Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Bias Vector: Mitigating Biases in Language Models with Task Arithmetic Approach

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that subtracting a Bias Vector—the weight difference from over-training on stereotyped text—from a pre-trained language model cuts measured bias by 0.177 points on average and preserves downstream task performance.

desk verdict Worth a serious referee, but the central claim needs a matched control before it carries the weight the abstract puts on it. read the letter →

arxiv 2412.11679 v1 pith:4VYVKCDQ submitted 2024-12-16 cs.CL cs.AI

Daiki Shirafuji , Makoto Takenaka , Shinya Taguchi This is my paper · ORCID
classification cs.CLcs.AI
keywords biasmitigationtaskarithmeticVectorweighteditingSEATGLUEStereoSetmaskedlanguagemodeling
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 tries to show that social bias in a language model can be treated like a task vector in weight space: continually train the same model on stereotype-saturated text, take the difference between the trained and original weights as a 'Bias Vector', and subtract that vector back off the original model. The claim is that this removes a measurable amount of bias without any manually curated debiasing dataset. Evaluated on the SEAT benchmark, BERT, ALBERT, and RoBERTa improved by an average of 0.177 effect-size points at scaling factor λ=1, while GLUE scores stayed essentially unchanged. The authors also argue that pushing λ high drives SEAT effect sizes toward zero not because bias is gone but because the model's representations collapse, so large-λ scores should not be read as successful debiasing.

What carries the argument

The central object is the Bias Vector, defined as the weight-space difference between a model continually trained on biased text and the original model, $V_{\mathrm{bias}} = \theta_{\mathrm{bias}} - \theta_{\mathrm{org}}$. It is applied through the simple rule $\theta_{\mathrm{debias}} = \theta_{\mathrm{org}} - \lambda V_{\mathrm{bias}}$, so the mechanism is vector addition and subtraction of parameter tensors under the assumption that models share an architecture; the scalar $\lambda$ controls how much bias is subtracted. The paper also excludes Layer Normalization parameters from the vector, arguing those layers only normalize activations and carry no bias information. The mechanism carries the argument because it turns debiasing into a one-line arithmetic operation: no debiasing corpus, no additional training at inference, and the same vector can be scaled up or down to trade bias reduction against preservation of language ability.

What would settle it

Repeat the continual training with several random seeds and check whether the Bias Vectors agree in direction (high pairwise cosine similarity) and whether the per-seed SEAT improvements bracket zero; if the vectors are seed-dependent or the SEAT spread exceeds the 0.177-point average gain, the claimed bias direction is not a stable property of the weights.

Watch

Extended reading notes

Core claim

The central discovery is that the arithmetic of task vectors transfers to bias: with $\theta_{\mathrm{org}}$ the pre-trained weights and $\theta_{\mathrm{bias}}$ the weights after masked-language-model continual training on stereotyped StereoSet sentences, the vector $V_{\mathrm{bias}} = \theta_{\mathrm{bias}} - \theta_{\mathrm{org}}$ encodes bias, and $\theta_{\mathrm{debias}} = \theta_{\mathrm{org}} - \lambda V_{\mathrm{bias}}$ yields a debiased model. At $\lambda=1$ this lowers average absolute SEAT effect sizes from 0.672/0.675/0.733 to 0.447/0.534/0.570 for BERT/ALBERT/RoBERTa, an average improvement of 0.177 points, while average GLUE scores move from about 0.776/0.779/0.794 to 0.779/0.785/0.792. The same experiment shows that subtracting too much (large λ) makes effect sizes collapse toward zero and GLUE scores fall sharply, which the paper reads as evidence that over-debiasing destroys representations rather than truly removing bias. Bias-type-specific vectors (race, profession, gender, religion, and all combined) all reduce SEAT scores at λ=1, and the Layer Normalization layers are excluded from the subtraction because they are treated as normalization-only.

Load-bearing premise

The load-bearing premise is that a model's bias is stored as a direction in parameter space that can be cleanly separated from general language ability, so that subtracting a fixed multiple of the Bias Vector removes bias without erasing other capabilities.

Editorial extensions

If this is right

  • With λ=1, all four bias-type vectors (race, profession, gender, religion) and the combined vector lower average absolute SEAT effect sizes for BERT, ALBERT, and RoBERTa, so a single subtraction step is a usable debiasing recipe across BERT-family models.
  • GLUE averages after debiasing at λ=1 are 0.779 (BERT), 0.785 (ALBERT), and 0.792 (RoBERTa), within about 0.2–0.6 points of the pre-trained baselines, so the paper's debiasing does not trade away downstream language understanding.
  • Because large λ drives both SEAT and GLUE scores down simultaneously, near-zero effect sizes on SEAT are not, by themselves, evidence of successful debiasing; task performance must be checked in the same experiment.
  • Debiasing with the profession vector alone also improved SEAT scores on other categories, which the paper attributes to overlapping bias directions; this implies that composing multiple bias vectors requires care to avoid double-subtracting shared components.

Reading between the lines

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

  • As the paper's own future-work section anticipates, the recipe should transfer to decoder-only LLMs; the open question is whether a Bias Vector built on 8,498 sentences remains a clean direction when subtracted from a much larger model.
  • Because the vector is built from StereoSet, whose stereotypes were annotated from a U.S. perspective, the method can only remove the bias directions present in that corpus; in another cultural context it could leave local stereotypes intact or distort them, so corpus coverage is a hidden boundary condition.
  • The bias-direction reversal seen at intermediate λ before collapse suggests that subtracting the vector moves the model along a continuum between the biased and original checkpoints; per-bias or per-layer λ tuning could exploit this to get finer control than the single global scalar used here.
  • A sharper evaluation than GLUE would be to compare layer-wise contributions of the Bias Vector; if only a few layers carry the bias direction, subtracting from those layers alone could cut SEAT with less collateral damage than the uniform subtraction the paper applies.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 a 'Bias Vector' method for debiasing pre-trained language models. The method first continually trains a model on biased text (StereoSet intrasentence sentences with stereotype fill-ins), computes the parameter difference between this biased model and the original pre-trained model (the bias vector), and then subtracts a scaled version of this vector from the original weights to obtain a debiased model. Experiments on BERT, ALBERT, and RoBERTa report an average improvement of 0.177 in absolute SEAT effect sizes at λ=1, while GLUE performance is reported as approximately unchanged. The paper also analyzes the effect of the scaling factor λ and shows that large λ values drive SEAT effect sizes toward zero, but attributes this to a collapse of representations rather than genuine debiasing.

Significance. If the central claim holds, the method would offer a simple, data-efficient debiasing approach that avoids constructing dedicated debiasing datasets, consistent with recent task-arithmetic ideas. The paper has notable strengths: it evaluates three model architectures, excludes StereoSet from the evaluation suite to reduce leakage concerns, constructs bias vectors over ten seeds, and honestly analyzes in Section 5.6 that large λ values can make SEAT scores decrease because of representation collapse rather than true bias removal. However, the main claim is not yet fully established. The paper lacks a matched control vector trained on non-stereotyped text, the per-bias-type vectors are inconsistent across models, the headline results rely on average absolute effect sizes, and no variance or significance information is provided for the GLUE comparison. These issues leave open the possibility that the observed improvements arise from generic fine-tuning effects or incipient representation degradation rather than from the removal of bias-specific information.

major comments (5)
  1. [Section 3.2, Section 5.1] The central claim that V_bias encodes bias-specific information is not verified by the experiments as presented. The only comparison is against the pre-trained baseline; there is no control vector constructed from an equal-sized corpus of non-stereotyped (or anti-stereotyped) text trained with identical hyperparameters and seeds. Without such a control, the observed SEAT reduction at λ=1 could be produced by any continual-MLM fine-tuning direction that slightly degrades embedding separability, especially given the paper's own finding (Section 5.6) that SEAT approaches zero when representations collapse. Please add a matched control experiment—for example, a vector trained on neutral text from the same domain—and show that the bias vector reduces SEAT significantly more than the control.
  2. [Section 5.1, Table 3] The headline 0.177 improvement is reported for the aggregated 'all' bias vector only, and the per-bias-type vectors are inconsistent. For ALBERT, BV(gender,1) and BV(religion,1) yield average absolute effect sizes of 0.736 and 0.735, which are above the pre-trained baseline of 0.675; for BERT, the individual vectors give much smaller improvements than BV(all,1). The paper should report signed effect sizes and standard deviations across the ten seeds for each vector, and explain why the aggregate vector behaves differently from its components; without this, the claim of consistent bias mitigation is not supported.
  3. [Section 5.1, Section 5.5, Table 5] The main tables report averages of absolute effect sizes. This can conflate genuine debiasing with bias reversal: if a debiasing step changes a positive effect size to a negative one of similar magnitude, the absolute value changes while the bias flips direction. Section 5.5 explicitly reports such reversal for intermediate λ, and Table 5 reports signed effect sizes. Please report signed effect sizes for all conditions and test whether the signed values move toward zero, not merely whether absolute values decrease.
  4. [Section 4.4, Table 4] The GLUE experiments are run with a single fixed seed, so the claim that performance is 'not degraded' at λ=1 is not statistically supported. The observed changes (BERT +0.3%, ALBERT +0.6%, RoBERTa −0.2%) are within the range of typical seed-to-seed variation for GLUE fine-tuning. Please report multiple seeds with means and standard deviations, or at least provide a paired bootstrap or significance test over several seeds for the λ=1 condition.
  5. [Appendix D.2, Table 5] The comparison with existing debiasing methods in Table 5 is indirect: the baseline and baseline-method rows are cited from Meade et al. (2022), while the Bias Vector rows are from the present paper's own evaluation pipeline. Differences in evaluation code, sentence templates, or data splits could account for part of the reported differences. The claim that the proposed method is 'sometimes competitive' with embedding-space methods would be more convincing if the existing methods were re-run under the same codebase (e.g., the bias-bench code the authors cite) with matched seeds and evaluation settings.
minor comments (5)
  1. [Throughout] There are several typos and grammatical issues, including 'ahve' (Section 2.1), 'neasurements' (Section 2.1), 'emperical' (Appendix D.2), 'Arrguments' (Appendix C), and the title line 'Bias V ector' in the first line of the manuscript.
  2. [Table 4, caption] The table caption says the results are for λ=1, but the table also includes rows for λ=10 and λ=100; the caption should be revised to describe the full table content.
  3. [Section 3.2] The exclusion of LayerNorm layers from the vector subtraction is stated without motivation or ablation. A sentence explaining why this choice was made, and ideally an ablation showing its effect, would be helpful.
  4. [Section 5.3] The paper acknowledges that SEAT does not directly evaluate profession bias, but the main table reports an average over gender, race, and religion subtests. It would be useful to also report the results separately by SEAT subtest to clarify which bias types drive the observed improvements.
  5. [Appendix D.2.1] The reference list for CDA includes multiple works that use different formulations (e.g., counterfactual data augmentation, dialogue debiasing); a sentence clarifying which variant was used in the cited results would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the debiasing result is an empirical finding evaluated on data disjoint from the training set, not a consequence of how the Bias Vector is defined.

full rationale

The Bias Vector is constructed as V_bias = θ_bias − θ_org, where θ_bias comes from continual MLM training on StereoSet stereotype sentences, and the debiased model is θ_debias = θ_org − λV_bias. The headline improvement is measured on SEAT, which is explicitly disjoint from the training data: 'It should be noted that the StereoSet dataset, used for the continual training (Section 3.1), is excluded from our evaluation experiments to prevent data leakage.' GLUE provides an independent capability check. λ is not a fitted parameter renamed as a prediction; it is swept from 0.01 to 10,000 and reported, with the λ=1 result being one point in that sweep. No load-bearing self-citation appears in the derivation; task arithmetic (Ilharco et al., 2023) is cited as inspiration, but the bias-vector construction and its evaluation are the paper's own empirical contribution. The absence of a non-stereotyped control vector is a validity concern (the improvement might be a generic fine-tuning artifact), not circularity, because no equation or fitted value in the paper makes the SEAT drop true by definition.

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

The paper introduces no new theoretical entities. Its contribution is an empirical recipe whose main cost is the assumption that bias directions in weight space are linearly removable. The free parameters are ordinary training hyperparameters plus the scaling factor; the central result depends on λ=1 and on the choice to overfit 30 epochs, neither of which is accompanied by sensitivity analysis or code.

free parameters (3)
  • scaling factor λ = 1, 10, 100 (main); 0.01 to 10,000 (analysis)
    Controls magnitude of bias vector subtraction in Equation (2); the paper's main claim uses λ=1, and high values cause representation collapse.
  • continual training epochs = 30
    Chosen to "effectively overfit the LMs toward biases" (Section 4.2.2); the amount of overfitting determines the bias vector.
  • continual training learning rate = 1e-4
    Used with AdamW; affects how the bias vector is formed, but the paper does not test sensitivity.
assumptions (3)
  • domain assumption Bias is a linearly separable direction in parameter space: θ_debias = θ_org - λ(θ_bias - θ_org) removes bias without removing other capabilities.
    Core hypothesis of the method, introduced in Section 2 and formalized in Equations (1)-(2); no proof or independent validation is given.
  • domain assumption Layer Normalization layers do not learn bias information and can be excluded from subtraction.
    Stated in Section 3.2 without empirical support; if false, the bias vector omits relevant parameters.
  • domain assumption StereoSet intrasentence stereotype sentences, with blanks filled by stereotype options, provide a valid "bias-only" training signal for the biases measured by SEAT.
    The training data construction in Section 4.2.1 assumes the stereotypes in StereoSet (US annotator views) align with the bias directions SEAT measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias Vector: Mitigating Biases in Language Models with Task Arithmetic Approach." pith.science (2026). https://pith.science/paper/4VYVKCDQ

@misc{pith2026241211679,
  author       = {Pith},
  title        = {Pith review of: Bias Vector: Mitigating Biases in Language Models with Task Arithmetic Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VYVKCDQ}},
  note         = {Machine review of arXiv:2412.11679}
}
read the original abstract

The use of language models (LMs) has increased considerably in recent years, and the biases and stereotypes in training data that are reflected in the LM outputs are causing social problems. In this paper, inspired by the task arithmetic, we propose the ``Bias Vector'' method for the mitigation of these LM biases. The Bias Vector method does not require manually created debiasing data. The three main steps of our approach involve: (1) continual training the pre-trained LMs on biased data using masked language modeling; (2) constructing the Bias Vector as the difference between the weights of the biased LMs and those of pre-trained LMs; and (3) subtracting the Bias Vector from the weights of the pre-trained LMs for debiasing. We evaluated the Bias Vector method on the SEAT across three LMs and confirmed an average improvement of 0.177 points. We demonstrated that the Bias Vector method does not degrade the LM performance on downstream tasks in the GLUE benchmark. In addition, we examined the impact of scaling factors, which control the magnitudes of Bias Vectors, with effect sizes on the SEAT and conducted a comprehensive evaluation of our debiased LMs across both the SEAT and GLUE benchmarks.

Figures

Figures reproduced from arXiv: 2412.11679 by the authors.

Figure 1
Figure 1. Overview of the Bias Vector method: (1) Training pre-trained LMs on biased data to create the biased [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Variation of effect sizes on the SEAT with the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Effect sizes on the gender-biased SEAT dataset (SEAT-8) with varying λ. The effect sizes are computed as the average of scores across ten different seed values. The closer the effect size is to zero, the smaller the bias. served results deviated from this trend. This section investigates whether the discrep￾ancy between our hypothesis and the SEAT results can be attributed to the Bias Vectors failing to ade￾quately … view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Effect sizes on gender bias tests in SEAT when [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Means and standard deviations of effect sizes on SEAT-8 with debiased BERT. 10 2 10 1 10 0 10 1 10 2 10 3 10 4 Effect Sizes on SEAT 1.5 1.0 0.5 0.0 0.5 1.0 1.5 lamda [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 8
Figure 8. Figure 8: Means and standard deviations of effect sizes on SEAT-8 with debiased ALBERT. 10 2 10 1 10 0 10 1 10 2 10 3 10 4 Effect Sizes on SEAT 1.0 0.5 0.0 0.5 1.0 lamda [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 10
Figure 10. Figure 10: Means and standard deviations of effect sizes on SEAT-8 with debiased RoBERTa. 10 2 10 1 10 0 10 1 10 2 10 3 10 4 Effect Sizes on SEAT 1.0 0.5 0.0 0.5 1.0 lamda [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Guard Vector: Beyond English LLM Guardrails with Task-Vector Composition and Streaming-Aware Prefix SFT

    cs.CL 2025-09 conditional novelty 6.0 of 10

    A task vector from an English guard model transfers safety classification to Korean, Chinese, and Japanese models, and a prefix-SFT variant maintains accuracy under streaming with a single-token classifier.

Reference graph

Works this paper leans on

41 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Samuel Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. 2023. https://openreview.net/forum?id=CQsmMYmlP5T Git R e- B asin: M erging M odels modulo P ermutation S ymmetries . In The Eleventh International Conference on Learning Representations

    work page 2023

  2. [2]

    Soumya Barikeri, Anne Lauscher, Ivan Vuli \'c , and Goran Glava s . 2021. https://doi.org/10.18653/v1/2021.acl-long.151 R eddit B ias: A R eal- W orld R esource for B ias E valuation and D ebiasing of C onversational L anguage M odels . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Jo...

    doi:10.18653/v1/2021.acl-long.151 2021

  3. [3]

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. https://proceedings.neurips.cc/paper_files/paper/2016/file/a486cd07e4ac3d270571622f4f316ec5-Paper.pdf Man is to C omputer P rogrammer as W oman is to H omemaker? D ebiasing W ord E mbeddings . In Advances in Neural Information Processing Systems, volume 29. Curran As...

    work page 2016

  4. [4]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : P re-training of D eep B idirectional T ransformers for L anguage U nderstanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (...

    doi:10.18653/v1/n19-1423 2019

  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 P owerful too: M itigating G ender B ias in D ialogue G eneration . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8173--8188, Online. Association f...

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

  6. [6]

    Shangbin Feng, Chan Young Park, Yuhan Liu, and Yulia Tsvetkov. 2023. https://doi.org/10.18653/v1/2023.acl-long.656 From P retraining D ata to L anguage M odels to D ownstream T asks: T racking the T rails of P olitical B iases L eading to U nfair NLP M odels . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volu...

    doi:10.18653/v1/2023.acl-long.656 2023

  7. [7]

    Hila Gonen and Yoav Goldberg. 2019. https://doi.org/10.18653/v1/N19-1061 Lipstick on a P ig: D ebiasing M ethods C over up S ystematic G ender B iases in W ord E mbeddings B ut do not R emove T hem . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 ...

    doi:10.18653/v1/n19-1061 2019

  8. [8]

    Shih-Cheng Huang, Pin-Zu Li, Yu-chi Hsu, Kuang-Ming Chen, Yu Tung Lin, Shih-Kai Hsiao, Richard Tsai, and Hung-yi Lee. 2024. https://aclanthology.org/2024.acl-long.590 Chat V ector: A S imple A pproach to E quip LLM s with I nstruction F ollowing and M odel A lignment in N ew L anguages . In Proceedings of the 62nd Annual Meeting of the Association for Com...

    work page 2024

Show all 41 references
  1. [9]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. https://openreview.net/forum?id=6t0Kwf8-jrj Editing M odels with T ask A rithmetic . In The Eleventh International Conference on Learning Representations

    2023

  2. [10]

    Bryson, and Arvind Narayanan

    Aylin Caliskan Islam, Joanna J. Bryson, and Arvind Narayanan. 2016. https://arxiv.org/abs/1608.07187 Semantics derived automatically from language corpora necessarily contain human biases . Science, 356(2)

    2016 arXiv

  3. [11]

    Sophie Jentzsch and Cigdem Turan. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.20 Gender B ias in BERT - M easuring and A nalysing B iases through S entiment R ating in a R ealistic D ownstream C lassification T ask . In Proceedings of the 4th Workshop on Gender Bias in Nat...

    2022 doi

  4. [12]

    Masahiro Kaneko and Danushka Bollegala. 2019. https://doi.org/10.18653/v1/P19-1160 Gender-preserving D ebiasing for P re-trained W ord E mbeddings . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1641--1650, Florence, Italy. A...

    2019 doi

  5. [13]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. https://arxiv.org/abs/1412.6980 Adam: A M ethod for S tochastic O ptimization . Preprint, arXiv:1412.6980

    2017 arXiv

  6. [14]

    Sachin Kumar, Vidhisha Balachandran, Lucille Njoo, Antonios Anastasopoulos, and Yulia Tsvetkov. 2023. https://doi.org/10.18653/v1/2023.eacl-main.241 Language G eneration M odels C an C ause H arm: S o W hat C an W e D o A bout I t? A n A ctionable S urvey . In Proceedings of t...

    2023 doi

  7. [15]

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. https://openreview.net/forum?id=H1eA7AEtvS ALBERT : A L ite BERT for S elf-supervised L earning of L anguage R epresentations . In International Conference on Learning Representations

    2020

  8. [16]

    Smith, and Luke Zettlemoyer

    Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A. Smith, and Luke Zettlemoyer. 2022. https://openreview.net/forum?id=SQgVgE2Sq4 Branch- T rain- M erge: E mbarrassingly P arallel T raining of E xpert L anguage M odels . In First Workshop on Interpol...

    2022

  9. [17]

    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 D ebiasing S entence R epresentations . In Proceedings of the 58th Annual Meeting of the Association for Co...

    2020 doi

  10. [18]

    Haochen Liu, Jamell Dacon, Wenqi Fan, Hui Liu, Zitao Liu, and Jiliang Tang. 2020. https://doi.org/10.18653/v1/2020.coling-main.390 Does G ender M atter? T owards F airness in D ialogue S ystems . In Proceedings of the 28th International Conference on Computational Linguistics,...

    2020 doi

  11. [19]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Ro BERT a: A R obustly O ptimized BERT P retraining A pproach . Preprint, arXiv:1907.11692

    2019 arXiv

  12. [20]

    Ilya Loshchilov and Frank Hutter. 2017. https://arxiv.org/abs/1711.05101 Fixing W eight D ecay R egularization in A dam . Preprint, arXiv:1711.05101

    2017 arXiv

  13. [21]

    Michael Matena and Colin Raffel. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/70c26937fbf3d4600b69a129031b66ec-Paper-Conference.pdf Merging M odels with F isher- W eighted A veraging . In Advances in Neural Information Processing Systems, volume 35, pages 1...

    2022

  14. [22]

    Bowman, and Rachel Rudinger

    Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019. https://doi.org/10.18653/v1/N19-1063 On M easuring S ocial B iases in S entence E ncoders . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computat...

    2019 doi

  15. [23]

    Nicholas Meade, Elinor Poole-Dayan, and Siva Reddy. 2022. https://doi.org/10.18653/v1/2022.acl-long.132 An E mpirical S urvey of the E ffectiveness of D ebiasing T echniques for P re-trained L anguage M odels . In Proceedings of the 60th Annual Meeting of the Association for C...

    2022 doi

  16. [24]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. https://arxiv.org/abs/1301.3781 Efficient E stimation of W ord R epresentations in V ector S pace . Preprint, arXiv:1301.3781

    2013 arXiv

  17. [25]

    Jiaqi Mu and Pramod Viswanath. 2018. https://openreview.net/forum?id=HkuGJ3kCb All-but-the- T op: S imple and E ffective P ostprocessing for W ord R epresentations . In International Conference on Learning Representations

    2018

  18. [26]

    Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. https://doi.org/10.18653/v1/2021.acl-long.416 S tereo S et: M easuring stereotypical bias in pretrained language models . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Int...

    2021 doi

  19. [27]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.154 C row S -pairs: A C hallenge D ataset for M easuring S ocial B iases in M asked L anguage M odels . In Proceedings of the 2020 Conference on Empirical Metho...

    2020 doi

  20. [28]

    Roberto Navigli, Simone Conia, and Bj\" o rn Ross. 2023. https://doi.org/10.1145/3597307 Biases in L arge L anguage M odels: O rigins, I nventory, and D iscussion . ACM Journal of Data and Information Quality, Volume 15, Issue 2

    2023 doi

  21. [29]

    OpenAI. 2022. https://openai.com/index/chatgpt/ https://openai.com/index/chatgpt/

    2022

  22. [30]

    OpenAI. 2024. https://arxiv.org/abs/2303.08774 GPT -4 T echnical R eport . Preprint, arXiv:2303.08774

    2024 arXiv

  23. [31]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf Language M odels are U nsupervised M ultitask L earners

    2019

  24. [32]

    Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. https://doi.org/10.18653/v1/2020.acl-main.647 Null I t O ut: G uarding P rotected A ttributes by I terative N ullspace P rojection . In Proceedings of the 58th Annual Meeting of the Association...

    2020 doi

  25. [33]

    Timo Schick, Sahana Udupa, and Hinrich Sch \"u tze. 2021. https://doi.org/10.1162/tacl_a_00434 Self- D iagnosis and S elf- D ebiasing: A P roposal for R educing C orpus- B ased B ias in NLP . Transactions of the Association for Computational Linguistics, 9:1408--1424

    2021 doi

  26. [34]

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/W18-5446 GLUE : A M ulti- T ask B enchmark and A nalysis P latform for N atural L anguage U nderstanding . In Proceedings of the 2018 EMNLP Workshop B lackbo...

    2018 doi

  27. [35]

    Tianlu Wang, Xi Victoria Lin, Nazneen Fatema Rajani, Bryan McCann, Vicente Ordonez, and Caiming Xiong. 2020. https://doi.org/10.18653/v1/2020.acl-main.484 Double- H ard D ebias: T ailoring W ord E mbeddings for G ender B ias M itigation . In Proceedings of the 58th Annual Meet...

    2020 doi

  28. [36]

    Kellie Webster, Xuezhi Wang, Ian Tenney, Alex Beutel, Emily Pitler, Ellie Pavlick, Jilin Chen, and Slav Petrov. 2020. https://arxiv.org/abs/2010.06032 Measuring and R educing G endered C orrelations in P re-trained M odels . Preprint, arXiv:2010.06032

    2020 arXiv

  29. [37]

    Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt

    Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. 2022. https://arxiv.org/abs/2203.05482 Model soups: averaging weights of multiple f...

    2022 arXiv

  30. [38]

    Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and Kai-Wei Chang. 2018. https://doi.org/10.18653/v1/D18-1521 Learning G ender- N eutral W ord E mbeddings . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4847--4853, Brussels, Belgi...

    2018 doi

  31. [39]

    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 D ata A ugmentation for M itigating G ender S tereotypes in L anguages with R ich M orphology . In Proceedings of the 57th Annual Meeting of the Associa...

    2019 doi

  32. [40]

    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...

  33. [41]

    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 11, 2026 · model on record in the stance chip above.