Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

LFTF: Locating First and Then Fine-Tuning for Mitigating Gender Bias in Large Language Models

T0 review · 4 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 Fine-tuning a single transformer block, chosen by how much its hidden states change, reduces measured profession-linked gender bias in LLMs by about 70 percent while leaving general ability nearly flat.

desk verdict The core debiasing claim is undercut by a loss function that cannot produce the reported balanced results; the datasets and locate-then-fine-tune idea have value but need major revision. read the letter →

arxiv 2505.15475 v1 pith:NF6GCKWY submitted 2025-05-21 cs.CL cs.AI

Zhanyue Qin , Yue Ding , Deyuan Liu , Qingbin Liu , Junxian Cai , Xi Chen , Zhiying Tu , Dianhui Chu
show 2 more authors
Cuiyun Gao Dianbo Sui
This is my paper
classification cs.CLcs.AI
keywords genderbiaslargelanguagemodelsdebiasingfine-tuninglocalizationfairnessevaluationtransformerblocksbenchmarkdatasets
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that profession-linked gender bias in a large language model can be substantially reduced by changing one transformer block. The LFTF algorithm first scores every block with BMI, defined as one minus the cosine similarity between the block's input and output hidden states, then fine-tunes the highest-scoring block with a loss built from the probabilities the model assigns to the tokens 'he' and 'she'. On Qwen2.5-7B the average AFGB-Score drops from 0.3305 to about 0.0955, the UB-Score on gender-hinted prompts rises from 0.5321 to 0.6704, and MMLU accuracy stays nearly flat at 0.7137 versus 0.7239; similar reductions are reported for Meta-Llama3-8B and Vicuna-7B-v1.5. The paper also introduces GenBiasEval and GenHintEval, a pair of datasets with two quantitative scores, as tools for checking both absolute gender fairness and consistency with explicit gender hints.

What carries the argument

The load-bearing object is the BMI score, defined for the $i$-th block as $\mathrm{BMI}_i = 1 - \frac{H_{i,l}^{\mathsf{T}} H_{i+1,l}}{\|H_{i,l}\|_2 \|H_{i+1,l}\|_2}$, that is, one minus the cosine similarity between the hidden state before and after the block. This score is the only criterion that selects which block to fine-tune, so the whole method depends on it ranking blocks by genuine gender-bias relevance rather than by generic representation change. The fine-tuning loss is $L = P(\text{he} \mid p, M) + P(\text{she} \mid p, M)$, which penalizes high next-token probability assigned to either gendered pronoun. The new datasets GenBiasEval and GenHintEval, with the AFGB-Score and UB-Score metrics, provide the measurements that the paper uses to show that the selected single block carries the debiasing effect.

What would settle it

Compute the BMI ordering on the gender-bias prompts and on gender-neutral or scrambled prompts of the same length and format; if the top-ranked block is the same in both orderings, the locating step is not responding to gender content, and the same debiasing would be obtained by fine-tuning that block without any locating stage.

Watch

Extended reading notes

Core claim

The paper's central claim is that the block in a transformer most responsible for gender bias can be found without any gender-labeled supervision, and that fine-tuning just that block is enough to debias the model. The locating stage computes BMI for every block; the block with the largest change between its input and output hidden states is declared most bias-relevant. The fine-tuning stage updates only that block using the loss $L = P(\text{he} \mid p, M) + P(\text{she} \mid p, M)$. According to the reported experiments, this one-block intervention lowers Qwen2.5-7B's average AFGB-Score on GenBiasEval from 0.3305 to about 0.0955, raises its UB-Score on GenHintEval from 0.5321 to 0.6704, and keeps MMLU accuracy essentially unchanged (0.7239 to 0.7137). The same recipe is reported to work on Meta-Llama3-8B and Vicuna-7B-v1.5, which the paper takes as evidence that block-level localization plus targeted fine-tuning generalizes across model families.

Load-bearing premise

The load-bearing premise is that a block's BMI score, one minus the cosine similarity of its input and output hidden states, measures that block's relevance to gender bias, even though the score contains no gender-specific information; if this proxy fails, the locating stage selects an arbitrary block and the method reduces to fine-tuning a single, probably last, layer.

Editorial extensions

If this is right

  • A practical debiasing pass could need only one block's weights, making it far cheaper than full fine-tuning and easier to audit than whole-model retraining.
  • The AFGB-Score and UB-Score pair lets downstream users check two things at once: that profession prompts no longer produce a large 'he' versus 'she' asymmetry, and that explicit gender hints are still followed.
  • Because the locating and loss logic does not depend on the model family, the method should transfer to other decoder-only transformers; the paper reports confirming results on Qwen, Llama, and Vicuna models.
  • Existing model-editing baselines, which can flip a biased model into an anti-biased one, fail on these metrics, so any successful debiasing method must balance the two pronoun probabilities rather than invert them.
  • The preserved general-ability claim is task-dependent: on MMLU, HellaSwag, BoolQ, RACE, CMMLU, C-EVAL, HumanEval, and MBPP the paper reports near-flat results, but on GSM8K and GSM-Plus it reports a noticeable decline.

Reading between the lines

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

  • An inference the paper does not itself draw: because BMI contains no gender-specific signal, the locating stage would likely rank the same block for almost any prompt distribution, so the reported effect may owe more to fine-tuning the final block than to discovering a genuinely bias-specific locus.
  • A testable consequence of the loss design is that a debiased model should not only equalize 'he' and 'she' but also shift probability mass away from gendered pronouns entirely; if instead the model simply flattens its output distribution, its scores on other next-token tasks should change in measurable ways.
  • The evaluation is binary and profession-centered, so the numbers do not speak to non-binary gender, gender-neutral constructions, or non-profession bias; replacing the target tokens and prompt templates with other protected attributes is a direct extension that the paper only sketches.
  • Because the method's compute cost is low, it could be combined with low-rank adapters or used repeatedly across different protected attributes, but such combinations are not tested in the paper.
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 / 5 minor

Summary. The paper proposes two evaluation datasets, GenBiasEval and GenHintEval, with associated metrics AFGB-Score and UB-Score, and a debiasing algorithm LFTF that first ranks transformer blocks by a Block Mitigating Importance Score (BMI) and then fine-tunes the top block with a loss defined as the sum of the next-token probabilities of "he" and "she". Experiments on Qwen2.5-7B, Meta-Llama3-8B, and Vicuna-7B-v1.5 report reduced AFGB-Score and roughly preserved general capability, with the claim that LFTF achieves balanced debiasing without significant shortcomings.

Significance. The datasets and metrics could be useful resources for the community, and the broad idea of locating bias-related internal blocks aligns with current modularity-based model editing research. The paper also reports an unusual empirical finding: fine-tuning a single block with a two-term probability loss yields balanced outputs. However, the central mechanism is not credible as described: the stated loss cannot produce the reported balanced behavior, and the BMI-based locating stage lacks gender-specific validation. If the loss and locating procedure were corrected and independently validated, the contribution could be meaningful, but in its current form the main debiasing claim is unsupported.

major comments (4)
  1. [Section 4.2, Eq. (5)] The loss L = P("he" | p, M) + P("she" | p, M) is a sum, not a balance term: its minimum is attained when both probabilities are driven to zero, and nothing in the expression rewards equality at nonzero probabilities. The reported results are the opposite: in Figure 3, the nurse example moves from P(he)=0.0478 and P(she)=0.5023 (sum 0.5501) to P(he)=0.5009 and P(she)=0.4982 (sum 0.9991), so the quantity that Eq. (5) minimizes actually increases. Table 8 confirms the expected collapse for single-term losses: both "w/o he" and "w/o she" produce AFGB-Score 0.9999 and UB-Score 0.0000, exactly the anti-bias collapse that minimizing one probability would cause; the full two-term loss should collapse both probabilities together. No additional term, weight, or constraint in the manuscript converts Eq. (5) into a balancing objective. As written, the method cannot produce the claimed balanced outputs; either the implemented loss differs from Eq. (5) or the reported numbers do not come from the stated procedure.
  2. [Section 4.1, Eq. (4)] BMI_i = 1 - cosine_similarity(H_{i,l}, H_{i+1,l}) measures how much the hidden state changes across a block, but it contains no gender-specific signal; the same score would rank blocks for any task or bias. The only validation in Appendix A.1 is the variance of BMI values across random samples, which checks stability, not whether the top-ranked block actually encodes gender stereotypes. Without a causal or correlational validation against gender-bias-related behavior, the locating stage may select an arbitrary high-change block, and the method reduces to fine-tuning the last block. The claim that the last block is "most related to gender bias" is therefore unsupported.
  3. [Section 5.1, Tables 4 and 6] The conclusion that LFTF achieves "very balanced results and no significant shortcomings" is not supported by the paper's own tables. For Qwen2.5-7B, GenHintEval UB-Score improves from 0.5321 to 0.6704, but for Meta-Llama3-8B it drops from 0.5265 to 0.3895 and for Vicuna-7B-v1.5 it drops from 0.7438 to 0.6613. The paper acknowledges these drops as "slight decline" in Section 5.3, but they are substantial relative to the claimed balance and are a clear shortcoming. Additionally, Table 5 shows large declines in mathematical reasoning (GSM8K: 0.5019 to 0.3692; GSM-Plus: 0.3182 to 0.2140), which undermines the blanket claim of no significant capability loss.
  4. [Section 3.1, Eq. (1) and Section 4.2, Eq. (5)] The evaluation metric AFGB-Score averages |P("he") - P("she")| over GenBiasEval, while the fine-tuning loss directly minimizes the sum of the same two probabilities on the training split. Improving a held-out split of the same template distribution is therefore partly by construction, and the paper does not demonstrate debiasing on an independent bias benchmark. GenHintEval measures consistency with gender hints, which is a different construct, and the results there are mixed. An evaluation on an external benchmark such as StereoSet or WinoBias, or under a distribution shift, would be needed to establish that LFTF generalizes beyond the training-template family.
minor comments (5)
  1. [Throughout] The acronym is inconsistent: "LFTF" and "LTFT" are used interchangeably (for example, in the Abstract, Section 4, and Section 6).
  2. [Table 8] The improvement value for "LFTF w/o ATT" on Word-Scale is shown as (+0.652) but should be (+0.0652) to match the difference 0.1456 - 0.0804.
  3. [Section 5.2 and Table 4] The text says that MMLU accuracy "does not decline," but Table 4 shows a drop from 0.7239 to 0.7137, a decline of 0.0102.
  4. [Section 3.1 and Section 2.2] There are several typos, including "similed" instead of "smiled" in Step (2) of the dataset construction and "mehthod" instead of "method" in Related Work.
  5. [Section 5.4] The dataset name "GenEvalBias" appears in the first sentence; it should be "GenBiasEval".

Circularity Check

1 steps flagged · score 4.0 of 10

AFGB reduction is partly built into the training objective, but external benchmarks keep the central claim from being fully circular.

  1. fitted input called prediction [Section 3.1 Eq. 1; Section 4.2 Eq. 5; Section 5.1 Table 4]
    "L = P (“he” | p, M) +P (“she” | p, M) (5) ... AF GB− Score = X p∈D |P (“he” | p, M) − P (“she” | p, M)| N um(D) (1) ... our proposed LFTF algorithm can achieve strong performance on both GenBiasEval and GenHintEval, with very balanced results and no significant shortcomings."

    Eq. 5's loss is a function of P(he|p,M) and P(she|p,M), and Eq. 1's AFGB-Score is the average absolute difference of those same two probabilities. Thus training with Eq. 5 can trivially reduce AFGB-Score by driving both probabilities toward zero (the global minimum of L), without establishing any balanced gender behavior. The reported post-LFTF AFGB drop (0.3305 to about 0.0955) is therefore partly a direct consequence of optimizing the very quantities the metric averages, rather than an independent validation of the located block. The paper's own case study shows P(he)=0.5009, P(she)=0.4982 after LFTF, which increases the sum minimized by Eq. 5; the balancing claim is not derivable from the stated loss.

full rationale

The LFTF pipeline is mostly self-contained: training and evaluation are separated (GenBiasEval 2:1:2 split; GenHintEval and MMLU/9-task suites are held out), and general-capability numbers are external. The self-citations to Qin et al. (2024) for modularity and for the loss design are inspirational, not load-bearing proofs, so they do not raise the score. One circularity-adjacent step is flagged: Eq. 5 trains directly on P(he) and P(she), the exact probabilities whose absolute difference defines the AFGB-Score in Eq. 1; a model minimizing Eq. 5 can therefore lower AFGB-Score by suppressing both tokens, and this is what makes the 'debiasing' result partly by construction rather than a discovered property of the located block. Note also that the reported balanced case (0.5009/0.4982) increases the sum minimized by Eq. 5, so if anything the stated loss predicts collapse, not balance; this is a correctness inconsistency rather than a circularity and is not double-counted in the score.

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

No new entities such as particles, forces, or dimensions are introduced. The BMI score is a derived statistic, not an entity. The free parameters are dataset construction choices and hyperparameters, plus the manual profession filter, which materially shape the benchmark. The most consequential axiom is that BMI's generic representation-change score corresponds to gender-bias relevance, which is asserted without gender-specific evidence.

free parameters (3)
  • manual profession filtering = 320 to 262 professions
    Subjective exclusion of gendered or composition-ambiguous professions (e.g., actress, salesman) shapes the benchmark distribution; the rule is not formalized (Section 3.1, Step 1).
  • prompt templates = 9 for GenBiasEval, 3 for GenHintEval
    Generated by GPT-4o-mini with a few examples; the number and content are hand-chosen, affecting both datasets (Section 3.1, Step 2 and Section 3.2).
  • fine-tuning hyperparameters = lr=1e-5, epochs=2, batch=32 for Qwen2.5-7B
    Reported for Qwen only; the other two models do not list hyperparameters, so the procedure is underspecified (Section 5.1).
assumptions (4)
  • domain assumption Gender is binary (male/female)
    The metrics and datasets only consider 'he' and 'she'; acknowledged in Section 7.
  • domain assumption LLMs are modular, with specific blocks handling specific tasks
    Invoked in Section 4 to justify locating a single bias-related block; cites Yu et al. (2023) and Qin et al. (2024) without independent verification.
  • ad hoc to paper BMI (1 - cosine similarity of block input/output hidden states) identifies gender-bias-relevant blocks
    Section 4.1, Eq. 4; no gender-specific signal is used, and the validation in Appendix A.1 only checks variance across samples, not whether the top block encodes gender bias.
  • ad hoc to paper Minimizing L = P(he) + P(she) reduces gender bias
    Section 4.2, Eq. 5; this is the stated objective, but its behavior is inconsistent with the reported results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LFTF: Locating First and Then Fine-Tuning for Mitigating Gender Bias in Large Language Models." pith.science (2026). https://pith.science/paper/NF6GCKWY

@misc{pith2026250515475,
  author       = {Pith},
  title        = {Pith review of: LFTF: Locating First and Then Fine-Tuning for Mitigating Gender Bias in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NF6GCKWY}},
  note         = {Machine review of arXiv:2505.15475}
}
read the original abstract

Nowadays, Large Language Models (LLMs) have attracted widespread attention due to their powerful performance. However, due to the unavoidable exposure to socially biased data during training, LLMs tend to exhibit social biases, particularly gender bias. To better explore and quantifying the degree of gender bias in LLMs, we propose a pair of datasets named GenBiasEval and GenHintEval, respectively. The GenBiasEval is responsible for evaluating the degree of gender bias in LLMs, accompanied by an evaluation metric named AFGB-Score (Absolutely Fair Gender Bias Score). Meanwhile, the GenHintEval is used to assess whether LLMs can provide responses consistent with prompts that contain gender hints, along with the accompanying evaluation metric UB-Score (UnBias Score). Besides, in order to mitigate gender bias in LLMs more effectively, we present the LFTF (Locating First and Then Fine-Tuning) algorithm.The algorithm first ranks specific LLM blocks by their relevance to gender bias in descending order using a metric called BMI (Block Mitigating Importance Score). Based on this ranking, the block most strongly associated with gender bias is then fine-tuned using a carefully designed loss function. Numerous experiments have shown that our proposed LFTF algorithm can significantly mitigate gender bias in LLMs while maintaining their general capabilities.

Figures

Figures reproduced from arXiv: 2505.15475 by the authors.

Figure 1
Figure 1. The detailed visualization of the construction [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The visualization of the BMI values for each [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. A case study of the Meta-Llama3-8B using [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in 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. Knowing Bias, Doing Better: Mitigating Social Bias in LLMs via Know-Bias Neuron Enhancement

    cs.AI 2026-01 conditional novelty 6.0 of 10

    Amplifying neurons that encode bias awareness, found with 45 yes/no questions, reduces gender/race/religion bias in three LLMs while roughly preserving general reasoning.

Reference graph

Works this paper leans on

70 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    https://api.semanticscholar.org/CorpusID:268232499 The claude 3 model family: Opus, sonnet, haiku

  2. [2]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  3. [3]

    Giuseppe Attanasio, Debora Nozza, Dirk Hovy, and Elena Baralis. 2022. Entropy-based attention regularization frees unintended bias mitigation from lists. arXiv preprint arXiv:2203.09192

    work page Pith review arXiv 2022

  4. [4]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. 2021. http://arxiv.org/abs/2108.07732 Program synthesis with large language models

    arXiv 2021

  5. [5]

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29

    2016

  6. [6]

    Shikha Bordia and SamuelR. Bowman. 2019. Identifying and reducing gender bias in word-level language models. Cornell University - arXiv,Cornell University - arXiv

    work page 2019

  7. [7]

    Bryson, and Arvind Narayanan

    Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. https://doi.org/10.1126/science.aal4230 Semantics derived automatically from language corpora contain human-like biases . Science, page 183–186

    doi:10.1126/science.aal4230 2017

  8. [8]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. 2024. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1--45

    2024

Show all 70 references
  1. [9]

    Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. 2020. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10800--10809

    2020

  2. [10]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott G...

    2021 arXiv

  3. [11]

    Pengyu Cheng, Wen Hao, Hsiang-Yu Yuan, Shijing Si, and Lawrence Carin. 2021. Fairfil: Contrastive neural debiasing method for pretrained text encoders. Learning,Learning

    2021

  4. [12]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. http://arxiv.org/abs/1905.10044 Boolq: Exploring the surprising difficulty of natural yes/no questions

    2019 arXiv

  5. [13]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. http://arxiv.org/abs/2110.14168 Training verifiers to solve math word problems

    2021 arXiv

  6. [14]

    Emily Dinan, Angela Fan, Adina Williams, Jack Urbanek, Douwe Kiela, and Jason Weston. 2019. Queens are powerful too: Mitigating gender bias in dialogue generation. arXiv preprint arXiv:1911.03842

    2019 arXiv

  7. [15]

    Tommaso Dolci, Fabio Azzalini, and Mara Tanelli. 2023. Improving gender-related fairness in sentence encoders: A semantics-based approach. Data Science and Engineering, 8(2):177--195

    2023

  8. [16]

    Xiangjue Dong, Yibo Wang, Philip S Yu, and James Caverlee. 2024. Disclosure and mitigation of gender bias in llms. arXiv preprint arXiv:2402.11190

    2024 arXiv

  9. [17]

    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 Zha...

    2024 arXiv

  10. [18]

    Satyam Dwivedi, Sanjukta Ghosh, and Shivam Dwivedi. 2023. Breaking the bias: Gender fairness in llms using prompt engineering and in-context learning. Rupkatha Journal on Interdisciplinary Studies in Humanities, 15(4)

    2023

  11. [19]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. 2024. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1--79

    2024

  12. [20]

    Michael Gira, Ruisu Zhang, and Kangwook Lee. 2022. Debiasing pre-trained language models via efficient fine-tuning. In Proceedings of the Second Workshop on Language Technology for Equality, Diversity and Inclusion, pages 59--69

    2022

  13. [21]

    Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. 2020. Mutant: A training paradigm for out-of-distribution generalization in visual question answering. arXiv preprint arXiv:2009.08566

    2020 arXiv

  14. [22]

    Akshat Gupta, Sidharth Baskaran, and Gopala Anumanchipalli. 2024. http://arxiv.org/abs/2403.07175 Rebuilding rome : Resolving model collapse during sequential model editing

    2024 arXiv

  15. [23]

    Xudong Han, Timothy Baldwin, and Trevor Cohn. 2021. https://doi.org/10.18653/v1/2021.findings-acl.41 Decoupling adversarial training for fair nlp . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021

    2021 doi

  16. [24]

    Zexue He, Bodhisattwa Prasad Majumder, and Julian McAuley. 2021. Detect and perturb: Neutral rewriting of biased and sensitive text via gradient-based decoding. arXiv preprint arXiv:2109.11708

    2021 arXiv

  17. [25]

    Dong Huang, Qingwen Bu, Jie Zhang, Xiaofei Xie, Junjie Chen, and Heming Cui. 2023 a . Bias assessment and mitigation in llm-based code generation. arXiv preprint arXiv:2309.14345

    2023 arXiv

  18. [26]

    Jianqiang Huang, Yu Qin, Jiaxin Qi, Qianru Sun, and Hanwang Zhang. 2022. Deconfounded visual grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 998--1006

    2022

  19. [27]

    Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. 2023 b . http://arxiv.org/abs/2305.08322 C-eval: A multi-level multi-discipline chinese evaluation suite for ...

    2023 arXiv

  20. [28]

    Przemyslaw Joniak and Akiko Aizawa. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.6 Gender biases and where to find them: Exploring gender bias in pre-trained transformer-based language models using movement pruning . In Proceedings of the 4th Workshop on Gender Bias in Natu...

    2022 doi

  21. [29]

    Jean Kaddour, Joshua Harris, Maximilian Mozes, Herbie Bradley, Roberta Raileanu, and Robert McHardy. 2023. Challenges and applications of large language models. arXiv preprint arXiv:2307.10169

    2023 arXiv

  22. [30]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. http://arxiv.org/abs/1412.6980 Adam: A method for stochastic optimization

    2017 arXiv

  23. [31]

    Camila Kolling, Martin More, Nathan Gavenski, Eduardo Pooch, Ot \'a vio Parraga, and Rodrigo C Barros. 2022. Efficient counterfactual debiasing for visual question answering. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3001--3010

    2022

  24. [32]

    Shachi H Kumar, Saurav Sahay, Sahisnu Mazumder, Eda Okur, Ramesh Manuvinakurike, Nicole Beckage, Hsuan Su, Hung-yi Lee, and Lama Nachman. 2024. Decoding biases: Automated methods and llm judges for gender bias detection in language models. arXiv preprint arXiv:2408.03907

    2024 arXiv

  25. [33]

    Black, and Yulia Tsvetkov

    Keita Kurita, Nidhi Vyas, Ayush Pareek, AlanW. Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. Cornell University - arXiv,Cornell University - arXiv

    2019

  26. [34]

    Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. http://arxiv.org/abs/1704.04683 Race: Large-scale reading comprehension dataset from examinations

    2017 arXiv

  27. [35]

    Anne Lauscher, Tobias Lüken, and Goran Glavaš. 2021. http://arxiv.org/abs/2109.03646 Sustainable modular debiasing of language models

    2021 arXiv

  28. [36]

    Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. 2024 a . http://arxiv.org/abs/2306.09212 Cmmlu: Measuring massive multitask language understanding in chinese

    2024 arXiv

  29. [37]

    Qintong Li, Leyang Cui, Xueliang Zhao, Lingpeng Kong, and Wei Bi. 2024 b . http://arxiv.org/abs/2402.19255 Gsm-plus: A comprehensive benchmark for evaluating the robustness of llms as mathematical problem solvers

    2024 arXiv

  30. [38]

    Tomasz Limisiewicz, David Mareček, and Tomáš Musil. 2024. http://arxiv.org/abs/2310.18913 Debiasing algorithm through model adaptation

    2024 arXiv

  31. [39]

    Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. 2022. A causal debiasing framework for unsupervised salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1610--1619

    2022

  32. [41]

    Haochen Liu, Jamell Dacon, Wenqi Fan, Hui Liu, Zitao Liu, and Jiliang Tang. 2020. http://arxiv.org/abs/1910.10486 Does gender matter? towards fairness in dialogue systems

    2020 arXiv

  33. [42]

    Kyle Mahowald, Anna A Ivanova, Idan A Blank, Nancy Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. 2024. Dissociating language and thought in large language models. Trends in Cognitive Sciences

    2024

  34. [43]

    Bodhisattwa Prasad Majumder, Zexue He, and Julian McAuley. 2022. Interfair: Debiasing with natural language feedback for fair interpretable predictions. arXiv preprint arXiv:2210.07440

    2022 arXiv

  35. [44]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

    2022

  36. [45]

    Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. 2023. http://arxiv.org/abs/2210.07229 Mass-editing memory in a transformer

    2023 arXiv

  37. [46]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. http://arxiv.org/abs/2110.11309 Fast model editing at scale

    2022 arXiv

  38. [47]

    Moin Nadeem, Anna Bethke, and Siva Reddy. 2020. Stereoset: Measuring stereotypical bias in pretrained language models. arXiv: Computation and Language,arXiv: Computation and Language

    2020

  39. [48]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff ...

    2024 arXiv

  40. [49]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and...

    2022 arXiv

  41. [50]

    SunYoung Park, Kyuri Choi, Haeun Yu, and Youngjoong Ko. 2023. Never too late to learn: Regularizing gender bias in coreference resolution. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, pages 15--23

    2023

  42. [51]

    Rebecca Qian, Candace Ross, Jude Fernandes, Eric Smith, Douwe Kiela, and Adina Williams. 2022. Perturbation augmentation for fairer nlp. arXiv preprint arXiv:2205.12586

    2022 arXiv

  43. [52]

    Zhanyue Qin, Haochuan Wang, Zecheng Wang, Deyuan Liu, Cunhang Fan, Zhao Lv, Zhiying Tu, Dianhui Chu, and Dianbo Sui. 2024. Mitigating gender bias in code large language models via model editing. arXiv preprint arXiv:2410.07820

    2024 arXiv

  44. [53]

    Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti, Dario Onorati, and Fabio Massimo Zanzotto. 2023. http://arxiv.org/abs/2305.13862 A trip towards fairness: Bias and de-biasing in large language models

    2023 arXiv

  45. [54]

    Lauren Rhue, Sofie Goethals, and Arun Sundararajan. 2024. Evaluating llms for gender disparities in notable persons. arXiv preprint arXiv:2403.09148

    2024 arXiv

  46. [55]

    Andrew D Selbst, Danah Boyd, Sorelle A Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and abstraction in sociotechnical systems. In Proceedings of the conference on fairness, accountability, and transparency, pages 59--68

    2019

  47. [56]

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

    2024

  48. [57]

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

    2023 arXiv

  49. [58]

    Jialu Wang, Yang Liu, and Xin Eric Wang. 2021. Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. arXiv preprint arXiv:2109.05433

    2021 arXiv

  50. [59]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. 2024 a . A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345

    2024

  51. [60]

    Peng Wang, Ningyu Zhang, Bozhong Tian, Zekun Xi, Yunzhi Yao, Ziwen Xu, Mengru Wang, Shengyu Mao, Xiaohan Wang, Siyuan Cheng, Kangwei Liu, Yuansheng Ni, Guozhou Zheng, and Huajun Chen. 2024 b . http://arxiv.org/abs/2308.07269 Easyedit: An easy-to-use knowledge editing framework...

    2024 arXiv

  52. [61]

    Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, and Jundong Li. 2023. Knowledge editing for large language models: A survey. ACM Computing Surveys

    2023

  53. [62]

    Robert Wu and Vardan Papyan. 2024. Linguistic collapse: Neural collapse in (large) language models. arXiv preprint arXiv:2405.17767

    2024 arXiv

  54. [63]

    Wanli Yang, Fei Sun, Jiajun Tan, Xinyu Ma, Du Su, Dawei Yin, and Huawei Shen. 2024. The fall of rome: Understanding the collapse of llms in model editing. arXiv preprint arXiv:2406.11263

    2024 arXiv

  55. [64]

    Zhiwen You, HaeJin Lee, Shubhanshu Mishra, Sullam Jeoung, Apratim Mishra, Jinseok Kim, and Jana Diesner. 2024. Beyond binary gender labels: Revealing gender biases in llms through gender-neutral name predictions. arXiv preprint arXiv:2407.05271

    2024 arXiv

  56. [65]

    Charles Yu, Sullam Jeoung, Anish Kasi, Pengfei Yu, and Heng Ji. 2023. Unlearning bias in language models by partitioning gradients. In Findings of the Association for Computational Linguistics: ACL 2023, pages 6032--6048

    2023

  57. [66]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. http://arxiv.org/abs/1905.07830 Hellaswag: Can a machine really finish your sentence?

    2019 arXiv

  58. [67]

    Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. http://arxiv.org/abs/2306.05685 Judging llm-as-a-judge with mt-bench and chatbot arena

    2023 arXiv

  59. [68]

    Fan Zhou, Yuzhou Mao, Liu Yu, Yi Yang, and Ting Zhong. 2023. Causal-debias: Unifying debiasing in pretrained language models and fine-tuning via causal invariant learning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Lon...

    2023

  60. [69]

    Ran Zmigrod, Sabrina J Mielke, Hanna Wallach, and Ryan Cotterell. 2019. Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology. arXiv preprint arXiv:1906.04571

    2019 arXiv

  61. [70]

    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...

  62. [71]

    write newline

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

Pith tools

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