Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair Language Modeling and Translation

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read 2DAMA is a post-hoc tuning method that removes stereotypical gender bias from language models and machine translation while preserving factual gender cues and most general performance.

desk verdict The LM side of 2DAMA is a credible incremental contribution, but the translation results are misread: the negative ΔG values are bias reversals, not factual-gender preservation. read the letter →

arxiv 2501.10150 v2 pith:QDKIUY4N submitted 2025-01-17 cs.CL cs.AI

classification cs.CLcs.AI
keywords genderbiasdebiasinglargelanguagemodelsmachinetranslationmodeleditingconcepterasurefactualfairness
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 introduces 2DAMA, a post-hoc adaptation method for pre-trained language models that erases stereotypical gender associations while deliberately keeping factual gender information. The authors argue that debiasing should not remove all gender signals, because pronouns and grammatical gender in morphologically rich languages are legitimate factual cues. They combine the DAMA model-editing approach with the closed-form LEACE concept erasure, then add a covariance-based threshold to nullify only directions that carry stereotype information. Experiments on Llama 2, Llama 3, and the ALMA-R translation model show reduced gender bias in English language modeling and in translation to German, Czech, and Russian, with modest impact on general quality. The paper claims this is among the first methods to debias machine translation while preserving factual gender.

What carries the argument

The central machinery is a linear projection applied to the weights of transformer feed-forward layers. The paper assumes the layer acts as a least-squares map $V = SU - \epsilon$, so the optimal debiasing matrix is $P^* = (I - W^\dagger P_{W\Sigma} W) S$, where $W$ whitens the key-value covariance and $P_{W\Sigma}$ projects onto the covariance direction of the bias signal. Dual Debiasing modifies the projection to keep, rather than nullify, directions whose covariance with factual gender exceeds a bias-to-feature threshold $t$. The authors apply this projection to roughly one-third of the model's mid-upper feed-forward layers.

What would settle it

Evaluate a debiased model on a held-out set of professions and on naturalistic sentences not used to build the projection; if the stereotypical coefficient stays close to the original level while the factual coefficient drops, the dual-debiasing claim would be contradicted.

Watch

Extended reading notes

Core claim

The paper claims that a modified LEACE projection applied to transformer feed-forward layer weights, guided by a bias-to-feature threshold, can suppress stereotypical gender associations while retaining factual gender information. Theoretically, it extends LEACE to linear transformations and proves that under conditional independence of bias and feature signals given the latent representation, the projection leaves factual gender covariance intact; in practice, when signals are correlated, it uses a threshold $t$ to keep dimensions whose covariance with factual gender is $t$ times larger than with bias. Empirically, with $t=0.05$, the method erases about 99% of stereotypical covariance while retaining over 30% of factual-gender covariance in each edited layer. The authors also find that factual gender is concentrated in a few high-covariance dimensions, whereas stereotype signal is spread across more dimensions with lower covariance.

Load-bearing premise

The load-bearing premise is that a transformer feed-forward layer behaves like a linear least-squares map from inputs to outputs, so a linear projection computed from prompt-derived vectors can erase bias in that layer and transfer to unseen text.

Editorial extensions

If this is right

  • 2DAMA reduces stereotypical gender bias in English language modeling on Llama 2 and Llama 3 while keeping perplexity and ARC performance close to the original.
  • It reduces stereotypical bias in English-to-German, English-to-Czech, and English-to-Russian translation with only small drops in COMET and chrF quality.
  • With a low bias-to-feature threshold, 2DAMA erases most of the stereotype covariance while keeping a substantial share of factual-gender covariance in edited layers.
  • Cross-lingual tests show that prompts in one language can reduce stereotypical bias in another (e.g., German prompts help Czech), while representational bias is better controlled with in-language prompts.
  • 2DAMA does not require a preset dimensionality for bias erasure; the LEACE-based projection learns the extent of bias signals implicitly.

Reading between the lines

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

  • Editorial inference: If stereotype and factual gender occupy largely orthogonal linear subspaces in feed-forward layers, the same covariance-threshold trick could be applied to other protected attributes where a legitimate factual component exists, such as age or nationality.
  • Editorial inference: The cross-lingual results suggest that stereotype direction learned in one language transfers to related target languages; a testable next step is to measure whether prompt-language similarity predicts transfer strength.
  • Editorial inference: The paper evaluates on profession prompts and coreference benchmarks; a natural stress test would be long-form generation or naturalistic text, where factual gender preservation might interact with narrative context.
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 / 6 minor

Summary. The paper proposes 2DAMA (Dual Debiasing Algorithm through Model Adaptation), a post-hoc model-editing method that aims to reduce stereotypical gender bias in large language models and in machine translation while preserving factual gender information. The method combines DAMA (Limisiewicz et al., 2024), LEACE (Belrose et al., 2023), and a new covariance-based decomposition that, given a hand-set threshold t, nullifies dimensions associated with stereotypical bias while retaining dimensions correlated with factual gender. Experiments on Llama 2 (7B, 13B), Llama 3 (8B), and ALMA-R (13B) report reductions on bias metrics (linear stereotype coefficients, WinoBias, WinoMT, BUG) with modest perplexity and ARC degradation. The paper also studies hyperparameter choices and cross-lingual transfer of debiasing signals.

Significance. If the central claim were fully supported, the paper would make a useful contribution: it addresses an understudied trade-off between debiasing and preserving legitimate gender cues, and it is among the first attempts to debias LLM-based machine translation post hoc. The paper presents a theorem relating concept erasure and feature preservation, a layer-wise analysis of where stereotypical and factual gender signals reside, and a cross-lingual generalization study. These analyses are valuable. However, the significance is currently tempered by two issues: the practical 2DAMA algorithm explicitly abandons the theorem's conditions, and the translation experiments use a signed ΔG metric whose negative values are interpreted as preservation rather than as bias reversal. The English language-modeling results are more solid, but they alone do not support the paper's distinctive dual-debiasing claim for machine translation.

major comments (4)
  1. [Section 5.1, Table 4] The interpretation of ΔG for 2DAMA (t = 1.00) is not supported by the numbers. The paper marks ΔG with a downward arrow (lower is better) but then presents a negative value for German (3.7 → −10.1) as evidence that higher t 'tends to preserve more factual signal' and yields 'equitable representation of factual gender.' Under the standard signed definition, a negative ΔG is a bias reversal, not a reduction; under an absolute-value reading, the bias magnitude increases from 3.7 to 10.1. For Czech, ΔG worsens from 2.1 to 7.2 even at t = 1.00. Moreover, ΔG measures the difference in translation accuracy between masculine and feminine entities; it does not measure factual-gender preservation. The paper needs either a corrected interpretation of these results or a direct metric of factual-gender retention (e.g., accuracy on held-out factually gendered sentences) to support the MT half of the dual-debiasing claim.
  2. [Section 2.4, Theorem 3] The practical Dual Debiasing algorithm is not covered by the theorem that precedes it. The authors state that the conditional-independence assumption (Zb ⊥ Zf | X) is 'strong and unlikely to hold' and that they 'abandon theoretical guarantees' in the actual algorithm, replacing the exact projection with a thresholded version that uses an empirically chosen t. Consequently, the introduction's promise of 'strong theoretical backing' for 2DAMA is not fulfilled for the method as implemented. The paper should either provide an approximation bound or a robustness analysis showing that the thresholded projection still approximately preserves Zf under near-conditional-independence, or explicitly reposition Theorem 3 as motivational rather than as a guarantee for the evaluated method. As it stands, the central dual-debiasing claim for the practical algorithm rests entirely on the empirical evaluation, which has the ΔG problems noted above.
  3. [Section 3.3, Table 1] The sign convention for the factual-gender coefficient af is inconsistent and obscures the factual-preservation claim. Table 1 places a downward arrow on af, implying that lower values are better, yet Section 4.1 interprets a higher af (e.g., 2DAMA's 0.042 vs. DAMA+LEACE's 0.028 for Llama 2 7B) as preserving factual gender information. The same ambiguity applies to the arrow on the intercept b. The authors should clarify the direction of all metrics in the table and, for af, report a measure that explicitly reflects preservation of factual gender rather than a generic regression coefficient that could be lower for other reasons.
  4. [Tables 1, 2, and 4] No error bars, confidence intervals, or significance tests are reported for any of the comparisons. Several key differences are small (e.g., Table 2: perplexity 21.51 vs. 21.28; COMET/chrf changes under 1 point in Table 4), so the repeated conclusion that general performance is 'not significantly affected' is not statistically supported. At minimum, the authors should report variance across multiple evaluation runs or seeds, or use a resampling-based test for the headline bias reductions, so that readers can assess whether the observed improvements are distinguishable from noise.
minor comments (6)
  1. [Abstract and Section 7] The abstract and the conclusion contain the misspelling 'Dabiasing'; change to 'Debiasing' throughout. Additional typos include 'Moonrose-Penrose' (Appendix A.1), 'Erarsuer' (Section 2.2 heading), 'Seterotypical' (Section 4.2 heading), 'psuedoinverse' (Appendix A.1), and 'Germna' (Figure 5 caption).
  2. [Section 2.4, Theorem 3 statement] The condition (B) in Theorem 3 is malformed: it is written as 'ΣX,ZbΣT X,Zf' without an equality or a stated property. The proof relies on the two covariance spaces being orthogonal, so the statement should be corrected, e.g., 'ΣX,Zb ΣT_{X,Zf} = 0'.
  3. [Section 2.4, Theorem 3 statement] The statement of Theorem 3 also contains a grammatical repetition: 'Zb and Zf Zb ⊥ Zf |X'. Please rephrase to 'Zb is conditionally independent of Zf given X.'
  4. [Section 6.2] The claim of being 'the first to propose a method for debiasing LLM in machine translation tasks' should be qualified in light of prior work on debiasing neural MT (e.g., Saunders and Byrne, 2020; Zmigrod et al., 2019), even if those works are not LLM-based. The phrasing should clarify the novelty relative to those lines of work.
  5. [Section 3.3] The description of the linear model (Eq. 2) says 'The slope coefficient tells how shifts in annotated scores across professions impact the difference in prediction probabilities.' This is correct, but the table header should specify whether the reported coefficients are on the original scale or multiplied (as in Figure 4, where they are 100·as and 100·af).
  6. [Section 5.1 and Table 4] The table reports both WinoMT and BUG columns, but the text mentions only ΔS and ΔG; it would be helpful to point out that the BUG results are also part of the evaluation and to note that BUG and WinoMT use different sentence distributions.

Circularity Check

2 steps flagged · score 5.0 of 10

Translation-side 'factual gender preservation' is partly forced by post-hoc threshold selection and a non-factual metric; LM-side claims retain independent empirical content.

  1. fitted input called prediction [Section 5.1 (Table 4), Section 2.4, and Appendix C.2 (Figure 5b)]
    "the threshold t (later referred to as bias-to-feature threshold) is empirically chosen. ... To remedy that, we use 2DAMA with higher values of feature-to-bias threshold (t = 1.00 instead of t = 0.05), which tends to preserve more factual signal. Factual gender understanding is especially essential for equitable representation of factual gender in morphologically rich languages, as evidenced by ∆G scores for t = 1 .00 setting."

    The threshold t is an empirically chosen hyperparameter. After observing that t = 0.05 worsens the representational-bias metric ∆G, the paper selects t = 1.00 specifically because it 'tends to preserve more factual signal', then offers the resulting ∆G values (German 3.7 → -10.1; Czech 2.1 → 7.2) as evidence of factual-gender preservation. The evidence is the same evaluation quantity that motivated the selection, so the reported outcome is a post-hoc fit rather than an independent confirmation. In addition, ∆G is a signed difference in translation F1 between masculine and feminine entities; it does not measure preservation of factual gender, and a negative value is a bias reversal, not a preservation win.

  2. self definitional [Section 2.4 and Table 3]
    "We propose to nullify dimensions of X with t times higher correlation with Zf than Zb ... In modified colspace projection ˜PWΣ, we only consider the column with t times higher variance with Zf than with Zb. Thus the final Dual Debiasing LEACE projection ... will to large extent preserve the protected feature while reliably erasing bias. ... Such intervention in the model erases ≈ 99% of covariance with a stereotypical signal while keeping over 30% of covariance with a factual gender signal."

    The algorithm defines 'preservation' as retaining columns whose factual-covariance exceeds stereotypical-covariance by a factor t. Reporting that the resulting projection keeps a certain fraction of factual covariance (Table 3) is therefore largely a restatement of the selection rule, not an independent discovery: the retained columns were chosen because they carry Zf. This makes the covariance-level 'factual preservation' evidence self-definitional. The behavioral af coefficients in Table 1 are more independent because they use held-out professions, so this step is partial rather than total.

full rationale

The mathematical core of the paper is not circular: DAMA-LEACE (Theorem 2) is proved from stated least-squares and exogeneity assumptions, and the Dual-Debiasing theorem (Theorem 3) explicitly abandons its conditional-independence guarantee in the practical algorithm. The English language-modeling results compare methods on held-out professions and external benchmarks, giving independent empirical content to the debiasing and general-performance claims. The main circularity is concentrated on the multilingual factual-preservation claim: the feature-to-bias threshold is selected after observing the ∆G metric, and the same metric is then cited as evidence, while ∆G does not actually measure factual-gender preservation. A secondary self-definitional element appears in Table 3, where covariance retention is largely a consequence of the projection construction. Overall the paper is not globally circular, but the translation-side dual claim is partially forced by hyperparameter choice and metric interpretation, warranting a score of 5.

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

The central claim rests on a linear model of gender encoding in feed-forward layers and on a covariance-based threshold that is tuned empirically. The theoretical guarantee for the dual-preservation claim holds only under assumptions the authors concede are unrealistic, so the main support is empirical.

free parameters (3)
  • bias-to-feature threshold t = 0.05 (main), 1.00 (translation experiments)
    Controls how many dimensions are preserved based on relative covariance with factual vs. stereotypical gender. Chosen empirically from hyperparameter analysis in Figures 4 and 5.
  • number of edited layers = 9 (7B/8B), 11 (13B), 12 in some analyses
    Layer count is selected from two-thirds of the stack onward, and the top two layers are excluded. The number is tuned per model size to balance bias reduction and performance.
  • starting layer (two-thirds of stack) = 26th for 13B, 21st for smaller models
    Chosen following Limisiewicz et al. (2024); affects which layers are edited.
assumptions (4)
  • domain assumption Gauss-Markov least-squares assumptions (no multicollinearity, exogeneity, homoscedasticity) hold for the regression V = S U - epsilon
    Used in Theorem 2 and Section 2.3 to justify setting S = Sigma_{U,V} Sigma^{-1}_{U,U}. The paper cites Eaton (1983) but does not verify these conditions for transformer feed-forward layers.
  • domain assumption Feed-forward layers approximate the least-squares solution, i.e., FF nearly equals S
    Stated in Section 2.3: 'Due to gradient optimization in the model's pre-training, we assume that the feed-forward layer approximates the least solution.' This is load-bearing for treating the layer as a linear map.
  • domain assumption Gender signals (stereotypical and factual) are linearly encoded in latent space and can be separated by covariance
    Dual Debiasing uses covariance matrices W Sigma_{X,[Zf,Zb]} and SVD to identify dimensions; this assumes linear representational structure of gender concepts.
  • ad hoc to paper Conditional independence Zb perpendicular to Zf given X in Theorem 3
    The theorem's guarantee requires this assumption, but Section 2.4 states 'the assumption of conditional independence is strong and unlikely to hold' and the practical method abandons it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair Language Modeling and Translation." pith.science (2026). https://pith.science/paper/QDKIUY4N

@misc{pith2026250110150,
  author       = {Pith},
  title        = {Pith review of: Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair Language Modeling and Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDKIUY4N}},
  note         = {Machine review of arXiv:2501.10150}
}
read the original abstract

Mitigation of biases, such as language models' reliance on gender stereotypes, is a crucial endeavor required for the creation of reliable and useful language technology. The crucial aspect of debiasing is to ensure that the models preserve their versatile capabilities, including their ability to solve language tasks and equitably represent various genders. To address this issue, we introduce a streamlined Dual Dabiasing Algorithm through Model Adaptation (2DAMA). Novel Dual Debiasing enables robust reduction of stereotypical bias while preserving desired factual gender information encoded by language models. We show that 2DAMA effectively reduces gender bias in English and is one of the first approaches facilitating the mitigation of stereotypical tendencies in translation. The proposed method's key advantage is the preservation of factual gender cues, which are useful in a wide range of natural language processing tasks.

Figures

Figures reproduced from arXiv: 2501.10150 by the authors.

Figure 1
Figure 1. Dual character of gender signals en￾coded in language models: stereotypical cues are shown on the left, and factual gender cues are shown on the right-hand side. “Die Ärztin” and “der Arzt” are respectively female and male Ger￾man translation for “the doctor”. et al., 2022; Gallegos et al., 2024). Biases are man￾ifested in multiple ways, including unequal repre￾sentation (models are more likely to generate men￾tions… view at source ↗
Figure 2
Figure 2. Schema (b) shows DAMA intervention in a language model layer. (a) We show the model’s behavior when presented with a stereotypical prompt in three languages. Specifically, (c) shows the projections of the feed-forward latent vector (⃗u) onto the output space. With DAMA (lower arrow), we nullify the gender component of the representation. It results in balanced probabilities of gendered tokens in the model’s output, … view at source ↗
Figure 3
Figure 3. Visualization of dimensions and their variances related to stereotypical and factual gen￾der signals identified by Dual Debiasing algorithm in 26th layer of Llama 2 13B. The red dots denote the bias-to-feature threshold t = 0.05. In 2DAMA, the dimension is preserved if stereotypical covari￾ance is below the threshold. 3.4 General Performance Evaluation Language Modeling We evaluate perplexity on general domain texts… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The hyperparameter analysis for 2DAMA applied to Llama 2 13B model on performance and bias in language modeling. We measured bias on gendered prompts by linear coefficients: as and af , the language modeling capabilities are measured by perplexity. Stars mark the perfo…
Figure 5
Figure 5. Figure 5: The hyperparameter analysis for 2DAMA applied to ALMA-R 13B model on performance and bias in translation to German. We measured bias via WinoMT metrics ∆S and ∆G. The translation quality to Germna is measured by chrf on WMT-22. Stars mark the performance of the best se…
Figure 6
Figure 6. Figure 6: Visualization of dimensions and their variances related to stereotypical and factual gen￾der signals identified by Dual Debiasing algorithm across different layers of Llama 2 13B [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 13 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  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]

    Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. https://openreview.net/forum?id=awIpKpwTwF LEACE : Perfect linear concept erasure in closed form . In Thirty-seventh Conference on Neural Information Processing Systems

  4. [4]

    Zou, Venkatesh Saligrama, and Adam Tauman Kalai

    Tolga Bolukbasi, Kai - Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Tauman Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4349--4357

  5. [5]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. http://arxiv.org/abs/1803.05457 Think you have solved question answering? try arc, the ai2 reasoning challenge

  6. [6]

    Wallach, Jennifer T

    Maria De - Arteaga, Alexey Romanov, Hanna M. Wallach, Jennifer T. Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Cem Geyik, Krishnaram Kenthapadi, and Adam Tauman Kalai. 2019. https://doi.org/10.1145/3287560.3287572 Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting . In Proceedings of the Conference on Fairness, A...

  7. [7]

    Pieter Delobelle, Ewoenam Tokpo, Toon Calders, and Bettina Berendt. 2022. https://doi.org/10.18653/v1/2022.naacl-main.122 Measuring fairness with biased rulers: A comparative study on bias metrics for pre-trained language models . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human La...

  8. [8]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany...

Show all 44 references
  1. [9]

    Morris L. Eaton. 1983. https://hdl.handle.net/11299/199431 The Gauss - Markov Theorem in Multivariate Analysis . Technical report, University of Minnesota

  2. [10]

    Chin - Lun Fu, Zih - Ching Chen, Yun - Ru Lee, and Hung - yi Lee. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.199 A dapter B ias: P arameter-efficient T oken-dependent R epresentation S hift for A dapters in NLP T asks . In Findings of the Association for Computation...

  3. [11]

    Gallegos, Ryan A

    Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://doi.org/10.1162/coli_a_00524 Bias and fairness in large language models: A survey . Computational Linguistics, 50(3):1097--1179

  4. [12]

    Hila Gonen, Shauli Ravfogel, and Yoav Goldberg. 2022. https://doi.org/10.18653/v1/2022.repl4nlp-1.8 Analyzing gender representation in multilingual models . In Proceedings of the 7th Workshop on Representation Learning for NLP, pages 67--77, Dublin, Ireland. Association for Co...

  5. [13]

    Marlis Hellinger and Hadumod Bußmann, editors. 2002. https://benjamins.com/catalog/impact.10 Gender Across Languages: The Linguistic Representation of Women and Men, Volume 2 , volume 2 of Impact: Studies in Language and Society. John Benjamins Publishing Company, Amsterdam/Ph...

  6. [14]

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. https://proceedings.mlr.press/v97/houlsby19a.html Parameter-efficient transfer learning for NLP . In Proceedings of the 36th In...

  7. [15]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

  8. [16]

    Bar Iluz, Tomasz Limisiewicz, Gabriel Stanovsky, and David Mare c ek. 2023. https://doi.org/10.18653/v1/2023.ijcnlp-main.57 Exploring the impact of training data distribution and subword tokenization on gender bias in machine translation . In Proceedings of the 13th Internatio...

  9. [17]

    Tom Kocmi, Tomasz Limisiewicz, and Gabriel Stanovsky. 2020. https://aclanthology.org/2020.wmt-1.39 Gender coreference and bias evaluation at WMT 2020 . In Proceedings of the Fifth Conference on Machine Translation, pages 357--364, Online. Association for Computational Linguistics

  10. [18]

    Hadas Kotek, Rikker Dockum, and David Q. Sun. 2023. https://api.semanticscholar.org/CorpusID:261276445 Gender bias and stereotypes in large language models . Proceedings of The ACM Collective Intelligence Conference

  11. [19]

    Shahar Levy, Koren Lazar, and Gabriel Stanovsky. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.211 Collecting a large-scale gender bias dataset for coreference resolution and machine translation . In Findings of the Association for Computational Linguistics: EMNLP 2021...

  12. [20]

    Tomasz Limisiewicz and David Mare c ek. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.3 Don ' t forget about pronouns: Removing gender bias in language models without losing factual gender information . In Proceedings of the 4th Workshop on Gender Bias in Natural Language Pr...

  13. [21]

    Tomasz Limisiewicz, David Mare c ek, and Tom \'a s Musil. 2024. https://openreview.net/forum?id=XIZEFyVGC9 Debiasing algorithm through model adaptation . In The Twelfth International Conference on Learning Representations

  14. [22]

    Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. 2023. https://api.semanticscholar.org/CorpusID:261031244 An empirical study of catastrophic forgetting in large language models during continual fine-tuning . ArXiv, abs/2308.08747

  15. [23]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/6f1d43d5a82a37e89b0665b33bf3a182-Abstract-Conference.html Locating and Editing Factual Associations in GPT . In NeurIPS

  16. [24]

    Andonian, Yonatan Belinkov, and David Bau

    Kevin Meng, Arnab Sen Sharma, Alex J. Andonian, Yonatan Belinkov, and David Bau. 2023. https://openreview.net/pdf?id=MkbcAHIYgyS Mass-Editing Memory in a Transformer . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 202...

  17. [25]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. http://arxiv.org/abs/1609.07843 Pointer sentinel mixture models

  18. [26]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022. https://openreview.net/pdf?id=0DcZxeWfOPt Fast model editing at scale . In International Conference on Learning Representations

  19. [27]

    Aur \'e lie N \'e v \'e ol, Yoann Dupont, Julien Bezan c on, and Kar \"e n Fort. 2022. https://doi.org/10.18653/v1/2022.acl-long.583 F rench C row S -pairs: Extending a challenge dataset for measuring social bias in masked language models to a language other than E nglish . In...

  20. [28]

    OECD. 2023. https://doi.org/https://doi.org/https://doi.org/10.1787/7a7afb35-en Health at a Glance 2023

  21. [29]

    Vaidehi Patil, Peter Hase, and Mohit Bansal. 2024. https://openreview.net/forum?id=7erlRDoaV8 Can sensitive information be deleted from LLM s? objectives for defending against extraction attacks . In The Twelfth International Conference on Learning Representations

  22. [30]

    Maja Popovi \'c . 2015. https://doi.org/10.18653/v1/W15-3049 chr F : character n-gram F -score for automatic MT evaluation . In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 392--395, Lisbon, Portugal. Association for Computational Linguistics

  23. [31]

    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...

  24. [32]

    Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan Cotterell. 2022. https://proceedings.mlr.press/v162/ravfogel22a.html L inear A dversarial C oncept E rasure . In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA , volume...

  25. [33]

    Ricardo Rei, Jos \'e G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and Andr \'e F. T. Martins. 2022. https://aclanthology.org/2022.wmt-1.52 COMET -22: Unbabel- IST 2022 submission for the metrics shared task . In Pro...

  26. [34]

    Danielle Saunders and Bill Byrne. 2020. https://doi.org/10.18653/v1/2020.acl-main.690 Reducing gender bias in neural machine translation as a domain adaptation problem . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7724--773...

  27. [35]

    Beatrice Savoldi, Marco Gaido, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2021. https://doi.org/10.1162/tacl_a_00401 Gender Bias in Machine Translation . Transactions of the Association for Computational Linguistics, 9:845--874

  28. [36]

    Karolina Stanczak and Isabelle Augenstein. 2021. http://arxiv.org/abs/2112.14168 A Survey on Gender Bias in Natural Language Processing . CoRR, abs/2112.14168

  29. [37]

    Smith, and Luke Zettlemoyer

    Gabriel Stanovsky, Noah A. Smith, and Luke Zettlemoyer. 2019. https://doi.org/10.18653/v1/P19-1164 Evaluating gender bias in machine translation . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1679--1684, Florence, Italy. Ass...

  30. [38]

    Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, D

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

  31. [39]

    Oskar Van Der Wal, Jaap Jumelet, Katrin Schulz, and Willem Zuidema. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.8 The birth of bias: A case study on the evolution of gender bias in an E nglish language model . In Proceedings of the 4th Workshop on Gender Bias in Natural La...

  32. [40]

    Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart M. Shieber. 2020. http://arxiv.org/abs/2004.12265 Causal Mediation Analysis for Interpreting Neural NLP: The Case of Gender Bias . CoRR, abs/2004.12265

  33. [41]

    Zuidema, and Katrin Schulz

    Oskar van der Wal, Dominik Bachmann, Alina Leidinger, Leendert van Maanen, Willem H. Zuidema, and Katrin Schulz. 2024. https://doi.org/10.1613/JAIR.1.15195 U ndesirable B iases in NLP: A ddressing C hallenges of M easurement . J. Artif. Intell. Res., 79:1--40

  34. [42]

    Haoran Xu, Amr Sharaf, Yunmo Chen, Weiting Tan, Lingfeng Shen, Benjamin Van Durme, Kenton Murray, and Young Jin Kim. 2024. http://arxiv.org/abs/2401.08417 Contrastive preference optimization: Pushing the boundaries of llm performance in machine translation

  35. [43]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. https://doi.org/10.18653/v1/N18-2003 Gender bias in coreference resolution: Evaluation and debiasing methods . In Proceedings of the 2018 Conference of the North A merican Chapter of the Associati...

  36. [44]

    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...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.