Pith. sign in

REVIEW 4 major objections 4 minor 38 references

KLAAD: Refining Attention Mechanisms to Reduce Societal Bias in Generative Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper seeks to establish that societal bias in decoder-only generative language models can be reduced by aligning attention distributions between stereotypical and anti-stereotypical sentence pairs, without directly editing attention…

desk verdict Plausible new attention-alignment objective for decoder-only models, but the evidence doesn't support the 'minimal impact' claim — Gemma's BBQ numbers look like an Unknown-answer artifact, and BOLD is only evaluated on Llama. read the letter →

arxiv 2507.19962 v1 pith:6IUIISFQ submitted 2025-07-26 cs.CL

Seorin Kim , Dongyoung Lee , Jaejin Lee This is my paper
classification cs.CL
keywords societalbiasattentionmechanismsKLdivergencegenerativelanguagemodelsdebiasingStereoSetBBQbenchmarkBOLD
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 a generative language model becomes less biased if it is fine-tuned to pay attention to stereotypical and anti-stereotypical sentences in the same way, rather than having its attention weights rewritten by arithmetic interventions. Its method, KLAAD, uses StereoSet sentence triplets and a composite loss made of cross-entropy, KL divergence on final-layer attention distributions, and a triplet loss on hidden states. The authors report improved ambiguous-context accuracy and near-zero bias scores on BBQ, more emotionally neutral open-ended generations on BOLD, and a narrower raw log-probability gap on CrowS-Pairs. The broader claim is that attention-level alignment is a viable debiasing lever for generative models, and that generation-based and distribution-level evaluations are more informative than a single Stereotype Score.

What carries the argument

The central mechanism is a KL-divergence loss between the softmax-normalized final-layer attention distributions of a stereotypical sentence and its anti-stereotypical counterpart, written as $\mathcal{L}_{KL} = D_{KL}(\mathrm{Attn}_{anti} \| \mathrm{Attn}_{stereo})$. It is paired with a cross-entropy loss over the coherent sentences to preserve language modeling ability and a triplet loss on final-layer hidden states that uses the stereotype as anchor, the anti-stereotype as positive, and an unrelated sentence as negative. This composite objective guides fine-tuning so the model reduces attention gaps on bias-sensitive tokens while preserving fluency and coherence, instead of directly modifying attention weights.

What would settle it

A reader could falsify the central claim by fine-tuning the same three models on StereoSet triplets and then measuring attention divergence on a held-out set of stereotype pairs that were not used during training; if the KL loss has collapsed on training templates but attention divergence on held-out pairs or bias scores on unseen tasks do not improve, the transfer assumption is false.

Watch

Extended reading notes

Core claim

The central discovery is that bias in a decoder-only language model can be reduced by training it to make its final-layer attention distributions similar for stereotypical and anti-stereotypical input sentences, without applying any arithmetic operation directly to pretrained attention matrices. On triplets built from StereoSet, KLAAD fine-tunes Llama-3.2-3B, GPT-Neo-2.7B, and Gemma-2-2B with a loss that combines cross-entropy for fluency, KL divergence between the softmax-normalized attention distributions of the stereotype and anti-stereotype sentences, and a triplet loss that pulls coherent sentence representations together while pushing unrelated ones apart. The authors report that this improves ambiguous-context accuracy and lowers bias scores on BBQ, produces more emotionally neutral continuations on BOLD across gender, race, profession, political ideology, and religious ideology, and narrows the mean absolute log-probability difference between stereotypical and anti-stereotypical pairs in CrowS-Pairs. They also argue that the standard CrowS-Pairs Stereotype Score can hide whether a model has developed a stronger one-sided preference, and that raw log-probability differences better reveal how a debiasing method actually changes model behavior.

Load-bearing premise

The load-bearing assumption is that making final-layer attention distributions match on short StereoSet sentence pairs transfers to reduced stereotypical behavior in very different tasks like ambiguous-context question answering and open-ended generation, and if that transfer fails the benchmark improvements would not reflect a genuine debiasing mechanism.

Editorial extensions

If this is right

  • Fine-tuning with attention alignment can reduce stereotype-driven attention differences on bias-sensitive tokens, as shown by the heatmap comparisons before and after KLAAD.
  • BBQ ambiguous-context accuracy can improve while disambiguated-context accuracy remains competitive, meaning the model is less likely to rely on group stereotypes when the relevant information is missing.
  • Open-ended generations on BOLD become more emotionally neutral across gender, race, profession, political ideology, and religious ideology categories, extending the effect beyond the training template pairs.
  • On CrowS-Pairs, KLAAD reduces the mean absolute log-probability difference between stereotypical and anti-stereotypical sentences, whereas some baseline methods increase it, so an Stereotype Score near 50 can coexist with a stronger one-sided preference.
  • Removing any one loss component degrades the combined result: cross-entropy loss sustains language ability, KL loss drives fairness, and the triplet loss refines contextual understanding.

Reading between the lines

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

  • If the attention-alignment effect generalizes, a similar KL-attention objective could be applied to other paired or contrastive bias datasets, such as those for toxicity or dialect prejudice, though the paper does not test those settings.
  • The reliance on StereoSet-style triplets suggests the method is most effective when a bias-sensitive token is localized and the contrast is lexical; biases expressed through longer contexts or stylistic patterns may require a different objective.
  • A direct test of the paper's transfer assumption would be to measure attention divergence on held-out contrast pairs that were not used during training; the paper evaluates downstream benchmarks but does not explicitly report whether the attention alignment itself transfers to unseen sentence structures.
  • Because KLAAD is a one-epoch fine-tuning with small auxiliary loss weights, it could plausibly be combined with parameter-efficient training or preference-optimization pipelines, although those combinations are not explored here.
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 / 4 minor

Summary. The paper proposes KLAAD, a debiasing framework for decoder-only generative language models that fine-tunes a model with a composite loss combining cross-entropy, KL divergence between final-layer attention distributions of stereotypical and anti-stereotypical sentence pairs, and a triplet loss over hidden states. Training pairs are constructed from StereoSet, and evaluation is performed on BBQ, BOLD, and CrowS-Pairs using Llama-3.2-3B, GPT-Neo-2.7B, and Gemma-2-2B. The central claim is that KLAAD improves bias mitigation on BBQ and BOLD with minimal impact on language modeling quality, and that attention-level alignment is a principled solution for generative models. The paper also argues that the CrowS-Pairs stereotype score is an unreliable metric for generative models.

Significance. If the claims were established, KLAAD would be a useful contribution: it targets decoder-only models, which are underexplored in attention-based debiasing, and the idea of aligning attention distributions as a soft training signal is plausible. The paper has notable strengths: it evaluates three different model families, provides a full ablation study (Table 8), and reports an unusually complete hyperparameter grid in Table 7. The CrowS-Pairs critique, supported by the raw log-probability differences in Table 4, is a thoughtful addition to the evaluation-metrics discussion. However, the current evidence does not support the abstract's claim of consistent improved bias mitigation across models and benchmarks. The Gemma BBQ result is confounded by a likely 'Unknown-response' artifact, BOLD results are reported only for Llama-3.2-3B, and the hyperparameters are selected on the evaluation benchmarks themselves. These issues affect the central claims, so the significance of the paper as written is limited until they are addressed.

major comments (4)
  1. [Table 2; Appendix C.1] The Gemma-2-2B row of Table 2 is not evidence of bias mitigation until a response-type artifact is ruled out. A.Amb rises from 5.11 to 52.56 while A.Dis falls from 45.19 to 30.71; since Appendix C.1 states that the correct response in ambiguous contexts is typically 'Unknown', this pattern is exactly what an 'always-Unknown' policy would produce. The paper does not report the distribution of predicted answer types, accuracy restricted to non-Unknown outputs, or an 'always-Unknown' baseline, so the headline BBQ improvement for this model may be spurious.
  2. [Appendix B.1; Section 4.1] The hyperparameters (lambda_1, lambda_2, lambda_3, and margin) are selected by grid search using the evaluation benchmarks, and this selection on the test set inflates the reported results. For Gemma, the text says the configuration was chosen 'using disambiguated context accuracy as the primary criterion,' yet the selected configuration still loses 14.5 points of A.Dis (45.19 to 30.71), and every configuration in Table 7 sacrifices substantial A.Dis. Similarly, for Llama the configuration was chosen by prioritizing fairness metrics including ambiguous-context accuracy. A held-out validation split or nested model selection is required before the reported benchmark comparisons can be interpreted as unbiased estimates of KLAAD's performance.
  3. [Section 5.2; Tables 5, 10–14] The BOLD evaluation is reported only for Llama-3.2-3B; no sentiment or psycholinguistic-norm results are given for GPT-Neo-2.7B or Gemma-2-2B. The abstract and Section 5.1 nevertheless claim improved bias mitigation on both BBQ and BOLD and consistent performance across all three models. Without per-model BOLD results, the cross-model generalization claim is unsupported by the evidence in the paper.
  4. [Section 3.2, Eq. (3)] The KL loss aligns final-layer attention distributions on short StereoSet templates, but the paper provides no quantitative test that this alignment transfers to the disambiguated-QA and open-ended-generation settings of BBQ and BOLD; the only transfer evidence is a few selected heatmaps (Figure 1b and Figure 3). This could be an acceptable mechanism hypothesis, but it is currently presented as the explanation for benchmark gains, while those gains are themselves confounded by the response-type and hyperparameter-selection issues above.
minor comments (4)
  1. [Abstract; Section 3.2; Section 2.2] The abstract says KLAAD works 'without directly modifying model weights,' but the method fine-tunes the full model and therefore updates all weights. The intended contrast is with directly editing attention matrices, as clarified in Section 2.2; please rephrase the abstract to avoid the misleading implication.
  2. [Section 4.2] There is a typo: 'Wikipedea' should be 'Wikipedia' in the description of the Dropout baseline.
  3. [Table 4; Section 5.3] The CrowS-Pairs log-probability analysis is reported only for Llama-3.2-3B and lacks standard deviations or significance tests. The claim that KLAAD 'narrows the log-probability difference' should be qualified as a single-model observation until replicated on the other two models.
  4. [Table 3] The qualitative BOLD examples are hand-picked and should be described as illustrative; they are not by themselves quantitative evidence, and the main quantitative BOLD results are limited to Llama-3.2-3B.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: KLAAD's benchmark claims rest on external evaluations, with only a benchmark-selection validity concern.

full rationale

The paper's derivation chain is self-contained. The KLAAD objective (Eq. 1: L = λ1·LCE + λ2·LKL + λ3·LTriplet) is a training loss defined on StereoSet triplets, and the central claims are evaluated on BBQ, BOLD, and CrowS-Pairs, which are external datasets not used to construct the training triplets. No equation defines any benchmark metric in terms of the training loss, and no fitted parameter from the grid search is renamed as a prediction: Table 7 reports hyperparameter configurations, while Table 2 reports benchmark results, so the reported outcomes are empirical measurements rather than consequences of the objective by construction. Figure 1(b)'s attention-difference reduction is a direct consequence of minimizing Eq. (3) on analogous sentence pairs, but the paper presents the benchmark numbers, not that heatmap, as its headline evidence, so this is not a self-definitional prediction. The paper contains no load-bearing self-citations by the present authors and invokes no uniqueness theorem or prior-work ansatz to force its choice of loss. The Limitations section candidly notes the English-only scope, the failure to address toxicity and microaggressions, and the risk that benchmark-defined fairness may marginalize communities; these are scope and ethical limitations, not admissions of circularity. The main validity concerns are the grid search over loss weights and margin conducted with the evaluation benchmarks in view (Section 4.1 and Appendix B.1, e.g., for Gemma 'using disambiguated context accuracy as the primary criterion for selection'), which risks inflating reported results through test-set selection, and the large Gemma A.Amb/A.Dis divergence, which may reflect an 'Unknown'-refusal artifact. Both are correctness and evaluation-validity concerns, not circular-derivation concerns under the stated hard rules.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several domain assumptions about the relationship between attention alignment and measured bias, and on free hyperparameters that are tuned on the evaluation benchmarks. No new physical or architectural entities are invented.

free parameters (4)
  • lambda_1 (CE loss weight) = Llama: 0.7; GPT-Neo: 0.5; Gemma: 0.9
    Selected by grid search on evaluation benchmarks (Appendix B.1, Table 7) to balance fairness and language ability; hence fitted to the test data.
  • lambda_2 (KL loss weight) = Llama: 0.15; GPT-Neo: 0.25; Gemma: 0.05
    Selected by grid search on evaluation benchmarks (Appendix B.1, Table 7) to balance fairness and language ability; hence fitted to the test data.
  • lambda_3 (Triplet loss weight) = Llama: 0.15; GPT-Neo: 0.25; Gemma: 0.05
    Selected by grid search on evaluation benchmarks (Appendix B.1, Table 7) to balance fairness and language ability; hence fitted to the test data.
  • Triplet margin = Llama: 0.3; GPT-Neo: 0.5; Gemma: 0.5
    Tuned in range [0.1, 0.5] via grid search on evaluation benchmarks (Appendix B.1, Table 7); hence fitted to the test data.
assumptions (5)
  • domain assumption Aligning attention distributions between stereotypical and anti-stereotypical sentences reduces stereotypical behavior in downstream generation.
    This assumption underlies the entire KL loss in Eq. (3) and is motivated only by the correlational evidence in Figure 1, not by a causal proof.
  • ad hoc to paper Final-layer softmax-normalized attention distributions are a sufficient and reliable signal for bias and a suitable target for alignment.
    Section 3.2 defines Attn_x as the softmax-normalized final-layer attention. The paper does not justify the choice of the final layer or the extra softmax normalization beyond saying it stabilizes training.
  • domain assumption StereoSet triplets (stereotype, anti-stereotype, unrelated) adequately represent the bias categories the model is trained on.
    The training set is built entirely from StereoSet (Section 3.1). The method assumes these templates are enough to debias the model on the categories evaluated in BBQ and BOLD.
  • domain assumption The evaluation metrics (BBQ bias score/accuracy, BOLD sentiment and VAD/BE5 norms, CrowS-Pairs stereotype score) capture the bias the method claims to reduce.
    Section 4.3 adopts these benchmarks as ground truth for bias. The paper later argues CrowS-Pairs is misleading, which raises a question about which benchmark should be trusted.
  • standard math KL divergence, cross-entropy, and triplet loss are standard and their gradients behave as expected in this fine-tuning setting.
    These are textbook objectives; no new mathematical assumptions are introduced beyond standard optimization practice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KLAAD: Refining Attention Mechanisms to Reduce Societal Bias in Generative Language Models." pith.science (2026). https://pith.science/paper/6IUIISFQ

@misc{pith2026250719962,
  author       = {Pith},
  title        = {Pith review of: KLAAD: Refining Attention Mechanisms to Reduce Societal Bias in Generative Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IUIISFQ}},
  note         = {Machine review of arXiv:2507.19962}
}
read the original abstract

Large language models (LLMs) often exhibit societal biases in their outputs, prompting ethical concerns regarding fairness and harm. In this work, we propose KLAAD (KL-Attention Alignment Debiasing), an attention-based debiasing framework that implicitly aligns attention distributions between stereotypical and anti-stereotypical sentence pairs without directly modifying model weights. KLAAD introduces a composite training objective combining Cross-Entropy, KL divergence, and Triplet losses, guiding the model to consistently attend across biased and unbiased contexts while preserving fluency and coherence. Experimental evaluation of KLAAD demonstrates improved bias mitigation on both the BBQ and BOLD benchmarks, with minimal impact on language modeling quality. The results indicate that attention-level alignment offers a principled solution for mitigating bias in generative language models.

Figures

Figures reproduced from arXiv: 2507.19962 by the authors.

Figure 1
Figure 1. (a) Attention differences between bias-sensitive and other tokens in pretrained Llama-3.2-3B model, (b) Attention weight differences from the final layer of Llama-3.2-3B on a CrowS-Pairs sentence pair: "Julie’s mom disowned her for being lesbian" (stereotypical) and "Julie’s mom disowned her for being straight" (anti￾stereotypical). Left: pretrained model; Right: model after applying KLAAD. The lighter color in the … view at source ↗
Figure 2
Figure 2. Overview of KLAAD. Embedding-based approaches, including FineDeb (Saravanan et al., 2023) and DeepSoftDe￾bias (Rakshit et al., 2025), attempt to neutralize bias within the embedding space. However, since they do not directly intervene in the generative process or attention mechanisms, their impact on final outputs can be limited or unclear. Other techniques include regularization or direct weight manipulation. For e… view at source ↗
Figure 3
Figure 3. Additional attention difference heatmaps showing KLAAD’s ability to equalize attention across identity [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 14 canonical work pages

  1. [1]

    Sid Black, Gao Leo, Phil Wang, Connor Leahy, and Stella Biderman. 2021. https://doi.org/10.5281/zenodo.5297715 GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow . If you use this software, please cite it using these metadata

    doi:10.5281/zenodo.5297715 2021

  2. [2]

    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 computer programmer as woman is to homemaker? debiasing word embeddings . In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc

    2016

  3. [3]

    Margaret M Bradley and Peter J Lang. 1994. Measuring emotion: the self-assessment manikin and the semantic differential. Journal of behavior therapy and experimental psychiatry, 25(1):49--59

    work page 1994

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, and 12 others. 2020. https://proceedings.neurips.cc/paper_fil...

    2020

  5. [5]

    Sven Buechel and Udo Hahn. 2016. Emotion analysis as a regression problem--dimensional models and their implications on emotion representation and metrical evaluation. In ECAI 2016, pages 1114--1122. IOS Press

    work page 2016

  6. [6]

    Ruizhe Chen, Yichen Li, Jianfei Yang, Yang Feng, Joey Tianyi Zhou, Jian Wu, and Zuozhu Liu. 2025. https://doi.org/10.18653/v1/2025.findings-naacl.39 Identifying and mitigating social bias knowledge in language models . In Findings of the Association for Computational Linguistics: NAACL 2025, pages 651--672, Albuquerque, New Mexico. Association for Computa...

    work page doi:10.18653/v1/2025.findings-naacl.39 2025

  7. [7]

    Pengyu Cheng, Weituo Hao, Siyang Yuan, Shijing Si, and Lawrence Carin. 2021. https://openreview.net/forum?id=N6JECD-PI5w Fairfil: Contrastive neural debiasing method for pretrained text encoders . In International Conference on Learning Representations

    work page 2021

  8. [8]

    Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. https://doi.org/10.1145/3442188.3445924 Bold: Dataset and metrics for measuring biases in open-ended language generation . In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, page 862–872, New ...

    arXiv 2021

Show all 38 references
  1. [9]

    Shaz Furniturewala, Surgan Jandial, Abhinav Java, Pragyan Banerjee, Simra Shahid, Sumit Bhatia, and Kokil Jaidka. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.13 thinking fair and slow: On the efficacy of structured prompts for debiasing language models . In Proceedings o...

    2024 doi

  2. [10]

    Yacine Gaci, Boualem Benatallah, Fabio Casati, and Khalid Benabdeslem. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.651 Debiasing pretrained text encoders by paying attention to paying attention . In Proceedings of the 2022 Conference on Empirical Methods in Natural Langu...

    2022 doi

  3. [11]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  4. [12]

    Yue Guo, Yi Yang, and Ahmed Abbasi. 2022. https://doi.org/10.18653/v1/2022.acl-long.72 Auto-debias: Debiasing masked language models with automated biased prompts . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Paper...

    2022 doi

  5. [13]

    Pengrui Han, Rafal Dariusz Kocielnik, Adhithya Prakash Saravanan, Roy Luoyao Jiang, Or Sharir, and Anima Anandkumar. 2024. https://openreview.net/forum?id=q5Ft9ZJtHm Chat GPT based data augmentation for improved parameter-efficient debiasing of LLM s . In First Conference on L...

    2024

  6. [14]

    Clayton Hutto and Eric Gilbert. 2014. Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Proceedings of the international AAAI conference on web and social media, volume 8, pages 216--225

    2014

  7. [15]

    Jingling Li, Zeyu Tang, Xiaoyu Liu, Peter Spirtes, Kun Zhang, Liu Leqi, and Yang Liu. 2024. Steering llms towards unbiased responses: A causality-guided debiasing framework. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models

    2024

  8. [16]

    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

    2024

  9. [17]

    Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Amancharla, and Anupam Datta. 2020. Gender bias in neural natural language processing. Logic, language, and security: essays dedicated to Andre Scedrov on the occasion of his 65th birthday, pages 189--202

    2020

  10. [18]

    Shenyu Lu, Yipei Wang, and Xiaoqian Wang. 2024. https://openreview.net/forum?id=jLIUfrAcMQ Debiasing attention mechanism in transformer without demographics . In The Twelfth International Conference on Learning Representations

    2024

  11. [19]

    Congda Ma, Tianyu Zhao, and Manabu Okumura. 2024. https://doi.org/10.18653/v1/2024.findings-acl.612 Debiasing large language models with structured knowledge . In Findings of the Association for Computational Linguistics: ACL 2024, pages 10274--10287, Bangkok, Thailand. Associ...

    2024 doi

  12. [20]

    Saif Mohammad. 2018. https://doi.org/10.18653/v1/P18-1017 Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 E nglish words . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1...

    2018 doi

  13. [21]

    Saif Mohammad and Peter Turney. 2010. https://aclanthology.org/W10-0204 Emotions evoked by common words and phrases: Using M echanical T urk to create an emotion lexicon . In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of E...

    2010

  14. [22]

    Saif M Mohammad. 2025. Nrc vad lexicon v2: Norms for valence, arousal, and dominance for over 55k english terms. arXiv preprint arXiv:2503.23547

    2025 arXiv

  15. [23]

    Mohammad and Peter D

    Saif M. Mohammad and Peter D. Turney. 2013. Crowdsourcing a word-emotion association lexicon. Computational Intelligence, 29(3):436--465

    2013

  16. [24]

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

    2021 doi

  17. [25]

    Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, C a g lar Gu l c ehre, and Bing Xiang. 2016. https://doi.org/10.18653/v1/K16-1028 Abstractive text summarization using sequence-to-sequence RNN s and beyond . In Proceedings of the 20th SIGNLL Conference on Computational Natural...

    2016 doi

  18. [26]

    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 challenge dataset for measuring social biases in masked language models . In Proceedings of the 2020 Conference on Empirical Methods in Na...

    2020 doi

  19. [27]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing sys...

    2022

  20. [28]

    Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel Bowman. 2022. https://doi.org/10.18653/v1/2022.findings-acl.165 BBQ : A hand-built bias benchmark for question answering . In Findings of the Association for ...

    2022 doi

  21. [29]

    Nirmalendu Prakash and Roy Ka-Wei Lee. 2023. https://doi.org/10.18653/v1/2023.blackboxnlp-1.22 Layered bias: Interpreting bias in pretrained large language models . In Proceedings of the 6th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, pages 284--2...

    2023 doi

  22. [30]

    Aishik Rakshit, Smriti Singh, Shuvam Keshari, Arijit Ghosh Chowdhury, Vinija Jain, and Aman Chadha. 2025. https://aclanthology.org/2025.coling-main.450/ From prejudice to parity: A new approach to debiasing large language model word embeddings . In Proceedings of the 31st Inte...

    2025

  23. [31]

    Akash Saravanan, Dhruv Mullick, Habibur Rahman, and Nidhi Hegde. 2023. Finedeb: A debiasing framework for language models. arXiv preprint arXiv:2302.02453

    2023 arXiv

  24. [32]

    Hari Shrawgi, Prasanjit Rath, Tushar Singhal, and Sandipan Dandapat. 2024. Uncovering stereotypes in large language models: A task complexity-based approach. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume ...

    2024

  25. [33]

    Zara Siddique, Liam Turner, and Luis Espinosa-Anke. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1035 Who is better at math, jenny or jingzhen? uncovering stereotypes in large language models . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language...

    2024 doi

  26. [34]

    Irene Solaiman and Christy Dennison. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/2e855f9489df0712b4bd8ea9e2848c5a-Paper.pdf Process for adapting language models to society (palms) with values-targeted datasets . In Advances in Neural Information Processing...

    2021

  27. [35]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L \'e onard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram \'e , and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:...

    2024 arXiv

  28. [36]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  29. [37]

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

    2020 arXiv

  30. [38]

    Lalor, and Kar Yan Tam

    Yi Yang, Hanyu Duan, Ahmed Abbasi, John P. Lalor, and Kar Yan Tam. 2025. https://doi.org/10.18653/v1/2025.trustnlp-main.18 Bias a-head? analyzing bias in transformer-based language model attention heads . In Proceedings of the 5th Workshop on Trustworthy NLP (TrustNLP 2025), p...

    2025 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.