs.CL
keywords multilingualclaimverificationfine-grainedveracityclassificationX-FactXLM-RlargelanguagemodelsLoRAfine-tuningfew-shotpromptingfact-checking
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 tests whether recent large decoder-only language models are better than smaller encoder models at fine-grained multilingual claim verification, a task closer to professional fact-checking than binary true/false classification. It evaluates five models on the X-Fact dataset, which spans 25 languages and seven veracity categories, using full fine-tuning for small models and LoRA fine-tuning plus seven-shot prompting for LLMs. The central claim is that XLM-R, a 270M-parameter encoder-only model, outperforms all tested 7B–12B LLMs by a wide margin, and also improves on the previous best result by 15.8 points of macro-F1. If true, this matters for deployment: smaller specialised models may be cheaper and more reliable than general-purpose large models for multilingual fact-checking.

What carries the argument

The central object is the X-Fact benchmark: 31,189 claims in 25 languages from 11 language families, labelled with seven veracity categories (true, mostly true, partly true/misleading, mostly false, false, complicated/hard to categorize, other) and accompanied by up to five evidence snippets per claim. The decisive comparison design is encoder-based fine-tuning (both full and classification-head-only) versus decoder-only LLMs under LoRA adaptation and seven-shot prompting, with macro-F1 as the headline metric. The core evidentiary mechanism is the frozen-head XLM-R configuration: because it updates only a classification head, similar to LoRA, it helps isolate architecture and pretraining objective from training method as contributors to the gap (51.4% for frozen XLM-R versus 16.9% for the best LLM).

What would settle it

Rerun the LLM configurations with full fine-tuning, larger LoRA rank, language-specific prompts, or more extensive hyperparameter search and check whether any 7B–12B model reaches or exceeds XLM-R's 57.7% macro-F1 on the X-Fact in-domain test set; if any does, the paper's conclusion that larger decoder-only models are systematically worse on this task is contradicted.

Watch

Extended reading notes

Core claim

The paper's central finding is that on X-Fact's seven-category verification task, XLM-R base (270M parameters, encoder-only, pretrained on 100 languages) reaches 57.7% macro-F1 on the in-domain test set after full fine-tuning, compared with 41.9% for the mBERT baseline and no more than 16.9% for any tested LLM (Qwen 2.5 7B, Llama 3.1 8B, Mistral Nemo 12B) in any prompt or fine-tuning configuration. The same ordering holds on the out-of-domain and zero-shot splits, where XLM-R scores 47.6% and 43.2%, while the best LLM scores are 15.1% and 15.4% respectively. The paper further claims that LLMs show systematic difficulty in leveraging evidence, with several models performing worse on claim-plus-evidence input than on claim-only input, and that LLMs' micro-F1 scores far exceed their macro-F1 scores, indicating predictions are driven by frequent categories while rare veracity labels are poorly handled.

Load-bearing premise

The paper's headline comparison assumes that the chosen LLM setups—one lightweight adapter fine-tuning configuration and a single seven-example prompt template—are fair stand-ins for what large models can do on this task, and the authors themselves flag that their small models were trained by a fundamentally different method.

Editorial extensions

If this is right

  • If the result holds, the state of the art on X-Fact's in-domain set is 57.7% macro-F1, not 41.9%, so future multilingual verification systems should be compared against this new number.
  • Encoder-only models with masked-language-model pretraining appear to handle fine-grained multilingual categories better than decoder-only LLMs, even when the LLMs are orders of magnitude larger.
  • Adding evidence text does not automatically help LLMs on this task; claim-only configurations sometimes score higher, so evidence presentation and integration need to be redesigned for decoder-only models.
  • LLM predictions on imbalanced seven-category data are biased toward frequent labels: the large gap between micro-F1 and macro-F1 means progress requires explicit attention to rare veracity categories.
  • For practical fact-checking systems, a small fully fine-tuned model can outperform much larger models while using fewer computational resources, changing the default model-selection trade-off.

Reading between the lines

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

  • Editorial inference: the 40-point gap could shrink under stronger LLM adaptation, such as full fine-tuning or larger adapters, because the paper's frozen-XLM-R comparison shows the gap is not purely an artifact of efficient-versus-full training, but does not prove architecture alone is decisive.
  • Editorial inference: a natural next test is whether LLMs improve when evidence is presented as separately retrievable snippets, when category definitions are translated into each target language, or when prompts are optimized per language rather than shared across all 25 languages.
  • Editorial inference: the micro-versus-macro gap results imply that reporting only micro-F1 in multilingual fact-checking can hide failure on rare labels, so deployed evaluations should report per-category and per-language breakdowns.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper benchmarks two small models (XLM-R base, mT5 base) and three decoder-only LLMs (Llama 3.1 8B, Qwen 2.5 7B, Mistral Nemo 12B) on the seven-way X-Fact multilingual claim verification dataset, covering in-domain, out-of-domain, and zero-shot test subsets. With full fine-tuning, XLM-R base achieves 57.7% macro-F1 on the in-domain test set, a 15.8-point improvement over the reported mBERT baseline of 41.9%, while the best LLM configuration (Qwen 2.5 with LoRA and claim-only input) reaches 16.9%. The authors further report that LLMs often perform worse with evidence than without, and that large macro-micro F1 gaps indicate bias toward frequent categories. The LLM outputs are publicly released, and Section 8 openly lists limitations including the training-method confound and the need for future parsing analysis.

Significance. If the headline performance gap survives output-parsing validation, the result is significant for both practical deployment and research methodology: it challenges the common assumption that larger generative models are preferable for fine-grained multilingual classification, and the released LLM outputs constitute a useful public resource for follow-up error analysis. The paper is also commendable for disclosing the LoRA-versus-full-fine-tuning confound in Section 8 and for including a frozen-head XLM-R control (51.4%) that partially mitigates that confound. However, the central claim currently rests on unvalidated parsing of LLM outputs and on an evidence-integration narrative that is partly contradicted by the paper's own Table 4.

major comments (4)
  1. [§5.2 and §8 (Output Analysis and Reproducibility)] The LLM macro-F1 scores depend on parsing free-form model outputs into the seven X-Fact labels, but the paper reports no parse-success rate, label-recovery validation, or handling of unparseable outputs. Because macro-F1 averages per-class F1, a parser that systematically misfires on long or rare labels such as 'complicated/hard to categorize' or 'other' could selectively depress LLM macro-F1 and widen the macro-micro gap that underlies the frequency-bias interpretation. The released outputs make this testable; please report per-configuration parse rates and label-mapping counts, and provide a robustness check (e.g., recompute scores after excluding or manually correcting unparseable outputs). Without such a check, the 40.8-point gap is not yet established as a model-capability gap rather than an artifact of output extraction.
  2. [§6 'Evidence Integration Patterns' and Table 4] The text claims that Qwen's preference for claim-only input is 'consistent across all evaluation sets,' but Table 4 shows the opposite for Qwen LoRA fine-tuning: claim+evidence beats claim-only on OOD (15.1 vs 11.1) and zero-shot (15.4 vs 11.7). In addition, Llama 3.1 few-shot is better with evidence on all three subsets (14.0 vs 10.8 on test, 13.3 vs 8.7 on OOD, 12.9 vs 8.7 on zero-shot). The statement that LLMs show 'systematic difficulties in leveraging evidence' is therefore not supported by the reported numbers as written; the claim should be restricted to the specific model-configuration combinations that actually exhibit the pattern, or the text should be revised.
  3. [§4.1, §4.3, and §8 'Training Methodology Differences'] The central comparison confounds architecture and scale with training method: SLMs use full fine-tuning with Bayesian hyperparameter search, while LLMs use LoRA (rank 16, alpha 32) with 4-bit quantization and a single 7-shot prompt template. The frozen-head XLM-R control (51.4%) is a useful partial mitigation, but it does not establish what a fully fine-tuned 7-12B model would achieve. Since the paper's broader conclusion is that 'smaller specialized models may be more effective than general-purpose large models,' this confound is load-bearing. Please add at least one fully fine-tuned LLM run on a subset (or an ablation on LoRA rank/alpha and quantization) to test whether the gap is robust, or alternatively temper the conclusion to apply specifically to parameter-efficient adaptation with a fixed prompt.
  4. [§6 'Fine-Grained Classification Challenges' and §8] The claim of 'pronounced biases toward frequent categories' is inferred solely from the aggregate gap between macro- and micro-F1 scores, but no per-label precision, recall, or confusion matrix is reported. The authors themselves defer 'detailed per-label performance analysis' to future work. Because the macro-micro gap could also be produced by the unvalidated parsing step (see first major comment) or by a trivial majority-label baseline, the bias claim needs direct per-class evidence (e.g., a confusion matrix or per-label F1 table) before it can be assessed.
minor comments (7)
  1. [Abstract] The sentence 'achieving 57.7% macro-F1 compared to the best LLM performance of 16.9%. This represents a 15.8% improvement over the previous state-of-the-art (41.9%)' juxtaposes the LLM baseline with the improvement over mBERT; please make explicit that the 15.8-point improvement is relative to the mBERT baseline, not to the best LLM score.
  2. [§4.2] The statement that claim-only SLM experiments were omitted because 'preliminary experiments confirmed the X-Fact paper's finding that claim-only setups yield worse performance' would be more useful with the actual preliminary numbers or a reference to a supporting figure.
  3. [§5.2] In the paragraph on LoRA fine-tuning and few-shot prompting, the phrase 'from from 12.7% ... to to 15.9%' contains duplicated prepositions; please fix the typo.
  4. [Table 4 and Appendix C.2] No variance information or number of seeds is reported for any configuration; please add standard deviations or at least state the number of runs. Also specify the LoRA training epochs/steps and batch size in Appendix C.2, as these are needed for reproducibility.
  5. [Appendix C.2] The reported temperature range (0.3 to 0.5) is not sufficient; please state the exact temperature used in the final experiments for each model and configuration.
  6. [§3] Figure 1 and Figure 2 would benefit from explicit counts or percentages in the text for the 'other' label and for low-resource languages such as Serbian, since these numbers are directly relevant to the later bias discussion.
  7. [§2.2 and Table 2] Consider replacing 'state-of-the-art LLMs' with 'recent open-weight models,' since no claim of overall state-of-the-art performance on this task is established for these models.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all headline results are external benchmark measurements; the flagged limitations are validity concerns, not circular reasoning.

full rationale

The paper's central claims are empirical scores computed against the fixed, externally published X-Fact test, out-of-domain, and zero-shot splits (Gupta and Srikumar, 2021). XLM-R, mT5, and the LLM configurations are evaluated on those held-out labels; no fitted parameter is relabeled as a prediction, and no claimed result is defined in terms of the quantity it is supposed to predict. The previous state-of-the-art comparison (41.9% mBERT) is an external result from the dataset authors, not from the present authors, so no self-citation chain is load-bearing. The manuscript's own limitations are explicitly non-circular: Section 8 notes that 'our comparison involves fundamentally different training approaches' and defers 'potential parsing issues' to future work, but those are concerns about the fairness or validity of the comparison, not about a derivation that reduces to its inputs. Similarly, the apparent inconsistency between the Section 5.2/6 statement that Qwen's claim-only advantage is 'consistent across all evaluation sets' and Table 4 (where claim+evidence beats claim-only on OOD and zero-shot) is a correctness or reporting issue, not circularity. Because every headline result is an external benchmark measurement and no circular step can be exhibited, the circularity score is 0.

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

This is an empirical benchmark paper with no theoretical model or invented entities. Its conclusions rest on dataset validity, metric choice, training-methodology equivalence, and comparability of the prior baseline. The authors explicitly flag the training-methodology assumption in Section 8, but the other assumptions are not independently verified.

free parameters (5)
  • LoRA rank = 16
    Chosen after systematic experimentation (Appendix C.2); determines adapter capacity for LLM fine-tuning and directly affects the LLM performance comparison.
  • LoRA alpha = 32
    Chosen alongside rank; controls the scaling of LoRA updates and affects adaptation strength.
  • 4-bit quantization = enabled
    Used for all LLM training and inference; may reduce LLM capabilities relative to full precision, potentially impacting the headline gap.
  • Temperature = 0.3-0.5
    Tuned for few-shot prompting; affects the randomness and correctness of LLM outputs.
  • Context window length = 2048/4096 for LLMs, 256-512 for SLMs
    Treated as a hyperparameter; the paper states evidence was fully accommodated, but window size could still affect LLM attention and evidence integration.
assumptions (4)
  • domain assumption X-Fact's seven veracity labels and evidence snippets are a valid benchmark for fine-grained multilingual fact verification.
    The paper relies on X-Fact as ground truth (Section 3) and cites Gupta & Srikumar (2021) for dataset quality, without independent validation of label reliability or evidence sufficiency.
  • domain assumption Macro-F1 is the appropriate primary metric for comparing models on this imbalanced 7-class task.
    Macro-F1 is used throughout Section 5 and is justified by label imbalance, but the paper does not show whether conclusions change under other metrics such as balanced accuracy or per-class F1.
  • domain assumption LoRA fine-tuning and 4-bit quantization represent a fair operationalization of LLM capability for this task.
    The LLM-vs-SLM comparison uses LoRA for LLMs and full fine-tuning for SLMs. The authors flag this in Section 8 as a potential confound, but it is load-bearing for the central claim.
  • domain assumption The mBERT baseline of 41.9% is directly comparable to the new runs.
    The claimed 15.8% improvement is computed against the Gupta & Srikumar (2021) number, which used a different model and possibly different hyperparameters and random seeds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Scale Meets Diversity: Evaluating Language Models on Fine-Grained Multilingual Claim Verification." pith.science (2026). https://pith.science/paper/D3UIFHFH

@misc{pith2026250720700,
  author       = {Pith},
  title        = {Pith review of: When Scale Meets Diversity: Evaluating Language Models on Fine-Grained Multilingual Claim Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3UIFHFH}},
  note         = {Machine review of arXiv:2507.20700}
}
read the original abstract

The rapid spread of multilingual misinformation requires robust automated fact verification systems capable of handling fine-grained veracity assessments across diverse languages. While large language models have shown remarkable capabilities across many NLP tasks, their effectiveness for multilingual claim verification with nuanced classification schemes remains understudied. We conduct a comprehensive evaluation of five state-of-the-art language models on the X-Fact dataset, which spans 25 languages with seven distinct veracity categories. Our experiments compare small language models (encoder-based XLM-R and mT5) with recent decoder-only LLMs (Llama 3.1, Qwen 2.5, Mistral Nemo) using both prompting and fine-tuning approaches. Surprisingly, we find that XLM-R (270M parameters) substantially outperforms all tested LLMs (7-12B parameters), achieving 57.7% macro-F1 compared to the best LLM performance of 16.9%. This represents a 15.8% improvement over the previous state-of-the-art (41.9%), establishing new performance benchmarks for multilingual fact verification. Our analysis reveals problematic patterns in LLM behavior, including systematic difficulties in leveraging evidence and pronounced biases toward frequent categories in imbalanced data settings. These findings suggest that for fine-grained multilingual fact verification, smaller specialized models may be more effective than general-purpose large models, with important implications for practical deployment of fact-checking systems.

Figures

Figures reproduced from arXiv: 2507.20700 by the authors.

Figure 1
Figure 1. Distribution of data in X-Fact by label across [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Distribution of data in X-Fact by language across subsets. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Details of the X-Fact dataset. Examples from [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Macro-F1 scores across test subset by model. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Macro-F1 scores across OOD subset by model. E Prompt Template In [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Macro-F1 scores across zero-shot subset by [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Prompt template used for LLMs [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparison of macro- and micro F1 scores across LLMs, evaluation subsets, and training methods. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Comparison of average macro- and micro-F1 scores across LLMs for each evaluation subset. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 16 canonical work pages

  1. [1]

    Giorgio Barnab \`o , Federico Siciliano, Carlos Castillo, Stefano Leonardi, Preslav Nakov, Giovanni Da San Martino, and Fabrizio Silvestri. 2022. Fbmultilingmisinfo: Challenging large-scale multilingual benchmark for misinformation detection. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1--8. IEEE

    work page 2022

  2. [2]

    Han Cao, Lingwei Wei, Mengyang Chen, Wei Zhou, and Songlin Hu. 2023. https://arxiv.org/abs/2311.17355 Are large language models good fact checkers: A preliminary study . Preprint, arXiv:2311.17355

    arXiv 2023

  3. [3]

    Recep Firat Cekinel, Pinar Karagoz, and C a g r C \"o ltekin. 2024. https://aclanthology.org/2024.lrec-main.368/ Cross-lingual learning vs. low-resource fine-tuning: A case study with fact-checking in T urkish . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pag...

    work page 2024

  4. [4]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. https://doi.org/10.18653/v1/2020.acl-main.747 Unsupervised cross-lingual representation learning at scale . In Proceedings of the 58th Annual Meeting of the Association for Comp...

    doi:10.18653/v1/2020.acl-main.747 2020

  5. [5]

    Michael Han Daniel Han and Unsloth team. 2023. http://github.com/unslothai/unsloth Unsloth

    2023

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

    doi:10.18653/v1/n19-1423 2019

  7. [7]

    Alphaeus Dmonte, Roland Oruche, Marcos Zampieri, Prasad Calyam, and Isabelle Augenstein. 2024. Claim verification in the age of large language models: A survey. arXiv preprint arXiv:2408.14317

    arXiv 2024

  8. [8]

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

    arXiv 2024

Show all 36 references
  1. [9]

    Yi R Fung, Kung-Hsiang Huang, Preslav Nakov, and Heng Ji. 2022. The battlefront of combating misinformation and coping with media bias. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4790--4791

    2022

  2. [10]

    Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos. 2022. https://doi.org/10.1162/tacl_a_00454 A survey on automated fact-checking . Transactions of the Association for Computational Linguistics, 10:178--206

    2022 doi

  3. [11]

    Ashim Gupta and Vivek Srikumar. 2021. https://doi.org/10.18653/v1/2021.acl-short.86 X -fact: A new benchmark dataset for multilingual fact checking . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Con...

    2021 doi

  4. [12]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. https://openreview.net/forum?id=XPZIaotutsD Deberta: Decoding-enhanced bert with disentangled attention . In International Conference on Learning Representations

    2021

  5. [13]

    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. 2022 a . http://dblp.uni-trier.de/db/conf/iclr/iclr2022.html#HuSWALWWC22 Lora: Low-rank adaptation of large language models. In ICLR. OpenReview.net

    2022

  6. [14]

    Yu, and Zhijiang Guo

    Xuming Hu, Junzhe Chen, Xiaochuan Li, Yufei Guo, Lijie Wen, Philip S. Yu, and Zhijiang Guo. 2023. https://arxiv.org/abs/2310.05177 Do large language models know about facts? Preprint, arXiv:2310.05177

    2023 arXiv

  7. [15]

    Xuming Hu, Zhijiang Guo, GuanYu Wu, Aiwei Liu, Lijie Wen, and Philip Yu. 2022 b . https://doi.org/10.18653/v1/2022.naacl-main.246 CHEF : A pilot C hinese dataset for evidence-based fact-checking . In Proceedings of the 2022 Conference of the North American Chapter of the Assoc...

    2022 doi

  8. [16]

    Zi-Hang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. 2020. https://proceedings.neurips.cc/paper/2020/file/96da2f590cd7246bbde0051047b0d6f7-Paper.pdf Convbert: Improving bert with span-based dynamic convolution . In Advances in Neural Information...

    2020

  9. [17]

    Yichuan Li, Bohan Jiang, Kai Shu, and Huan Liu. 2020. https://arxiv.org/abs/2011.04088 Mm-covid: A multilingual and multimodal data repository for combating covid-19 disinformation . Preprint, arXiv:2011.04088

    2020 arXiv

  10. [18]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  11. [19]

    Mistral AI Team . 2024. Mistral nemo. https://mistral.ai/en/news/mistral-nemo. Accessed: 14-Feb-2025

    2024

  12. [20]

    Salar Mohtaj, Ata Nizamoglu, Premtim Sahitaj, Vera Schmitt, Charlott Jakob, and Sebastian M\" o ller. 2024. https://doi.org/10.1145/3643491.3660290 Newspolyml: Multi-lingual european news fake assessment dataset . In Proceedings of the 3rd ACM International Workshop on Multime...

    2024

  13. [21]

    Jeppe N rregaard and Leon Derczynski. 2021. https://aclanthology.org/2021.nodalida-main.47/ D an FEVER : claim verification dataset for D anish . In Proceedings of the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa), pages 422--428, Reykjavik, Iceland (Online). ...

    2021

  14. [22]

    Kellin Pelrine, Anne Imouza, Camille Thibault, Meilina Reksoprodjo, Caleb Gupta, Joel Christoph, Jean-Fran c ois Godbout, and Reihaneh Rabbany. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.395 Towards reliable misinformation mitigation: Generalization, uncertainty, and GP...

    2023 doi

  15. [23]

    Mat \'u s Pikuliak, Ivan Srba, Robert Moro, Timo Hromadka, Timotej Smole n , Martin Meli s ek, Ivan Vykopal, Jakub Simko, Juraj Podrou z ek, and Maria Bielikova. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.1027 Multilingual previously fact-checked claim retrieval . In Pr...

    2023 doi

  16. [24]

    Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. https://doi.org/10.18653/v1/S18-2023 Hypothesis only baselines in natural language inference . In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics,...

    2018 doi

  17. [25]

    Dietram A Scheufele and Nicole M Krause. 2019. Science audiences, misinformation, and fake news. Proceedings of the National Academy of Sciences, 116(16):7662--7669

    2019

  18. [26]

    Gautam Kishore Shahi and Durgesh Nandini. 2020. https://doi.org/10.36190/2020.14 FakeCovid- A Multilingual Cross-domain Fact Check News Dataset for COVID-19 . ICWSM

    2020 doi

  19. [27]

    Gautam Kishore Shahi, Julia Maria Stru , and Thomas Mandl. 2021. Overview of the clef-2021 checkthat! lab task 3 on fake news detection. Working Notes of CLEF

    2021

  20. [28]

    Aryan Singhal, Thomas Law, Coby Kassner, Ayushman Gupta, Evan Duan, Aviral Damle, and Ryan Luo Li. 2024. https://aclanthology.org/2024.nlp4pi-1.2 Multilingual fact-checking using LLM s . In Proceedings of the Third Workshop on NLP for Positive Impact, pages 13--31, Miami, Flor...

    2024

  21. [29]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, W...

    2023 arXiv

  22. [30]

    Herbert Ullrich, Jan Drchal, Martin Rýpar, Hana Vincourová, and Václav Moravec. 2023. https://doi.org/10.1007/s10579-023-09654-3 Csfever and ctkfacts: acquiring czech data for fact verification . Language Resources and Evaluation, 57(4):1571–1605

    2023 doi

  23. [31]

    William Yang Wang. 2017. https://doi.org/10.18653/v1/P17-2067 liar, liar pants on fire : A new benchmark dataset for fake news detection . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 422--426, Vanco...

    2017 doi

  24. [32]

    Xinyu Wang, Wenbo Zhang, and Sarah Rajtmajer. 2024. https://arxiv.org/abs/2410.18390 Monolingual and multilingual misinformation detection for low-resource languages: A comprehensive survey . Preprint, arXiv:2410.18390

    2024 arXiv

  25. [33]

    Chi, Quoc V

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processi...

    2022

  26. [34]

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. https://doi.org/10.18653/v1/2021.naacl-main.41 m T 5: A massively multilingual pre-trained text-to-text transformer . In Proceedings of the 2021 Conferenc...

    2021 doi

  27. [35]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  28. [36]

    Caiqi Zhang, Zhijiang Guo, and Andreas Vlachos. 2024. https://aclanthology.org/2024.emnlp-main.113 Do we need language-specific fact-checking models? the case of C hinese . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1899--1...

    2024

Pith tools

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