Pith. sign in

REVIEW 5 major objections 5 minor 36 references

Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Training small LLMs with a curriculum of synthetic hallucinated answers lets a 3B detector rival GPT-4o at spotting lies.

desk verdict Plausible recipe, sloppy manuscript: three incompatible descriptions of the curriculum undercut the headline claim, but the idea is worth a serious referee. read the letter →

arxiv 2505.17558 v1 pith:OZIN3JT2 submitted 2025-05-23 cs.CL cs.AI

Shrey Pandit , Ashwin Vinod , Liu Leqi , Ying Ding This is my paper
classification cs.CLcs.AI
keywords hallucinationdetectiondirectpreferenceoptimizationcurriculumlearningsyntheticnegativesMiniCheckgroundingzero-shottransfersmalllanguagemodelsfactverification
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

This paper argues that the quality of the rejected examples matters more than the size of the model for aligning a hallucination detector. Instead of using ordinary failed generations as negatives in Direct Preference Optimization, the authors use carefully engineered hallucinated answers scored by an external fact-checker, and feed them to the model in easy-to-hard order. If the claim holds, small 1B–3B detectors can match or beat much larger models such as GPT-4o on medical and open-domain QA benchmarks, and can transfer zero-shot to unseen datasets, making reliable hallucination checking cheap.

What carries the argument

The central object is the MiniCheck grounding probability $p_l$, computed by asking the fact-checker how well a hallucinated answer is supported by its context; lower $p_l$ means the hallucination is easier to catch because it is poorly grounded. The paper sorts all hallucinated negatives by ascending $p_l$, discards the very easy ones below 0.25, splits the rest into S bins, and runs DPO stage-wise on each bin, so the model sees progressively harder lies. The DPO loss in Eq. (1) is the standard preference optimization objective with true answers as chosen and hallucinated answers as rejected, with the trust-region parameter set to 0.1 and LoRA adapters on the query, key, value, and output projections.

What would settle it

Compute the correlation between MiniCheck's grounding probability and human-annotated difficulty of the hallucinated answers in MedHallu; if the correlation is near zero, the curriculum is not actually ordering by difficulty and the reported gains would have to come from the curated negative samples alone, not from the curriculum. A second direct test is to replace MiniCheck ordering with random ordering of the same filtered negatives on a held-out split and measure whether the curriculum's F1 advantage disappears.

Watch

Extended reading notes

Core claim

HaluCheck trains Llama-3.2 1B and 3B backbones with LoRA under the standard DPO objective, using gold answers as chosen completions and curated hallucinated answers as rejected completions. Each hallucination is scored by MiniCheck's grounding probability $p_l$; samples with $p_l < 0.25$ are discarded, and the rest are sorted by ascending $p_l$ and split into stages so the model first learns to reject obvious falsehoods and then progressively subtler ones. The authors report that this recipe yields F1 0.759 on MedHallu and 0.753 on HaluEval for the 3B model, outperforming the backbone and several larger open models while staying competitive with GPT-4o, and that the 3B model transfers zero-shot to DROP, CovidQA, and PubMedQA with higher average accuracy than both Llama-3.2 3B and GPT-3.5-Turbo.

Load-bearing premise

The load-bearing premise is that MiniCheck's grounding probability is an accurate and sufficient measure of how hard a hallucinated answer is to detect, and that discarding samples below 0.25 plus sorting the rest by ascending $p_l$ creates the intended easy-to-hard curriculum; no independent check of difficulty is used.

Editorial extensions

If this is right

  • HaluCheck 3B improves up to 24% over its Llama-3.2 3B backbone on difficult benchmarks like MedHallu and HaluEval, with F1 0.759 and 0.753 respectively.
  • HaluCheck 3B outperforms GPT-3.5-Turbo on CovidQA accuracy (62.50 vs 56.70) and beats it on average across DROP, CovidQA, and PubMedQA in a strict zero-shot setting.
  • Curriculum ordering of the same curated negatives beats random selection: on MedHallu F1 rises from 0.528 to 0.664 for the 1B model and from 0.694 to 0.759 for the 3B model.
  • Curated hallucinated negatives outperform standard failed generations as DPO negatives, giving MedHallu F1 0.729 vs 0.691 for the 3B model.
  • Training jointly on MedHallu and HaluEval DPO sets preserves cross-dataset transfer, whereas training on either dataset alone overfits and degrades the other benchmark.

Reading between the lines

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

  • The same curriculum-DPO recipe could be iterated: HaluCheck's own detections could be re-scored by MiniCheck and fed back as harder negatives, turning the verifier and detector into a co-training loop without new data.
  • Because training is binary and QA-focused, extending the recipe to span-level annotations should let the curriculum teach partial hallucination detection, a fine-grained capability the paper explicitly sets aside.
  • If MiniCheck's grounding score is the bottleneck, swapping in a stronger fact-checker should directly raise the ceiling of the curriculum, since difficulty ordering and negative filtering both depend on that score.
  • The zero-shot gains suggest HaluCheck learns a general 'lie detector' skill, but the paper's own limitation note warns this could instead be overfitting to adversarial patterns in MedHallu and HaluEval; testing on out-of-distribution non-QA domains would separate the two.
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

5 major / 5 minor

Summary. The paper proposes HaluCheck, a family of 1B and 3B Llama-3.2 models aligned with Direct Preference Optimization (DPO), where the negative examples are hallucinated answers scored by MiniCheck grounding probability and presented in a claimed easy-to-hard curriculum. The method is evaluated on MedHallu and HaluEval, with zero-shot transfer to DROP, CovidQA, and PubMedQA, and is compared with larger open and closed models. The main claimed contributions are the curriculum-based sampling strategy and the demonstration that small aligned detectors can outperform much larger baselines.

Significance. If the results are reproducible, this is a practically useful contribution: it shows that small open models can be turned into strong hallucination detectors by using an external fact-checker to rank negatives, and the paper includes informative ablations over training-set composition, random-vs-curriculum ordering, and difficulty cutoffs. The Limitations section is honest about the reliance on MiniCheck and the QA-only scope. However, inconsistencies in the description of the curriculum and in the reported numbers mean that the central claim is not yet established.

major comments (5)
  1. [Algorithm 1 / Section C.2 / Section D] The paper gives three incompatible descriptions of the curriculum. Algorithm 1 (lines 14-19) partitions the data by p_l into S bins and optimizes stage-by-stage, so hard examples are withheld until later stages; Section C.2 instead states that "DPO training proceeds batch wise on the sorted data for four epochs, with all batches trained per epoch"; Section D specifies "total epochs = 25" and "sequential sampling". Since no code is provided, it is impossible to tell whether the actual run was the stage-wise curriculum claimed in the abstract or merely sorted-batch DPO in which hard examples are seen from the first epoch. This is load-bearing because 'curriculum DPO' is the headline contribution.
  2. [Section 5.1 / Tables 1, 4, 5, 7] Reported numbers for the same model and benchmark disagree. HaluCheck 3B MedHallu F1 is 0.759 in Table 1, 0.759 in Table 4 and 75.90 in Table 5, but 0.729 in Table 7; HaluCheck 1B MedHallu F1 is 0.664 in Table 1 but 0.711 in Section 5.1. The reader cannot tell which run is the final model, and the abstract's "24%" improvement is not traceable to a specific table.
  3. [Section 3 / Section 4] The difficulty ordering is described inconsistently. Section 3 says "Lower pl indicates easier hallucination cases, ensuring the curriculum starts with easy (high-grounding) and gradually moves to harder ones"; Section 4 says negatives are sorted by "decreasing MiniCheck confidence drop"; Algorithm 1 and Section C.2 sort by ascending p_l. Low p_l cannot simultaneously be "easy" and "high-grounding" in the same ordering. This ambiguity affects which samples are seen first and whether the intended curriculum was actually implemented.
  4. [Section A.2 / Table 5] The random-vs-curriculum ablation cannot establish the staged-curriculum mechanism if training was in fact sorted-batch DPO repeated for all epochs. Under the Section C.2/D description, the curriculum arm differs from the random arm only in within-epoch batch order, and hard examples are seen from epoch 1; under the Algorithm 1 description, exposure is staged. The paper does not state which description generated Table 5, so the abstract's claim that the gains come from a curriculum is not supported.
  5. [Section 3 / Limitations] The entire difficulty signal comes from MiniCheck's p_l, but the paper provides no validation that p_l ranks perceptual difficulty in a way that corresponds to learning order. The Limitations section acknowledges reliance on MiniCheck, but the curriculum claim requires p_l to be an adequate difficulty measure. A concrete test would be to correlate p_l with an independent fact-checker or human judgments, or to show that staged ordering by p_l outperforms random ordering of the same samples under staged training.
minor comments (5)
  1. [Abstract / Table 1] The abstract says "upto" instead of "up to", and Table 1 uses the spelling "HalluCheck-Llama" where the rest of the paper uses "HaluCheck".
  2. [Section 3] There is an incomplete sentence: "For each sample i we define Let x(i) denote...".
  3. [References] The references for Bengio et al., Elman, and Zhang et al. are duplicated as a/b variants.
  4. [Section C.2 / Section 5.3] Section C.2 refers to "Table 6 in the main paper" and Section 5.3 refers to "Table 7", but both tables appear only in the appendix.
  5. [Table 5] Table 5 reports F1 in percentage units (e.g., 75.90) while Tables 1 and 4 use decimals (0.759); the units should be unified.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: MiniCheck difficulty scoring, DPO training, and benchmark evaluation are externally grounded; the coauthored MedHallu benchmark is a minor, non-load-bearing self-citation.

full rationale

The claimed derivation chain is: (i) MiniCheck (Tang et al., 2024, external to the authors) assigns grounding probabilities to hallucinated negatives; (ii) negatives are filtered (p_l >= 0.25) and sorted by p_l; (iii) standard DPO fine-tunes Llama-3.2 backbones; (iv) HaluCheck models are scored on MedHallu pqa_labeled, the HaluEval test partition, and zero-shot DROP/CovidQA/PubMedQA. No step is defined in terms of the downstream metric: MiniCheck is not fitted to the MedHallu/HaluEval labels, the DPO objective (Eq. 1) is the standard fixed-reference loss, and training/evaluation splits are disjoint (9,000 MedHallu pqa_artificial + 8,000 HaluEval train vs. 1,000 pqa_labeled + 2,000 HaluEval test). The coauthored MedHallu benchmark (Pandit et al., 2025) is a self-citation and supplies both training negatives and one evaluation set, but the central claim also rests on external HaluEval and zero-shot results, so the citation is not load-bearing. Table 3 uses MiniCheck scores only descriptively, and the genuine DPO head-to-head (Table 7) independently confirms the negative-sample choice. Two concerns are flagged for the correctness pass rather than as circularity: (a) Table 6 selects the p_l cutoff that maximizes test F1, a model-selection-on-test concern; (b) the curriculum is described inconsistently — Algorithm 1 is stage-wise over S bins, Section C.2 says batch-wise training on sorted data for four epochs with all batches per epoch, Section D says 25 total epochs with sequential sampling, Section 4 sorts by 'decreasing MiniCheck confidence drop', and Section 3 conflates low p_l with 'high-grounding'. These are reproducibility/soundness holes; nothing in the paper makes a prediction or result equivalent, by construction or by fitted-input renaming, to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on DPO's preference objective, on MiniCheck's grounding scores as an ordering and filtering signal, and on the premise that curated hallucinated benchmark answers are superior DPO negatives. No new theoretical entity is introduced, but the method depends on data-tuned hyperparameters and an under-specified curriculum staging.

free parameters (3)
  • MiniCheck grounding cutoff = 0.25
    Chosen from ablations in Table 6; Section C.2 states the 0.25-1.0 range yields the highest F1, so the filter is tuned to benchmark performance.
  • Curriculum stage count S / bin construction = not specified
    Algorithm 1 requires S stages, but Section C.2 only describes sorted batches over epochs with no stage count, leaving the curriculum under-specified.
  • DPO beta = 0.1
    Fixed in Section D with no sensitivity analysis; the loss in Eq. (1) and therefore the training behavior depend on it.
assumptions (4)
  • standard math DPO's preference optimization objective is valid for the chosen/rejected pairs used here (Rafailov et al. 2023).
    Invoked in Eq. (1); the method inherits the Bradley-Terry preference assumption underlying DPO.
  • domain assumption MiniCheck grounding probability p_l is a valid and accurate measure of hallucination quality and difficulty.
    Used in Sections 3 and C.2 to score, filter, and order all negatives; if MiniCheck is biased, the curriculum and negative selection break.
  • domain assumption Hallucinated answers from MedHallu and HaluEval are better DPO negatives than standard failed generations.
    Core to the method; Table 3 supports this only for MedHallu means and medians, with no independent validation across domains.
  • domain assumption Binary classification of whole answers as hallucinated or factual is an adequate task formulation for hallucination detection.
    The authors' Limitations admit partial and span-level hallucinations are not covered, so the claim is scoped to binary QA detection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection." pith.science (2026). https://pith.science/paper/OZIN3JT2

@misc{pith2026250517558,
  author       = {Pith},
  title        = {Pith review of: Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZIN3JT2}},
  note         = {Machine review of arXiv:2505.17558}
}
read the original abstract

Aligning large language models (LLMs) to accurately detect hallucinations remains a significant challenge due to the sophisticated nature of hallucinated text. Recognizing that hallucinated samples typically exhibit higher deceptive quality than traditional negative samples, we use these carefully engineered hallucinations as negative examples in the DPO alignment procedure. Our method incorporates a curriculum learning strategy, gradually transitioning the training from easier samples, identified based on the greatest reduction in probability scores from independent fact checking models, to progressively harder ones. This structured difficulty scaling ensures stable and incremental learning. Experimental evaluation demonstrates that our HaluCheck models, trained with curriculum DPO approach and high quality negative samples, significantly improves model performance across various metrics, achieving improvements of upto 24% on difficult benchmarks like MedHallu and HaluEval. Additionally, HaluCheck models demonstrate robustness in zero-shot settings, significantly outperforming larger state-of-the-art models across various benchmarks.

Figures

Figures reproduced from arXiv: 2505.17558 by the authors.

Figure 1
Figure 1. Illustration of the qualitative difference be [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure showing the pipeline for selecting high [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure showing the grounded factuality of the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 8 canonical work pages

  1. [1]

    Deema Alnuhait, Neeraja Kirtane, Muhammad Khalifa, and Hao Peng. 2024. Factcheckmate: Preemptively detecting and mitigating hallucinations in lms. arXiv preprint arXiv:2410.02899

    arXiv 2024

  2. [2]

    Anusha Balakrishnan, Jinfeng Rao, Kartikeya Upasani, Michael White, and Rajen Subba. 2019. https://doi.org/10.18653/v1/P19-1080 Constrained decoding for neural NLG from compositional representations in task-oriented dialogue . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 831--844, Florence, Italy. Assoc...

    doi:10.18653/v1/p19-1080 2019

  3. [3]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009 a . Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41--48

    work page 2009

  4. [4]

    Yoshua Bengio, J \'e r \^o me Louradour, Ronan Collobert, and Jason Weston. 2009 b . Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41--48

    work page 2009

  5. [5]

    Monojit Choudhury, Kalika Bali, Sunayana Sitaram, and Ashutosh Baheti. 2017. Curriculum design for code-switching: Experiments with language identification and language modeling with deep neural networks. In Proceedings of the 14th International Conference on Natural Language Processing (ICON-2017), pages 65--74

    work page 2017

  6. [6]

    Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. Drop: A reading comprehension benchmark requiring discrete reasoning over paragraphs. arXiv preprint arXiv:1903.00161

    arXiv 2019

  7. [7]

    Esin Durmus, He He, and Mona Diab. 2020. https://doi.org/10.18653/v1/2020.acl-main.454 FEQA : A question answering evaluation framework for faithfulness assessment in abstractive summarization . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5055--5070, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.454 2020

  8. [8]

    Jeffrey L Elman. 1993 a . Learning and development in neural networks: The importance of starting small. Cognition, 48(1):71--99

    work page 1993

Show all 36 references
  1. [9]

    Jeffrey L Elman. 1993 b . Learning and development in neural networks: The importance of starting small. Cognition, 48(1):71--99

    1993

  2. [10]

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

    2022

  3. [11]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. https://doi.org/10.1145/3571730 Survey of hallucination in natural language generation . ACM Computing Surveys, 55(12):1–38

    2023 doi

  4. [12]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W Cohen, and Xinghua Lu. 2019. Pubmedqa: A dataset for biomedical research question answering. arXiv preprint arXiv:1909.06146

    2019 arXiv

  5. [13]

    Madhava Krishna, Bhagesh Gaur, Arsh Verma, and Pankaj Jalote. 2024. https://doi.org/10.1109/RE59067.2024.00056 Using llms in software requirements specifications: An empirical evaluation . In 2024 IEEE 32nd International Requirements Engineering Conference (RE), pages 475--483

    2024

  6. [14]

    Jinqi Lai, Wensheng Gan, Jiayang Wu, Zhenlian Qi, and Philip S Yu. 2024. Large language models in law: A survey. AI Open

    2024

  7. [15]

    Katherine Lee, Orhan Firat, Ashish Agarwal, Clara Fannjiang, and David Sussillo. 2019. https://openreview.net/forum?id=SkxJ-309FQ Hallucinations in neural machine translation

    2019

  8. [16]

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. https://arxiv.org/abs/2305.11747 Halueval: A large-scale hallucination evaluation benchmark for large language models . Preprint, arXiv:2305.11747

    2023 arXiv

  9. [17]

    Tianyu Liu, Yizhe Zhang, Chris Brockett, Yi Mao, Zhifang Sui, Weizhu Chen, and Bill Dolan. 2022. https://arxiv.org/abs/2104.08704 A token-level reference-free hallucination detection benchmark for free-form text generation . Preprint, arXiv:2104.08704

    2022 arXiv

  10. [18]

    Potsawee Manakul, Adian Liusie, and Mark JF Gales. 2023. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. arXiv preprint arXiv:2303.08896

    2023 arXiv

  11. [19]

    Meta. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

    2024 arXiv

  12. [20]

    Timo M \"o ller, Anthony Reina, Raghavan Jayakumar, and Malte Pietsch. 2020. Covid-qa: A question answering dataset for covid-19. In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL 2020

    2020

  13. [21]

    OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024....

    2024 arXiv

  14. [22]

    Shrey Pandit, Jiawei Xu, Junyuan Hong, Zhangyang Wang, Tianlong Chen, Kaidi Xu, and Ying Ding. 2025. https://arxiv.org/abs/2502.14302 Medhallu: A comprehensive benchmark for detecting medical hallucinations in large language models . Preprint, arXiv:2502.14302

    2025 arXiv

  15. [23]

    Pulkit Pattnaik, Rishabh Maheshwary, Kelechi Ogueji, Vikas Yadav, and Sathwik Tejaswi Madhusudhan. 2024. Curry-dpo: Enhancing alignment using curriculum learning & ranked preferences. arXiv preprint arXiv:2403.07230

    2024 arXiv

  16. [24]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728--53741

    2023

  17. [25]

    Selvan Sunitha Ravi, Bartosz Mielczarek, Anand Kannappan, Douwe Kiela, and Rebecca Qian. 2024. Lynx: An open source hallucination evaluation model. arXiv preprint arXiv:2407.08488

    2024 arXiv

  18. [26]

    Mrinmaya Sachan and Eric Xing. 2016. Easy questions first? a case study on curriculum learning for question answering. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 453--463

    2016

  19. [27]

    Thibault Sellam, Dipanjan Das, and Ankur P. Parikh. 2020. https://arxiv.org/abs/2004.04696 Bleurt: Learning robust metrics for text generation . Preprint, arXiv:2004.04696

    2020 arXiv

  20. [28]

    Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, and 1 others. 2022. Large language models encode clinical knowledge. arXiv preprint arXiv:2212.13138

    2022 arXiv

  21. [29]

    Liyan Tang, Philippe Laban, and Greg Durrett. 2024. Minicheck: Efficient fact-checking of llms on grounding documents. arXiv preprint arXiv:2404.10774

    2024 arXiv

  22. [30]

    Yi Tay, Shuohang Wang, Luu Anh Tuan, Jie Fu, Minh C Phan, Xingdi Yuan, Jinfeng Rao, Siu Cheung Hui, and Aston Zhang. 2019. Simple and effective curriculum pointer-generator networks for reading comprehension over long narratives. arXiv preprint arXiv:1905.10847

    2019 arXiv

  23. [31]

    Qwen Team. 2024. https://qwenlm.github.io/blog/qwen2.5/ Qwen2.5: A party of foundation models

    2024

  24. [32]

    Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, and Yongdong Zhang. 2020. Curriculum learning for natural language understanding. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 6095--6104

    2020

  25. [33]

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019 a . Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675

    2019 arXiv

  26. [34]

    Xuan Zhang, Pamela Shapiro, Gaurav Kumar, Paul McNamee, Marine Carpuat, and Kevin Duh. 2019 b . Curriculum learning for domain adaptation in neural machine translation. arXiv preprint arXiv:1905.05816

    2019 arXiv

  27. [35]

    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...

  28. [36]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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