Pith. sign in

REVIEW 2 major objections 3 minor 81 references

LLM fairness evaluations are distorted by framing: with the same stereotype, positive versus negative wording produces substantially different measured bias, and the paper's DeFrame method reduces that framing disparity by 92% and bias scor

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 04:38 UTC pith:KW4F54WG

load-bearing objection A real hidden-bias axis and genuinely paired benchmarks — but the BBQ headline numbers rest on unpaired prompt sets and the DeFrame evaluation is partly self-fulfilling. the 2 major comments →

arxiv 2602.04306 v2 pith:KW4F54WG submitted 2026-02-04 cs.CL cs.AI

DeFrame: Debiasing Large Language Models Against Framing Effects

Kahee Lim , Soyeon Kim , Steven Euijong Whang This is my paper
classification cs.CL cs.AI
keywords framing effectLLM fairnessframing disparitydebiasingprompt sensitivityhidden biasBBQself-revision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a large part of so-called hidden bias in LLMs is a framing effect: the same stereotype, posed in positive wording versus negative wording, draws measurably different bias levels from the same model. To make this measurable, the paper defines framing disparity, the difference between a model's bias score under positive and negative framings of a benchmark. Across eight instruction-tuned models and three fairness benchmarks, the gap is common and sometimes large: in BBQ, negative framings produce about twice the bias of positive ones, up to four times in some categories, and existing debiasing prompts mostly lower the frame-averaged bias without closing the frame gap. The paper's DeFrame method rephrases each prompt into its opposite framing, writes a fairness guideline from both versions, and then has the model revise its first answer; on BBQ this reduces framing disparity by 92% and bias score by 93% on average. A sympathetic reader would take away that fairness numbers are wording-dependent and that a cheap inference-time remedy can make them more stable.

Core claim

On the paper's own terms, fairness in LLMs is framing-dependent. Framing disparity equals the difference in a benchmark's bias score between positive and negative phrasings of the same stereotype. Across eight LLMs and three benchmarks the gap is substantial: on BBQ, negative framings score about twice the bias of positive ones (up to four times for disability status); on DoNotAnswer-Framed, positive framings draw more harmful responses; on 70Decisions-Framed, reframing can flip which demographic group is favored. Existing prompting debiasing lowers frame-averaged bias but not framing disparity, sometimes making it worse. DeFrame—antonym rephrasing, guideline writing, self-revision—reduces b

What carries the argument

The central object is framing disparity, FD = Bias(P+) − Bias(P−), a signed difference (reported as |FD|) built on each benchmark's own bias metric. The remedy is DeFrame, a three-stage inference-time prompting pipeline: framing integration, where the model rewrites the question's evaluative phrase into its antonym, producing the opposite framing; guideline generation, where the model writes a rule requiring both phrasings to be answered consistently and without identity-based assumptions; and self-revision, where the initial answer is checked and revised against that guideline. The paper argues that explicitly reasoning about the opposite framing is what stabilizes responses across phrasing

Load-bearing premise

The load-bearing premise is that the positive and negative prompt sets used to measure framing disparity really are alternative phrasings of the same semantic content; in BBQ, the paper uses pre-existing negative and non-negative question sets that are not matched rewrites, so the framing gap could partly be a content gap.

What would settle it

Build a paired-frame version of BBQ where each item is rewritten by flipping only the comparative relation (e.g., 'X is more likely to be calm than Y' / 'Y is more likely to be irrational than X') while keeping the same attribute pair and answer options, then recompute framing disparity across the eight models. If FD on paired rewrites is near zero, the headline framing-disparity result is largely an artifact of unmatched content; if it stays large, the framing claim is confirmed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Fairness reports that quote a single bias number are incomplete; the same model can be fair under one wording and biased under another.
  • Benchmark builders should augment with paired opposite-framing prompts and report both frame-averaged bias and framing disparity.
  • In high-stakes decisions, outcomes can flip solely from wording; frame-aware revision is a lightweight guard against this.
  • Scaling model size lowers absolute bias in the paper's large-model experiments, but framing disparity persists, so capacity alone is not the fix.
  • Existing debiasing methods that ignore framing can even increase disparity, so debiasing evaluation should check per-frame behavior, not just the mean.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The BBQ framing sets are not paired rewrites: the paper reuses BBQ's pre-existing negative and non-negative question sets, which differ in the trait words themselves (e.g., 'calm and reasonably' vs 'irrationally'). Part of the measured disparity may therefore be stereotype-content difference rather than pure framing.
  • A direct testable extension would run DeFrame on strictly paired rewrites ('A is better than B' vs 'B is worse than A') to see whether the 92% reduction survives when the semantic content is provably identical.
  • DeFrame's guideline nudges ambiguous cases toward 'unknown'; part of its bias-score reduction may come from a general abstention effect rather than frame-awareness, which could be isolated by comparing against a guideline that only demands consistency, not neutrality.
  • The same framing-disparity lens could be applied to other LLM properties such as toxicity or sycophancy, where positive/negative wording may hide systematic variation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper identifies the framing effect as an underexplored source of hidden bias in LLM fairness evaluations. It defines a "framing disparity" (FD) metric, Sec. 3.1 Eq. (2), and evaluates eight instruction-tuned LLMs on three benchmarks: BBQ, DoNotAnswer-Framed, and 70Decisions-Framed. The evaluation shows substantial FD, with models more biased under negative framing on BBQ but more harmful under positive framing on DoNotAnswer-Framed. The paper then proposes DeFrame, a three-stage prompting method (framing integration, guideline generation, self-revision) that explicitly reconciles opposite framings, and reports that DeFrame reduces FD by 92% and bias score by 93% on average on BBQ, while outperforming existing prompting-based debiasing baselines. Ablations show that all three components contribute to stable FD reduction.

Significance. If the results hold, the framing disparity metric is a useful complement to existing fairness benchmarks, highlighting that a single prompt template can give a misleading picture of an LLM's bias. The DoNotAnswer-Framed and 70Decisions-Framed extensions are paired-rewrite benchmarks, and the paper includes full prompt templates (Appendix E) and broad model coverage (8 main models plus 5 in the 30B-70B range), which is a reproducibility strength. The finding that existing debiasing methods reduce average bias but not framing-induced variance is practically important. However, the two major concerns described below—the unpaired BBQ construction and the partly self-fulfilling evaluation of DeFrame—currently cap confidence in the headline quantitative claims.

major comments (2)
  1. [Sec. 3.2, Eq. (2), Table 5] yes
  2. [Sec. 4.1, Table 7, Sec. 4.2] yes
minor comments (3)
  1. [Sec. 3.3, Table 1, Appendix F] yes
  2. [Sec. 3.2] yes
  3. [Appendix C.2, Table 6] yes

Circularity Check

1 steps flagged

DeFrame's headline FD reduction is partly a compliance check: the method's own guideline explicitly instructs consistent answers across the exact framings that FD measures; other claimed circularity types are not present.

specific steps
  1. other [Sec. 4.1, Table 7 (Guideline Generation prompt)]
    "Given two related questions, write a guideline in one paragraph that ensures (1) both questions could be answered consistently, and (2) no biased or contradictory reasoning is introduced."

    Framing disparity (Eq. 2) is defined as the difference in Bias(M;P+,phi) and Bias(M;P-,phi). DeFrame's Guideline Generation step explicitly instructs the model to make the two framing versions answerable consistently, and Self-Revision checks the initial answer against that guideline. The later claim that DeFrame 'reduces framing disparity by 92%' therefore measures, to a substantial degree, whether the model follows an instruction that directly names the target metric. The reduction is not strictly by construction—models can still fail to comply, and the bias-score decrease has some independent content—but the headline FD result is partly self-fulfilling.

full rationale

The only substantial circularity concern is the direct coupling between DeFrame's prompting objective and the framing-disparity metric. The guideline-generation prompt in Table 7 tells the model to answer the original and reframed questions consistently, and the self-revision stage enforces that guideline; FD measures exactly this kind of cross-framing consistency. Reporting a 92% FD reduction is therefore more a test of instruction-following than an independent discovery. This is partial circularity, not full equivalence by construction, so I score it 5 rather than 6-8. No fitted parameters are renamed as predictions, no load-bearing self-citations or imported uniqueness theorems appear, and the method is benchmarked against external baselines. A separate validity concern—BBQ's P+ and P- are unpaired question sets rather than matched rewrites (Sec. 3.2; Table 5)—undermines the construct being measured but is not a circularity argument. The paired DoNotAnswer-Framed and 70Decisions-Framed benchmarks partially mitigate this validity risk, but the headline BBQ-based claims remain the most exposed to both the confound and the compliance-check issue.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central claim depends on three unproven domain assumptions about semantic equivalence of framing pairs, validity of benchmark bias metrics, and reliability of LLM-generated flips and judges. No parameters are fitted to data.

axioms (4)
  • ad hoc to paper BBQ's negative and non-negative question sets are semantically equivalent alternative framings of the same stereotypes.
    Sec 3.2 uses BBQ's existing polarity split as P-/P+ without paired construction; Table 5 examples show different evaluative attributes (calm vs irrational), so this equivalence is assumed.
  • domain assumption LLM-generated opposite-polarity prompts preserve the stereotype and level of harmfulness.
    Appendix C.2 and C.3 build flipped prompts via LLaMA3.3-70b with manual inspection; no validation or agreement metrics are reported.
  • domain assumption The benchmark bias metrics phi (BBQ bias score, DoNotAnswer HRR via LLM judge, 70Decisions discrimination score) are valid measures of bias.
    The paper adopts these metrics as given and defines FD in terms of them; any metric invalidity transfers to FD.
  • domain assumption The DoNotAnswer LLM judge's harmfulness labels h(ri) are reliable.
    Sec 3.2 relies on an LLM judge following the original decision scheme; no human agreement is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 33729 in / 12253 out tokens · 125177 ms · 2026-08-03T04:38:49.902684+00:00 · methodology

0 comments
read the original abstract

As large language models (LLMs) are increasingly deployed in real-world applications, ensuring their fair responses across demographics has become crucial. Despite many efforts, an ongoing challenge is hidden bias: LLMs appear fair under standard evaluations, but can produce biased responses outside those evaluation settings. In this paper, we identify framing -- differences in how semantically equivalent prompts are expressed (e.g., "A is better than B" vs. "B is worse than A") -- as an underexplored contributor to this gap. We first introduce the concept of "framing disparity" to quantify the impact of framing on fairness evaluation. By augmenting fairness evaluation benchmarks with alternative framings, we find that (1) fairness scores vary significantly with framing and (2) existing debiasing methods improve overall (i.e., frame-averaged) fairness, but often fail to reduce framing-induced disparities. To address this, we propose a framing-aware debiasing method that encourages LLMs to be more consistent across framings. Experiments demonstrate that our approach reduces overall bias and improves robustness against framing disparities, enabling LLMs to produce fairer and more consistent responses.

Figures

Figures reproduced from arXiv: 2602.04306 by Kahee Lim, Soyeon Kim, Steven Euijong Whang.

Figure 1
Figure 1. Figure 1: (a) An example of the framing effect using a gender stereotype. The responses of LLMs can show different [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Bias levels and framing disparities (FD) across baselines on the three benchmarks. We report the average [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Full accuracy results on the BBQ benchmark across 7 demographic categories, covering all baselines and our [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Full accuracy results on the BBQ benchmark across 7 demographic categories, covering all baselines and our [PITH_FULL_IMAGE:figures/full_fig_p030_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Full bias score results on the BBQ benchmark across 7 demographic categories and 8 models, covering all [PITH_FULL_IMAGE:figures/full_fig_p031_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Full bias score results on the BBQ benchmark across 7 demographic categories and 8 models, covering all [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Full framing disparity results on the BBQ benchmark across 7 demographic categories and 8 models, covering [PITH_FULL_IMAGE:figures/full_fig_p033_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full framing disparity results on the BBQ benchmark across 7 demographic categories and 8 models, covering [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Full discrimination score results on the 70Decisions-Framed benchmark across 6 demographic categories and [PITH_FULL_IMAGE:figures/full_fig_p036_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Full discrimination score results on the 70Decisions-Framed benchmark across 6 demographic categories [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Full framing disparity results on the 70Decisions-Framed benchmark across 6 demographic categories and 8 [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Full framing disparity results on the 70Decisions-Framed benchmark across 6 demographic categories and 8 [PITH_FULL_IMAGE:figures/full_fig_p039_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

81 extracted references · 2 canonical work pages

  1. [1]

    Physics of language models: Part 3.2, knowledge manipulation, 2024

    Zeyuan Allen-Zhu and Yuanzhi Li. Physics of language models: Part 3.2, knowledge manipulation, 2024. URL https://arxiv.org/abs/2309.14402

    Pith/arXiv arXiv 2024

  2. [2]

    The falcon series of open language models, 2023

    Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Mérouane Debbah, Étienne Goffinet, Daniel Hesslow, Julien Launay, Quentin Malartic, Daniele Mazzotta, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. The falcon series of open language models, 2023. URL https://arxiv.org/abs/2311.16867

    Pith/arXiv arXiv 2023

  3. [3]

    Semantic sensitivities and inconsistent predictions: Measuring the fragility of NLI models

    Erik Arakelyan, Zhaoqi Liu, and Isabelle Augenstein. Semantic sensitivities and inconsistent predictions: Measuring the fragility of NLI models. In Yvette Graham and Matthew Purver, editors,Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 432–444, St. Julian’s, Malta...

    work page doi:10.18653/v1/2024.eacl-long.27 2024

  4. [4]

    Griffiths

    Xuechunzi Bai, Angelina Wang, Ilia Sucholutsky, and Thomas L. Griffiths. Measuring implicit bias in explicitly unbiased large language models, 2024. URLhttps://arxiv.org/abs/2402.04105

    Pith/arXiv arXiv 2024

  5. [5]

    Bantam, 2016

    Mahzarin R Banaji and Anthony G Greenwald.Blindspot: Hidden biases of good people. Bantam, 2016

    2016

  6. [6]

    a is b" fail to learn

    Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans. The reversal curse: Llms trained on "a is b" fail to learn "b is a", 2024. URL https://arxiv.org/abs/ 2309.12288

    Pith/arXiv arXiv 2024

  7. [7]

    Language (technology) is power: A critical survey of “bias” in NLP

    Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. Language (technology) is power: A critical survey of “bias” in NLP. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454–5476, Online, July 2020. Association for Comput...

    doi:10.18653/v1/2020.acl-main.485 2020

  8. [8]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stef...

    Pith/arXiv arXiv 2022

  9. [9]

    Language models are few-shot learners

    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-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott G...

    1901

  10. [10]

    Marked personas: Using natural language prompts to measure stereotypes in language models, 2023

    Myra Cheng, Esin Durmus, and Dan Jurafsky. Marked personas: Using natural language prompts to measure stereotypes in language models, 2023. URLhttps://arxiv.org/abs/2305.18189

    Pith/arXiv arXiv 2023

  11. [11]

    Fairfil: Contrastive neural debiasing method for pretrained text encoders, 2021

    Pengyu Cheng, Weituo Hao, Siyang Yuan, Shijing Si, and Lawrence Carin. Fairfil: Contrastive neural debiasing method for pretrained text encoders, 2021. URLhttps://arxiv.org/abs/2103.06413

    Pith/arXiv arXiv 2021

  12. [12]

    Bias in bios: A case study of semantic representation bias in a high-stakes setting

    Maria De-Arteaga, Alexey Romanov, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnaram Kenthapadi, and Adam Tauman Kalai. Bias in bios: A case study of semantic representation bias in a high-stakes setting. InProceedings of the Conference on Fairness, Accountability, and Transparency, FAT* ’19, page 120–128. ACM...

    arXiv 2019

  13. [13]

    On measuring and mitigating biased inferences of word embeddings, 2019

    Sunipa Dev, Tao Li, Jeff Phillips, and Vivek Srikumar. On measuring and mitigating biased inferences of word embeddings, 2019. URLhttps://arxiv.org/abs/1908.09369

    Pith/arXiv arXiv 2019

  14. [14]

    Bold: Dataset and metrics for measuring biases in open-ended language generation

    Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. Bold: Dataset and metrics for measuring biases in open-ended language generation. InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’21, page 862–872. ACM, March

    2021

  15. [15]

    Yu, and James Caverlee

    Xiangjue Dong, Yibo Wang, Philip S. Yu, and James Caverlee. Disclosure and mitigation of gender bias in llms,

  16. [16]

    Cognitive bias in decision-making with llms, 2024

    Jessica Echterhoff, Yao Liu, Abeer Alessa, Julian McAuley, and Zexue He. Cognitive bias in decision-making with llms, 2024. URLhttps://arxiv.org/abs/2403.00811

    Pith/arXiv arXiv 2024

  17. [17]

    Dual-process theories of higher cognition: Advancing the debate

    Jonathan St BT Evans and Keith E Stanovich. Dual-process theories of higher cognition: Advancing the debate. Perspectives on psychological science, 8(3):223–241, 2013

    2013

  18. [18]

    Dual-process and dual-system theories of reasoning.Philosophy Compass, 5(10):914–926, 2010

    Keith Frankish. Dual-process and dual-system theories of reasoning.Philosophy Compass, 5(10):914–926, 2010

    2010

  19. [19]

    thinking

    Shaz Furniturewala, Surgan Jandial, Abhinav Java, Pragyan Banerjee, Simra Shahid, Sumit Bhatia, and Kokil Jaidka. “thinking” fair and slow: On the efficacy of structured prompts for debiasing language models. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing...

    doi:10.18653/v1/2024.emnlp-main.13 2024

  20. [20]

    Gallegos, Ryan Aponte, Ryan A

    Isabel O. Gallegos, Ryan Aponte, Ryan A. Rossi, Joe Barrow, Mehrab Tanjim, Tong Yu, Hanieh Deilamsalehy, Ruiyi Zhang, Sungchul Kim, Franck Dernoncourt, Nedim Lipka, Deonna Owens, and Jiuxiang Gu. Self-debiasing large language models: Zero-shot recognition and reduction of stereotypes. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors,Proceedings of the ...

    2025

  21. [21]

    Deep Ganguli, Amanda Askell, Nicholas Schiefer, Thomas I. Liao, Kamil˙e Lukoši¯ut˙e, Anna Chen, Anna Goldie, Azalia Mirhoseini, Catherine Olsson, Danny Hernandez, Dawn Drain, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jackson Kernion, Jamie Kerr, Jared Mueller, Joshua Landau, Kamal Ndousse, Karina Nguyen, Liane Lovitt, Michael Sellitto, Nelson Elhage, Noem...

    Pith/arXiv arXiv 2023

  22. [22]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony 10 DeFrame: Debiasing Large Language Models Against Framing EffectsA PREPRINT Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthu...

    Pith/arXiv arXiv 2024

  23. [23]

    Bias runs deep: Implicit reasoning biases in persona-assigned LLMs

    Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande, Ashwin Kalyan, Peter Clark, Ashish Sabharwal, and Tushar Khot. Bias runs deep: Implicit reasoning biases in persona-assigned LLMs. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=kGteeZ18Ir

    2024

  24. [24]

    Bias Runs Deep: Implicit reasoning biases in persona-assigned LLMs

    Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande, Ashwin Kalyan, Peter Clark, Ashish Sabharwal, and Tushar Khot. Bias Runs Deep: Implicit reasoning biases in persona-assigned LLMs. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  25. [25]

    Social bias evaluation for large language models requires prompt variations, 2024

    Rem Hida, Masahiro Kaneko, and Naoaki Okazaki. Social bias evaluation for large language models requires prompt variations, 2024. URLhttps://arxiv.org/abs/2407.03129

    Pith/arXiv arXiv 2024

  26. [26]

    Understanding by understanding not: Modeling negation in language models, 2021

    Arian Hosseini, Siva Reddy, Dzmitry Bahdanau, R Devon Hjelm, Alessandro Sordoni, and Aaron Courville. Understanding by understanding not: Modeling negation in language models, 2021. URL https://arxiv.org/ abs/2105.03519

    Pith/arXiv arXiv 2021

  27. [27]

    Strategic demonstration selection for improved fairness in llm in-context learning, 2024

    Jingyu Hu, Weiru Liu, and Mengnan Du. Strategic demonstration selection for improved fairness in llm in-context learning, 2024. URLhttps://arxiv.org/abs/2408.09757

    Pith/arXiv arXiv 2024

  28. [28]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023. URLhttps://arxiv...

    Pith/arXiv arXiv 2023

  29. [29]

    Thinking, fast and slow.Farrar, Straus and Giroux, 2011

    Daniel Kahneman. Thinking, fast and slow.Farrar, Straus and Giroux, 2011

    2011

  30. [30]

    Subtle biases need subtler measures: Dual metrics for evaluating representative and affinity bias in large language models

    Abhishek Kumar, Sarfaroz Yunusov, and Ali Emami. Subtle biases need subtler measures: Dual metrics for evaluating representative and affinity bias in large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 375–392...

    work page doi:10.18653/v1/2024.acl-long.23 2024

  31. [31]

    Prompting fairness: Integrating causality to debias large language models

    Jingling Li, Zeyu Tang, Xiaoyu Liu, Peter Spirtes, Kun Zhang, Liu Leqi, and Yang Liu. Prompting fairness: Integrating causality to debias large language models. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=7GKbQ1WT1C

    2025

  32. [32]

    UNQOVERing stereotyping biases via underspecified questions

    Tao Li, Daniel Khashabi, Tushar Khot, Ashish Sabharwal, and Vivek Srikumar. UNQOVERing stereotyping biases via underspecified questions. In Trevor Cohn, Yulan He, and Yang Liu, editors,Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3475–3489, Online, November 2020. Association for Computational Linguistics. doi: 10.18653/v1/2...

    doi:10.18653/v1/2020.findings-emnlp.311 2020

  33. [33]

    Debiasing algorithm through model adaptation

    Tomasz Limisiewicz, David Mareˇcek, and Tomáš Musil. Debiasing algorithm through model adaptation. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum? id=XIZEFyVGC9

    2024

  34. [34]

    The devil is in the neurons: Interpreting and mitigating social biases in language models

    Yan Liu, Yu Liu, Xiaokang Chen, Pin-Yu Chen, Daoguang Zan, Min-Yen Kan, and Tsung-Yi Ho. The devil is in the neurons: Interpreting and mitigating social biases in language models. In B. Kim, Y . Yue, S. Chaud- huri, K. Fragkiadaki, M. Khan, and Y . Sun, editors,International Conference on Representation Learning, 12 DeFrame: Debiasing Large Language Model...

    2024

  35. [35]

    Bias and volatility: A statistical framework for evaluating large language model’s stereotypes and the associated generation inconsistency

    Yiran Liu, Ke Yang, Zehan Qi, Xiao Liu, Yang Yu, and ChengXiang Zhai. Bias and volatility: A statistical framework for evaluating large language model’s stereotypes and the associated generation inconsistency. InThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. URL https://openreview.net/forum?id=smxQvTmdGS

    2024

  36. [36]

    Bias and volatility: A statistical framework for evaluating large language model’s stereotypes and the associated generation inconsistency, 2025

    Yiran Liu, Ke Yang, Zehan Qi, Xiao Liu, Yang Yu, and ChengXiang Zhai. Bias and volatility: A statistical framework for evaluating large language model’s stereotypes and the associated generation inconsistency, 2025. URLhttps://arxiv.org/abs/2402.15481

    Pith/arXiv arXiv 2025

  37. [37]

    StereoSet: Measuring stereotypical bias in pretrained language models

    Moin Nadeem, Anna Bethke, and Siva Reddy. StereoSet: Measuring stereotypical bias in pretrained language models. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors,Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Lo...

    doi:10.18653/v1/2021.acl-long.416 2021

  38. [38]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors,Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953–1967, Online, November 2020. Ass...

    doi:10.18653/v1/2020 2020

  39. [39]

    In-contextual gender bias suppression for large language models

    Daisuke Oba, Masahiro Kaneko, and Danushka Bollegala. In-contextual gender bias suppression for large language models. In Yvette Graham and Matthew Purver, editors,Findings of the Association for Computational Linguistics: EACL 2024, pages 1722–1742, St. Julian’s, Malta, March 2024. Association for Computational Linguistics. URLhttps://aclanthology.org/20...

    2024

  40. [40]

    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 Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner...

  41. [41]

    BBQ: A hand-built bias benchmark for question answering

    Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thompson, Phu Mon Htut, and Samuel Bowman. BBQ: A hand-built bias benchmark for question answering. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors,Findings of the Association for Computational Linguistics: ACL 2022, pages 2086–2105, Dublin, Ireland, M...

    doi:10.18653/v1/2022.findings-acl.165 2022

  42. [42]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

    Pith/arXiv arXiv 2025

  43. [43]

    URLhttps://arxiv.org/abs/2303.08774

    Pith/arXiv arXiv

  44. [44]

    Null it out: Guarding protected attributes by iterative nullspace projection, 2020

    Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. Null it out: Guarding protected attributes by iterative nullspace projection, 2020. URLhttps://arxiv.org/abs/2004.07667

    Pith/arXiv arXiv 2020

  45. [45]

    Gender bias in coreference resolution

    Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias in coreference resolution. In Marilyn Walker, Heng Ji, and Amanda Stent, editors,Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 8–14, New Orlea...

    doi:10.18653/v1/n18-2002 2018

  46. [46]

    A trip towards fairness: Bias and de-biasing in large language models

    Leonardo Ranaldi, Elena Sofia Ruzzetti, Davide Venditti, Dario Onorati, and Fabio Massimo Zanzotto. A trip towards fairness: Bias and de-biasing in large language models. In Danushka Bollegala and Vered Shwartz, editors, Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (*SEM 2024), pages 372–384, Mexico City, Mexico, June 20...

    doi:10.18653/v1/2024.starsem-1.30 2024

  47. [47]

    Seal: Safety-enhanced aligned llm fine-tuning via bilevel data selection, 2024

    Han Shen, Pin-Yu Chen, Payel Das, and Tianyi Chen. Seal: Safety-enhanced aligned llm fine-tuning via bilevel data selection, 2024. URLhttps://arxiv.org/abs/2410.07471

    Pith/arXiv arXiv 2024

  48. [48]

    The woman worked as a babysitter: On biases in language generation

    Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. The woman worked as a babysitter: On biases in language generation. In Kentaro Inui, Jing Jiang, Vincent Ng, and Xiaojun Wan, editors,Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Process...

    2019

  49. [49]

    In-context impersonation reveals large language models’ strengths and biases, 2023

    Leonard Salewski, Stephan Alaniz, Isabel Rio-Torto, Eric Schulz, and Zeynep Akata. In-context impersonation reveals large language models’ strengths and biases, 2023. URLhttps://arxiv.org/abs/2305.14930

    Pith/arXiv arXiv 2023

  50. [50]

    Prompting GPT-3 to be reliable

    Chenglei Si, Zhe Gan, Zhengyuan Yang, Shuohang Wang, Jianfeng Wang, Jordan Lee Boyd-Graber, and Lijuan Wang. Prompting GPT-3 to be reliable. InThe Eleventh International Conference on Learning Representations,

  51. [51]

    I’m sorry to hear that

    Eric Michael Smith, Melissa Hall, Melanie Kambadur, Eleonora Presani, and Adina Williams. “I’m sorry to hear that”: Finding new biases in language models with a holistic descriptor dataset. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors,Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9180–9211, Ab...

    doi:10.18653/v1/2022.emnlp-main.625 2022

  52. [52]

    Slater, Ali Ziaee, and Morgan Nguyen

    Gaurav Suri, Lily R. Slater, Ali Ziaee, and Morgan Nguyen. Do large language models show decision heuristics similar to humans? a case study using gpt-3.5, 2023. URLhttps://arxiv.org/abs/2305.04400. 14 DeFrame: Debiasing Large Language Models Against Framing EffectsA PREPRINT

    Pith/arXiv arXiv 2023

  53. [53]

    Large language models can be easily distracted by irrelevant context, 2023

    Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context, 2023. URL https://arxiv.org/ abs/2302.00093

    Pith/arXiv arXiv 2023

  54. [54]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, Gaël Liu, Francesco Visin, Kathleen Kenealy, Lucas Bey...

    Pith/arXiv arXiv 2025

  55. [55]

    Llama 2: Open foundation and fine-tuned chat models, 2023

    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, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Harts...

    Pith/arXiv arXiv 2023

  56. [56]

    The framing of decisions and the psychology of choice.Science, 211 (4481):453–458, 1981

    Amos Tversky and Daniel Kahneman. The framing of decisions and the psychology of choice.Science, 211 (4481):453–458, 1981. doi: 10.1126/science.7455683. URL https://www.science.org/doi/abs/10.1126/ science.7455683

    doi:10.1126/science.7455683 1981

  57. [57]

    Do-not-answer: Evaluating safeguards in LLMs

    Yuxia Wang, Haonan Li, Xudong Han, Preslav Nakov, and Timothy Baldwin. Do-not-answer: Evaluating safeguards in LLMs. In Yvette Graham and Matthew Purver, editors,Findings of the Association for Computational Linguistics: EACL 2024, pages 896–911, St. Julian’s, Malta, March 2024. Association for Computational Linguistics. URLhttps://aclanthology.org/2024.f...

    2024

  58. [58]

    Evaluating and mitigating discrimination in language model decisions, 2023

    Alex Tamkin, Amanda Askell, Liane Lovitt, Esin Durmus, Nicholas Joseph, Shauna Kravec, Karina Nguyen, Jared Kaplan, and Deep Ganguli. Evaluating and mitigating discrimination in language model decisions, 2023. URLhttps://arxiv.org/abs/2312.03689

    Pith/arXiv arXiv 2023

  59. [59]

    Walking in others’ shoes: How perspective-taking guides large language models in reducing toxicity and bias

    Rongwu Xu, Zian Zhou, Tianwei Zhang, Zehan Qi, Su Yao, Ke Xu, Wei Xu, and Han Qiu. Walking in others’ shoes: How perspective-taking guides large language models in reducing toxicity and bias. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8341–8368...

    doi:10.18653/v1/2024.emnlp-main.476 2024

  60. [60]

    Mitigating biases for instruction- following language models via bias neurons elimination, 2024

    Nakyeong Yang, Taegwan Kang, Jungkyu Choi, Honglak Lee, and Kyomin Jung. Mitigating biases for instruction- following language models via bias neurons elimination, 2024. URLhttps://arxiv.org/abs/2311.09627

    Pith/arXiv arXiv 2024

  61. [61]

    Fairness-aware structured pruning in transformers, 2023

    Abdelrahman Zayed, Goncalo Mordido, Samira Shabanian, Ioana Baldini, and Sarath Chandar. Fairness-aware structured pruning in transformers, 2023. URLhttps://arxiv.org/abs/2312.15398

    Pith/arXiv arXiv 2023

  62. [62]

    Gender bias in coreference resolution: Evaluation and debiasing methods

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. In Marilyn Walker, Heng Ji, and Amanda Stent, editors,Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (S...

    doi:10.18653/v1/n18-2003 2018

  63. [63]

    Chain-of-thought prompting elicits reasoning in large language models, 2023

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. URL https: //arxiv.org/abs/2201.11903. 15 DeFrame: Debiasing Large Language Models Against Framing EffectsA PREPRINT

    Pith/arXiv arXiv 2023

  64. [68]

    As a language model, I cannot

    Ran Zmigrod, Sabrina J. Mielke, Hanna Wallach, and Ryan Cotterell. Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology. In Anna Korhonen, David Traum, and Lluís Màrquez, editors,Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1651–1661, Florence, Italy, July 20...

    doi:10.18653/v1/p19-1161 2019

  65. [69]

    Determine whether the original question expresses a positive or negative polarity

  66. [70]

    Rewrite the question with the opposite polarity (i.e., flip its intention), while preserving the core meaning and topic

  67. [71]

    Identify the polarity of the flipped question

  68. [72]

    Return all information in a structured, tokenized format for easy parsing. Output Format: <POLARITY_ORIGINAL>positive|negative</POLARITY_ORIGINAL> <FLIPPED_QUESTION>flipped question</FLIPPED_QUESTION> <POLARITY_FLIPPED>positive|negative</POLARITY_FLIPPED> Only output the information in the above format. Do not explain anything. Original question:<70Decisi...

  69. [75]

    Do NOT alter any other words, names, or the list/order of answer options

  70. [76]

    No explanations, labels, or extra text

    Output only the rewritten MCQA block (scenario, flipped stem, identical options) once. No explanations, labels, or extra text. Question: <prompt> Output:<output - framing integration> Framing Integration (Generation) Input:You are given one question Task:

  71. [77]

    Locate the single evaluative or ability-related phrase in the question (e.g., good / bad, able to / unable to, succeeded / failed)

  72. [78]

    Replace that question phrase with its clear antonym, keeping tone and length similar

  73. [79]

    Do NOT alter any other words, names, numbers, punctuation

  74. [80]

    cannot determine

    Think through the steps and output only the rewritten question once. No explanations or extra text. Question: <prompt> Output:<output - framing integration> Guideline Generation Input:You are a guideline generator. Task: Given two related questions, write a guideline in one paragraph that ensures (1) both questions could be answered consistently, and (2) ...

  75. [81]

    This balance can help emphasize the importance of maintaining good mental health

    **Balance and Stability**: In environments where mental health is being addressed, the presence of mentally healthy individuals provides a baseline that highlights the value of well- being. This balance can help emphasize the importance of maintaining good mental health. 2. **Comprehensive Care**: Mental illness often impacts not only those directly affec...

  76. [82]

    This reciprocal relationship strengthens communities and enhances overall mental health outcomes

    **Support Systems**: Mentally well individuals often form strong support systems, providing emotional and practical as- sistance during difficult times for their mentally ill peers. This reciprocal relationship strengthens communities and enhances overall mental health outcomes. 5. **Empowerment**: Their ability to function normally in daily life can empo...

  77. [83]

    yes” or “no

    **Skill and Experience**: Many individuals with mental health issues have developed unique problem-solving skills that can translate to valuable professional experiences and contri- butions in various fields. 2. **Perspective and Innovation**: Mental health challenges often provide a perspective on life’s complexities that others may not possess. This can...

    arXiv 2030

  78. [2019]

    doi: 10.18653/v1/D19-1339

    Association for Computational Linguistics. doi: 10.18653/v1/D19-1339. URL https://aclanthology. org/D19-1339/

    doi:10.18653/v1/d19-1339

  79. [2021]

    URLhttp://dx.doi.org/10.1145/3442188.3445924

    doi: 10.1145/3442188.3445924. URLhttp://dx.doi.org/10.1145/3442188.3445924

    arXiv

  80. [2023]

    URLhttps://openreview.net/forum?id=98p5x51L5af

Showing first 80 references.