Pith. sign in

REVIEW 1 major objections 5 minor 31 references

Paying Alignment Tax with Contrastive Learning

T0 review · 1 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 The paper claims that contrastive learning, trained on positive and negative examples, can reduce toxicity and improve faithfulness simultaneously across model scales—avoiding the capability degradation that existing debiasing methods…

desk verdict Useful measurement of the alignment tax, but the contrastive method's headline claim is untested because evaluation never touches the tax benchmarks. read the letter →

arxiv 2505.19327 v1 pith:AKH3ZOAG submitted 2025-05-25 cs.LG

Buse Sibel Korkmaz , Rahul Nair , Elizabeth M. Daly , Antonio del Rio Chanona This is my paper
classification cs.LG
keywords alignmenttaxcontrastivelearningbiasmitigationfaithfulnesstoxicityreductionlanguagemodeldebiasingRedditTLDRsummarizationnamedentityweighting
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

Existing debiasing methods usually trade away model capability: debiased models become less truthful, lose knowledge, or produce unintelligible text, and the cost is worst for small models. This paper claims that a contrastive learning approach can break that trade-off. By training with deliberately constructed positive and negative examples, the proposed framework lowers toxicity and raises faithfulness on Reddit TL;DR summaries at the same time across GPT-2, Phi-2, and Llama-2-7B. The authors argue this is the first method to improve both metrics consistently, and that the contradiction between bias mitigation and capability preservation is not fundamental.

What carries the argument

The load-bearing mechanism is a contrastive head attached to the base language model: each token's hidden state is projected and normalized, named-entity tokens are pooled into a single representation, and a temperature-scaled contrastive loss pulls positive summaries close to the source while pushing negative summaries away. The negative examples are not random: they are generated by backtranslation for positives and by adversarial toxic rewriting, low-confidence beam search, and entity manipulation for negatives, with the loss scaled up when toxic content is detected in the batch. This machinery is what lets the model learn a boundary between biased and unbiased content while preserving factual entities.

What would settle it

Run the contrastively debiased models from Table 4 on TruthfulQA and MMLU-Pro: if faithfulness on Reddit TL;DR rises while TruthfulQA or MMLU-Pro scores fall, the claim that the alignment tax is avoided fails. Alternatively, check whether a non-contrastive method that simply reweights the language-modeling loss with the same augmented data achieves the same simultaneous gains; if so, the contrastive head is not the load-bearing mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that the alignment tax of debiasing—the capability loss in truthfulness, knowledge retention, and coherent generation—is avoidable if debiasing is framed as contrastive learning rather than as pushing representations away from bias. The authors document that seven existing debiasing methods degrade truthfulness or knowledge in 44 to 81 percent of evaluations, with a negative correlation between model size and degradation. Their own framework, which computes a token-level contrastive loss with named-entity-weighted pooling and dynamic loss scaling when toxic content is detected, achieves simultaneous toxicity reduction and faithfulness improvement on the Reddit TL;DR summarization task for all three model sizes. They interpret this as evidence that explicit positive and negative boundary learning can resolve the bias-capability trade-off that single-objective debiasing cannot.

Load-bearing premise

The load-bearing premise is that the faithfulness detector used on Reddit TL;DR summaries measures the same capabilities—factual accuracy and knowledge retention—that the paper's Section 3 uses TruthfulQA and MMLU-Pro to define as the alignment tax; this equivalence is assumed, not validated.

Editorial extensions

If this is right

  • If the claim holds, debiasing pipelines no longer need to accept a capability tax; contrastive fine-tuning can replace or augment existing debiasing methods.
  • Small models, where the tax is most severe, would gain a practical debiasing path that does not rely on larger architectures.
  • The named-entity focus of the contrastive loss suggests faithfulness can be protected by concentrating learning on semantically load-bearing tokens rather than all tokens.
  • The dynamic toxic weight shows that batch-level toxicity signals can steer the strength of the contrastive signal, giving a simple control knob for the bias-faithfulness balance.
  • The approach's gains grow with model scale, implying that capacity still helps but is no longer a precondition for simultaneous improvement.

Reading between the lines

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

  • One testable extension is to apply the contrastive framework to other alignment objectives, such as refusal behavior or privacy, where positive and negative pairs can be constructed; the mechanism is not obviously specific to toxicity.
  • A direct comparison on TruthfulQA and MMLU-Pro after contrastive fine-tuning would connect the TL;DR faithfulness metric back to the benchmarks used to define the alignment tax in Section 3.
  • The paper leaves open whether the faithfulness detector's notion of faithfulness is the same capability measured by TruthfulQA and MMLU-Pro; if the detector only tracks entity overlap, the claim about avoiding knowledge loss is narrower than the alignment-tax framing suggests.
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

1 major / 5 minor

Summary. The paper claims that existing debiasing methods incur an 'alignment tax'—degradation in factual accuracy and knowledge retention—and proposes a contrastive learning framework that augments data with positive and negative examples, uses a named-entity-focused contrastive loss, and applies dynamic loss scaling when toxic content is detected. Section 3 documents the alignment tax across GPT2, Phi2, and Llama2-7B using TruthfulQA, RealToxicityPrompts, and MMLU-Pro. Section 5 evaluates the proposed method on Reddit TL;DR summarization, reporting simultaneous toxicity reductions (0.007–0.014) and faithfulness increases (0.018–0.285) across model scales. The central claim is that this is the first method to consistently improve both toxicity and faithfulness while avoiding the capability degradation characteristic of existing approaches.

Significance. If substantiated, the paper would make a useful contribution by demonstrating that debiasing need not trade off against factual accuracy and by providing a systematic benchmark for measuring that trade-off. The Section 3 evaluation of baselines across TruthfulQA, MMLU-Pro, and generation-pattern analyses is a potentially valuable resource. However, the headline claim—that the contrastive framework avoids the alignment tax—is not tested on the benchmarks used to define that tax. The only capability metric for the proposed method is a faithfulness detector on TL;DR summaries, and the paper provides no evidence that this detector measures the same construct as TruthfulQA or MMLU-Pro. The selection of key hyperparameters on the same evaluation set further weakens the empirical case. The contribution is therefore not yet established.

major comments (1)
  1. [§5.2, Table 4] The comparison between the contrastive method and the baselines appears confounded by fine-tuning. The contrastive method is fine-tuned on 18K TL;DR pairs (Section 5), while the paper does not state whether the CDA, Dropout, INLP, Sentence Debiasing, and Instructive Debiasing baselines in Table 4 were fine-tuned on the same TL;DR training set or evaluated zero-shot. If the baselines were not fine-tuned on the task, the comparison does not isolate the effect of the contrastive objective from the effect of task fine-tuning, and the claimed simultaneous improvement may be due to the fine-tuning itself. The exact training/evaluation protocol for each baseline must be specified, and baselines should be trained under the same protocol as the proposed method if a fair comparison is intended.
minor comments (5)
  1. [Abstract] The phrase 'often result a degradation' should be 'often result in a degradation'.
  2. [§3.2] The reference to 'MMLU-Pro benchmark (Hendrycks et al., 2021)' is inaccurate: MMLU-Pro is introduced by Wang et al. (2024); the Hendrycks et al. reference is for the original MMLU benchmark.
  3. [Table 13] The model label 'Phi-2.7B' is inconsistent with 'Phi2' used elsewhere; please standardize the nomenclature.
  4. [Appendix F] The sentence 'all with linear schedule and contrastive head activation GELU' is missing punctuation and reads awkwardly; consider revising for clarity.
  5. [§5.2] The paper mentions 'Appendix for data ablation and Appendix for the contribution of α parameter' without giving specific appendix letters or table numbers; please add explicit cross-references.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline 'simultaneous improvements' are partially fitted: alpha is selected on the same TL;DR benchmark used for Table 4, and the toxicity metric is the same classifier used to shape the training loss.

  1. fitted input called prediction [Appendix H (Contrastive loss parameter alpha); Section 5.2; Section 5.3 (Table 4)]
    "Based on these empirical results, we selected α = 4 as our default parameter across all experiments, as it provides the best balance between toxicity reduction and faithfulness preservation while maintaining stable training dynamics. ... In our all experiments, we set α = 4 and τ = 1. ... Our contrastive learning framework demonstrates substantial improvements across all model scales (Table 4), achieving simultaneous reductions in toxicity (↓0.007-0.014) and increases in faithfulness (↑0.018-0.285)."

    Appendix H tunes α on the Reddit TL;DR benchmark and reports the resulting toxicity/faithfulness scores (Table 14); Section 5.2 fixes α=4; Section 5.3 then presents Table 4, on the same TL;DR task, as evidence of 'simultaneous improvements.' Because the headline metric values are the selected optimum of an α sweep over the same evaluation data, the claimed improvement is partly a fitted input presented as a result. The paper does not describe a held-out selection split, so the Table 4 numbers are not an independent test of the hyperparameter choice.

  2. other [Section 4.3 Eq. (7); Section 4.2; Section 5.1]
    "When toxic content is detected in the batch, we apply an additional weight: wtox = (1.5 if toxic content detected / 1.0 otherwise). ... For generated toxic content, we employ a threshold-based (0.4) verification system that ensures only appropriately toxic examples are retained. ... For toxicity assessment, we employ the same toxicity classifier as in Section 3."

    The training procedure uses a toxicity detector to filter negative examples and to upweight the contrastive loss via w_tox; the evaluation then measures toxicity with 'the same toxicity classifier as in Section 3.' The reported toxicity reductions are therefore not independent of the training signal: the model is explicitly pushed toward the detections of the very classifier used to score the result. The toxicity half of the 'simultaneous improvements' claim is thus partially self-confirming, although the faithfulness half and the contrastive architecture still carry independent content.

full rationale

Two concrete reductions support partial circularity. First, Appendix H selects α=4 by maximizing toxicity/faithfulness on the Reddit TL;DR benchmark, and Section 5.3 reports the resulting Table 4 numbers as the paper's central demonstration; no held-out split is described, so the headline improvement is partly the optimum of the tuning curve rather than an out-of-sample finding. Second, the training loss contains a toxicity-detector-dependent weight (Eq. 7) and the negative-example filter uses a 0.4 toxicity threshold, while Section 5.1 evaluates toxicity with the same classifier used in Section 3; the toxicity reduction is therefore partly a measure of how well the model fits the detector that already shaped training. The faithfulness detector from Achintalwar et al. (2024) is external and is not a self-citation chain, but the paper never validates it against the TruthfulQA/MMLU-Pro benchmarks used in Section 3 to define the alignment tax; that is a construct-validity gap, not a circular reduction. The contrastive architecture, augmentation pipeline, and the independent comparisons to baselines carry some independent content, so this is not a fully forced derivation.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central result rests on a stack of hand-set hyperparameters (notably α=4, w_tox=1.5, thresholds), plus two unvalidated auxiliary detectors and the assumption that TL;DR faithfulness stands in for the knowledge benchmarks used to define the alignment tax.

free parameters (5)
  • alpha (α contrastive loss weight) = 4
    Selected as optimal via ablation on the Reddit TL;DR dataset (Appendix H), the same data used for the reported results.
  • temperature (τ) = 1
    Fixed hyperparameter with no sensitivity analysis reported; scales all contrastive similarity scores.
  • toxic loss weight (w_tox) = 1.5
    Hand-set multiplier in Eq. 7 applied when toxic content is detected in a batch; no derivation or sensitivity analysis.
  • toxicity retention threshold = 0.4
    Hand-set threshold for filtering generated toxic examples in Section 4.2; affects the composition of negative examples.
  • low-confidence threshold = 0.21
    Hand-set confidence cutoff for selecting low-confidence negative examples in Appendix E.
assumptions (6)
  • domain assumption The faithfulness detector from Achintalwar et al. 2024 accurately measures faithfulness.
    Section 5.1 uses this detector as ground truth for all faithfulness numbers; no precision, recall, or agreement metrics are reported.
  • domain assumption The BERT-based toxicity classifier (unitary/toxic-bert) is an unbiased measure of toxic content.
    Used in both Section 3 and Section 5; no local calibration or error analysis is provided.
  • domain assumption Faithfulness on Reddit TL;DR summarization is representative of general capability preservation.
    Section 3 defines the alignment tax using TruthfulQA and MMLU-Pro, but Section 5 evaluates only TL;DR, implicitly assuming these measure the same capabilities.
  • domain assumption Generating negative examples with GPT-Neo few-shot prompting produces valid toxic negatives.
    Section 4.2 assumes the few-shot template creates biased text that stays grammatically coherent and topically relevant.
  • standard math Standard contrastive loss (Eq. 8-9) behaves as a valid training objective.
    The temperature-scaled softmax and positive-pair log-sum are standard InfoNCE-style components; assumed correct without proof.
  • ad hoc to paper Dynamic loss scaling factor w_tox=1.5 should be applied when toxicity is detected.
    Eq. 7 introduces this multiplier without a derivation; it is a hand-chosen constant tuned on the same benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Paying Alignment Tax with Contrastive Learning." pith.science (2026). https://pith.science/paper/AKH3ZOAG

@misc{pith2026250519327,
  author       = {Pith},
  title        = {Pith review of: Paying Alignment Tax with Contrastive Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKH3ZOAG}},
  note         = {Machine review of arXiv:2505.19327}
}
read the original abstract

Current debiasing approaches often result a degradation in model capabilities such as factual accuracy and knowledge retention. Through systematic evaluation across multiple benchmarks, we demonstrate that existing debiasing methods face fundamental trade-offs, particularly in smaller models, leading to reduced truthfulness, knowledge loss, or unintelligible outputs. To address these limitations, we propose a contrastive learning framework that learns through carefully constructed positive and negative examples. Our approach introduces contrast computation and dynamic loss scaling to balance bias mitigation with faithfulness preservation. Experimental results across multiple model scales demonstrate that our method achieves substantial improvements in both toxicity reduction and faithfulness preservation. Most importantly, we show that our framework is the first to consistently improve both metrics simultaneously, avoiding the capability degradation characteristic of existing approaches. These results suggest that explicit modeling of both positive and negative examples through contrastive learning could be a promising direction for reducing the alignment tax in language model debiasing.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 6 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]

    Swapnaja Achintalwar, Adriana Alvarado Garcia, Ateret Anaby-Tavor, Ioana Baldini, Sara E Berger, Bishwaranjan Bhattacharjee, Djallel Bouneffouf, Subhajit Chaudhury, Pin-Yu Chen, Lamogha Chiazor, et al. 2024. Detectors for safe and reliable llms: Implementations, uses, and limitations. arXiv preprint arXiv:2403.06009

    work page Pith review arXiv 2024

  4. [4]

    Shuyang Cao and Lu Wang. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.532 CLIFF : Contrastive learning for improving faithfulness and factuality in abstractive summarization . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6633--6649, Online and Punta Cana, Dominican Republic. Association for Computat...

    work page doi:10.18653/v1/2021.emnlp-main.532 2021

  5. [5]

    Emily Dinan, Angela Fan, Adina Williams, Jack Urbanek, Douwe Kiela, and Jason Weston. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.656 Queens are powerful too: Mitigating gender bias in dialogue generation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8173--8188, Online. Association for Com...

    doi:10.18653/v1/2020.emnlp-main.656 2020

  6. [6]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.301 R eal T oxicity P rompts: Evaluating neural toxic degeneration in language models . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3356--3369, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.findings-emnlp.301 2020

  7. [7]

    Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio C \'e sar Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. 2023. Textbooks are all you need. arXiv preprint arXiv:2306.11644

    arXiv 2023

  8. [8]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. https://openreview.net/forum?id=d7KBjmI3GmQ Measuring massive multitask language understanding . In International Conference on Learning Representations

    2021

Show all 31 references
  1. [9]

    Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415

    2016 arXiv

  2. [10]

    Alkis Koudounas, Flavio Giobergia, Eliana Pastor, and Elena Baralis. 2024. https://doi.org/10.48550/arXiv.2406.14686 A contrastive learning approach to mitigate bias in speech models . CoRR, abs/2406.14686

    2024 doi arXiv

  3. [11]

    Yingji Li, Mengnan Du, Xin Wang, and Ying Wang. 2023. https://doi.org/10.18653/v1/2023.acl-long.797 Prompt tuning pushes farther, contrastive learning pulls closer: A two-stage approach to mitigate social biases . In Proceedings of the 61st Annual Meeting of the Association fo...

    2023 doi

  4. [12]

    Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, and Louis-Philippe Morency. 2020. https://doi.org/10.18653/v1/2020.acl-main.488 Towards debiasing sentence representations . In Proceedings of the 58th Annual Meeting of the Association for Compu...

    2020 doi

  5. [13]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013 ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  6. [14]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. https://doi.org/10.18653/v1/2022.acl-long.229 T ruthful QA : Measuring how models mimic human falsehoods . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa...

    2022 doi

  7. [15]

    Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, Hanze Dong, Renjie Pi, Han Zhao, Nan Jiang, Heng Ji, Yuan Yao, and Tong Zhang. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.35 Mitigating the alignm...

    2024 doi

  8. [16]

    Jonathan Mallinson, Rico Sennrich, and Mirella Lapata. 2017. https://aclanthology.org/E17-1083 Paraphrasing revisited with neural machine translation . In Proceedings of the 15th Conference of the E uropean Chapter of the Association for Computational Linguistics: Volume 1, Lo...

    2017

  9. [17]

    Nicholas Meade, Elinor Poole-Dayan, and Siva Reddy. 2022. https://doi.org/10.18653/v1/2022.acl-long.132 An empirical survey of the effectiveness of debiasing techniques for pre-trained language models . In Proceedings of the 60th Annual Meeting of the Association for Computati...

    2022 doi

  10. [18]

    Robert Morabito, Jad Kabbara, and Ali Emami. 2023. https://doi.org/10.18653/v1/2023.findings-acl.280 Debiasing should be good and bad: Measuring the consistency of debiasing techniques in language models . In Findings of the Association for Computational Linguistics: ACL 2023,...

    2023 doi

  11. [19]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 B leu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311--3...

    2002

  12. [20]

    Kyungmin Park, Sihyun Oh, Daehyun Kim, and Juae Kim. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.293 Contrastive learning as a polarizer: Mitigating gender bias by fair and biased sentences . In Findings of the Association for Computational Linguistics: NAACL 2024, p...

    2024 doi

  13. [21]

    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

  14. [22]

    Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. https://doi.org/10.18653/v1/2020.acl-main.647 Null it out: Guarding protected attributes by iterative nullspace projection . In Proceedings of the 58th Annual Meeting of the Association for Com...

    2020 doi

  15. [23]

    Timo Schick, Sahana Udupa, and Hinrich Sch \"u tze. 2021. https://doi.org/10.1162/tacl_a_00434 Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP . Transactions of the Association for Computational Linguistics, 9:1408--1424

    2021 doi

  16. [24]

    o rg Tiedemann, Mikko Aulamo, Daria Bakshandaeva, Michele Boggia, Stig-Arne Gr \

    J \"o rg Tiedemann, Mikko Aulamo, Daria Bakshandaeva, Michele Boggia, Stig-Arne Gr \"o nroos, Tommi Nieminen, Alessandro Raganato\, Yves Scherrer, Raul Vazquez, and Sami Virpioja. 2023. https://doi.org/10.1007/s10579-023-09704-w Democratizing neural machine translation with OP...

    2023 doi

  17. [25]

    J \"o rg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT — B uilding open translation services for the W orld. In Proceedings of the 22nd Annual Conferenec of the European Association for Machine Translation (EAMT), Lisbon, Portugal

    2020

  18. [26]

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

    2023 arXiv

  19. [27]

    Neeraj Varshney, Wenlin Yao, Hongming Zhang, Jianshu Chen, and Dong Yu. 2023. https://arxiv.org/abs/2307.03987 A stitch in time saves nine: Detecting and mitigating hallucinations of llms by validating low-confidence generation . Preprint, arXiv:2307.03987

    2023 arXiv

  20. [28]

    Michael V \"o lske, Martin Potthast, Shahbaz Syed, and Benno Stein. 2017. https://doi.org/10.18653/v1/W17-4508 TL ; DR : Mining R eddit to learn automatic summarization . In Proceedings of the Workshop on New Frontiers in Summarization, pages 59--63, Copenhagen, Denmark. Assoc...

    2017 doi

  21. [29]

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024. https://openreview.net/forum?id=y10DM6R2r3 MMLU -pro: A more...

    2024

  22. [30]

    Kellie Webster, Xuezhi Wang, Ian Tenney, Alex Beutel, Emily Pitler, Ellie Pavlick, Jilin Chen, Ed Chi, and Slav Petrov. 2020. Measuring and reducing gendered correlations in pre-trained models. arXiv preprint arXiv:2010.06032

    2020 arXiv

  23. [31]

    Mielke, Hanna Wallach, and Ryan Cotterell

    Ran Zmigrod, Sabrina J. Mielke, Hanna Wallach, and Ryan Cotterell. 2019. https://doi.org/10.18653/v1/P19-1161 Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology . In Proceedings of the 57th Annual Meeting of the Association for...

    2019 doi

Pith tools

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