Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

This paper argues that CLIP-style vision-language encoders, tested on four architectures, systematically place face embeddings closer to occupation and activity statements along gender lines, with observed bias exceeding chance-level nulls

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 ↗

Using cosine similarity in CLIP embedding space, the paper finds that male and female face sets are differentially associated with occupation and activity statements across all four tested models.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A clean measurement idea sunk by sloppy numbers and uncontrolled face galleries; the broad CLIP-bias claim is already known. the 4 major comments →

arxiv 2508.11262 v1 pith:ES6XL5TP submitted 2025-08-15 cs.CV cs.AI

Vision-Language Models display a strong gender bias

Aiswarya Konavoor , Raj Abhijit Dandekar , Rajat Dandekar , Sreedath Panat This is my paper
classification cs.CV cs.AI
keywords gender biasvision-language modelsCLIPembedding geometrycosine similaritystereotype measurementrepresentation biaszero-shot alignment
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.

The reading

The paper tries to establish that the geometry of contrastive vision-language encoders carries a consistent gender-linked association between faces and statements about work and daily activity. Using 220 face photographs split by perceived binary gender and a set of occupation and activity statements, it defines each statement's bias as the difference between its mean cosine similarity to male faces and its mean cosine similarity to female faces, then compares the observed absolute bias to a label-swap null distribution. Across all four CLIP-style encoders tested, the observed bias exceeds the null by factors of 1.84 to 2.00, and the statement-level directions track real-world occupational stereotypes, with statements like 'firefighter' and 'CEO' landing near male faces and 'nurse' and 'teacher' near female faces. A sympathetic reader would care because the probe isolates the encoder itself: no decoding and no downstream task, just the shared embedding space that retrieval and zero-shot systems are built on.

Core claim

The paper's central discovery claim is that gender-linked associations between faces and occupation or activity statements are present in all four contrastive dual encoders studied — ViT-B/32, ViT-L/14, RN50, and RN101 — and that the effect is well above what a label-swap null would produce. The headline evidence is the observed-to-null ratio in Table 3: 2.00 for ViT-B/32, 1.95 for ViT-L/14, 1.85 for RN50, and 1.84 for RN101. Category-level aggregation shows directional splits that are consistent across models: domestic labour (+1.18), professional roles (+0.84), and physical labour (+0.30) lean male, while technical labour (-0.90), cognitive labour (-0.41), and emotional labour (-0.18) lean

What carries the argument

The load-bearing object is the statement-level association score $Bias(s) = \frac{1}{N_m}\sum_i t_s^\top v_i^m - \frac{1}{N_f}\sum_j t_s^\top v_j^f$, the difference between a statement embedding's mean cosine similarity to a male face gallery and to a female face gallery, equivalently a dot product because all embeddings are L2-normalized. Around it sit two calibration devices: a bootstrap over images (and over statements at category level) that attaches 95% confidence intervals, and a label-swap null that randomly re-partitions the pooled face embeddings to estimate how much mean absolute bias would appear with no gender structure. The observed-to-null ratio is what the paper reports as the

Load-bearing premise

The load-bearing premise is that the male and female face galleries differ only in perceived gender; the 220 photographs are balanced in number and cropped to faces, but they are not matched on age, ethnicity, expression, background, or other visual attributes, so any systematic visual difference between the two groups would be scored as gender bias.

What would settle it

Take the same statements and the same encoders, but replace the face galleries with synthetic faces generated in matched pairs that differ only in perceived gender (same age, ethnicity, expression, lighting, background), then recompute the observed-to-null ratios; if the ratios fall from roughly 1.84-2.00 to near 1.0, the original signal was carried by non-gender visual attributes rather than by gender association. A cheaper version: heavily blur or mask the facial region and re-embed — if substantial bias remains without facial information, the score is not measuring the gender of the face.

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

If this is right

  • If the bias lives in the embedding geometry itself, every retrieval and zero-shot classifier built on these encoders inherits it: a face-to-text retrieval system will rank stereotype-coded statements higher for a face of the corresponding perceived gender.
  • The observed-to-null ratios (2.00, 1.95, 1.85, 1.84) give a concrete calibration baseline: audits of CLIP-style models that land near 1.0 are near chance, while ratios near 2.0 replicate this paper's finding.
  • The reported category directions (domestic, professional, and physical labour male-leaning; technical, cognitive, and emotional labour female-leaning) predict where stereotypes concentrate for these model families.
  • The protocol — embeddings only, cosine similarity, bootstrap confidence intervals, label-swap null — is reusable as a cheap, deterministic audit for any new dual encoder before deployment.

Where Pith is reading between the lines

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

  • If the claim is right, the bias is not an artifact of decoding or prompting but is stored in the joint image-text geometry, so it would be shared by every downstream head and mitigation would have to act on the embeddings themselves.
  • A natural extension the paper leaves implicit: swap the face galleries for synthetic faces matched pairwise on age, ethnicity, expression, lighting, and background, differing only in perceived gender; if the observed-to-null ratios collapse toward 1.0, the measured association is partly carried by non-gender visual attributes rather than by gender per se.
  • The category-level sign for technical labour (female-leaning, -0.90) sits in tension with the statement-level examples (engineer, mechanic, electrician listed as male-associated); untangling whether that inversion is driven by particular statements or by the aggregation is a concrete next check.
  • The same protocol could be applied to attribute dimensions orthogonal to gender — age, perceived race, attire — whenever a balanced and ethically sourced probe gallery is available, turning the framework into a general geometric audit of demographic associations.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 6 minor

Summary. The paper proposes a lightweight audit for gender bias in CLIP-style vision-language encoders. For each of 220 face images (split by perceived binary gender) and a set of occupation/activity statements, the authors compute statement-level association scores as the difference between mean cosine similarity to male and female face embeddings (Eq. 1). They add bootstrap confidence intervals and a label-swap permutation null, and report observed-to-null ratios above 1.8 for all four tested encoders (ViT-B/32, ViT-L/14, RN50, RN101). The paper concludes that these encoders exhibit consistent, stereotypically patterned gender associations.

Significance. If the result holds, the proposed method is a simple and transparent way to probe gender associations in contrastive multimodal encoders. Its strengths are that the bias score is a direct mean difference with no fitted parameters, the null model is a clean label-swap permutation, and the paper explicitly attempts to attach uncertainty to every estimate. The central claim — that all four CLIP-style encoders place gender-coded face embeddings closer to stereotype-coded statements — is plausible and consistent with prior work. However, the manuscript as written contains load-bearing reporting errors (an inconsistent statement count, impossible confidence intervals) and a potentially confounded face-gallery design that prevents the current experiments from establishing the claim. The methodology is promising but the reported results are not yet reliable.

major comments (4)
  1. [Abstract, §2.1, Table 2] The abstract states there are 150 unique statements, but Table 2 lists 120 activities plus 200 occupations, totaling 320 statements. Moreover, the six activity categories in Table 2 (Domestic & Caregiving, Mobility & Transport, etc.) do not match the six labor categories named in the abstract and Section 1 (emotional, cognitive, domestic, technical, professional, physical). This is not a minor typo: it makes the dataset description ambiguous and affects every aggregate result. Please reconcile the counts and clarify the relationship between the activity set and the occupation set.
  2. [Table 4] The reported 95% confidence intervals are internally inconsistent. For example, Domestic labour has Bias=1.180 with CI [0.097, 0.080], whose lower bound exceeds its upper bound; Professional roles has CI [-0.284, 0.110] that excludes the point estimate 0.835; Technical labour has CI [-0.012, 0.080] that excludes -0.898. Since uncertainty quantification is a stated contribution, these impossible intervals invalidate the category-level conclusions as reported. Recompute and verify every interval, and state the resampling unit and number of bootstrap resamples used for each row.
  3. [Eq. (1), §2.1] The two face galleries are balanced only in number; the paper reports no matching or statistical control for age, ethnicity, facial expression, head pose, background, or image quality. CLIP embeddings encode these attributes, so the mean cosine-similarity difference in Eq. (1) can be driven by non-gender visual differences between the two galleries. The label-swap null does not address this issue because it randomizes over the same pooled embeddings and preserves any systematic non-gender differences. The central claim that the measured associations are specifically gender-linked therefore requires additional evidence, e.g., per-attribute breakdowns, stratification on available face attributes, or validation on a second, tightly matched face dataset.
  4. [§2.4, Table 3] The null-calibration comparison in Table 3 is incompletely specified. The paper does not state how many label-swap trials were run, what statistic is averaged (mean absolute Bias over statements? over categories?), or what the dispersion of the null distribution is. Without this information, ratios such as 2.00 cannot be meaningfully assessed. Please define the observed and null statistics precisely, report the null distribution (e.g., mean and 95% interval), and state the number of permutations.
minor comments (6)
  1. [§2.2] The prompt-template expansion is described only with two examples. Please list the full set of templates and report a sensitivity analysis across template sets, since averaged text embeddings can be sensitive to this choice.
  2. [Table 4] It is unclear which model Table 4 refers to. Table 5 suggests the values may be for ViT-B/32, but this is never stated. Add a model column or say explicitly that the table is for one representative model.
  3. [§2.5] The reproducibility section mentions a fixed environment and caching, but no code or data repository is provided. Please include a link or state that materials will be released.
  4. [Table 5] The header "A vg. Bias" contains a typo; should be "Avg. Bias".
  5. [Throughout] Spelling of "labour" vs. "labor" is inconsistent. Also, the citation style is inconsistent (e.g., bracket numbers with varying formatting).
  6. [Figure 5/6] Ensure all figures are legible in print, with labeled axes and clear legends; the current text refers to patterns "shown in Figure 6" without listing the model for each subplot.

Circularity Check

0 steps flagged

No circularity: the bias measurement is a direct cosine-similarity contrast with a permutation null, not a fitted or self-referential construction.

full rationale

The paper's central quantity, Bias(s) in Eq. (1), is defined directly as the difference between the mean cosine similarity of a text embedding to male face embeddings and the mean cosine similarity to female face embeddings. Nothing in that definition is fitted to the outcome it predicts; the observed values in Table 3 are direct measurements from pretrained CLIP encoders. The null model is a label-swap permutation of the same face embeddings, which provides an independent baseline by construction, and the reported ratios compare observed bias to that permutation expectation. There is no parameter fitted to a subset of statements and then used to 'predict' those same statements, no self-definitional relationship between the probe statements and the bias score, and no load-bearing self-citation. The most significant validity concern is the uncontrolled face gallery (Section 2.1), which may differ in age, ethnicity, expression, or background across perceived-gender groups; if so, the bias scores could reflect those visual attributes rather than gender alone. That is a measurement-design threat to internal validity, not circularity: the derivation chain does not reduce to its own inputs, and the conclusion is not forced by definition or by a self-referential fit.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper fits no numerical parameters; the bias score is an arithmetic mean difference. The subjective choices, such as face gallery, statement list, prompt templates, and bootstrap count, are design decisions rather than fitted parameters. The main unstated premises are the validity of the shared embedding space, the absence of visual confounds in the face galleries, and the correctness of the statement categorization.

axioms (5)
  • standard math L2-normalized embeddings make cosine similarity equal to dot product in a shared CLIP space.
    Used in Equation (1) and Section 2.2; this is standard for CLIP-style encoders.
  • domain assumption CLIP-style embeddings of faces and text are semantically comparable in a shared space.
    Section 2.2 assumes the shared embedding space supports meaningful association measurement.
  • domain assumption Face galleries labeled by perceived binary gender isolate gender as the only systematic difference.
    Section 2.1 and Figure 2; no controls are reported for age, ethnicity, expression, or background.
  • domain assumption The statement set is gender-neutral and correctly grouped into six labor categories.
    Section 2.1 and Table 2; no list of all 150 or 320 statements is provided.
  • domain assumption A label-swap null distribution of mean absolute bias is an appropriate significance baseline.
    Section 2.4 uses this null to claim that ratios greater than 1 indicate bias beyond chance.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision-Language Models display a strong gender bias." pith.science (2026). https://pith.science/paper/ES6XL5TP

@misc{pith2026250811262,
  author       = {Pith},
  title        = {Pith review of: Vision-Language Models display a strong gender bias},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ES6XL5TP}},
  note         = {Machine review of arXiv:2508.11262}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-language models (VLM) align images and text in a shared representation space that is useful for retrieval and zero-shot transfer. Yet, this alignment can encode and amplify social stereotypes in subtle ways that are not obvious from standard accuracy metrics. In this study, we test whether the contrastive vision-language encoder exhibits gender-linked associations when it places embeddings of face images near embeddings of short phrases that describe occupations and activities. We assemble a dataset of 220 face photographs split by perceived binary gender and a set of 150 unique statements distributed across six categories covering emotional labor, cognitive labor, domestic labor, technical labor, professional roles, and physical labor. We compute unit-norm image embeddings for every face and unit-norm text embeddings for every statement, then define a statement-level association score as the difference between the mean cosine similarity to the male set and the mean cosine similarity to the female set, where positive values indicate stronger association with the male set and negative values indicate stronger association with the female set. We attach bootstrap confidence intervals by resampling images within each gender group, aggregate by category with a separate bootstrap over statements, and run a label-swap null model that estimates the level of mean absolute association we would expect if no gender structure were present. The outcome is a statement-wise and category-wise map of gender associations in a contrastive vision-language space, accompanied by uncertainty, simple sanity checks, and a robust gender bias evaluation framework.

Figures

Figures reproduced from arXiv: 2508.11262 by Aiswarya Konavoor, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat.

Figure 1
Figure 1. Figure 1: Geometric illustration of association in the shared embedding space. The image vectors for a female face and a male face and the text vectors for ”Cooking” and ”CEO” are shown as unit directions, where a smaller angle implies a larger cosine similarity, and since the ”Cooking” vector is closer to the female image than to the male image we have β < α. While the ”CEO” vector is closer to the male image than … view at source ↗
Figure 2
Figure 2. Figure 2: Example face images used in the study. Representa￾tive subset of the balanced male and female face galleries used for bias measurement. Each image is cropped to focus on the face, and individuals are labeled only by perceived binary gender (Male 1–4, Female 1–4) for clarity. These samples illustrate the diversity of appearance within each group while excluding contextual ele￾ments that could confound simil… view at source ↗
Figure 3
Figure 3. Figure 3: Methodology overview. We create a taxonomy of occu￾pations and activities, pair it with male and female face galleries, embed both using contrastive VLMs, and quantify bias as the dif￾ference in mean similarities with bootstrap confidence intervals. Bias Score Calculation For a statement s with text embedding ts and two galleries of face embeddings Vm = {v m i } Nm i=1 and Vf = {v f j } Nf j=1, we define t… view at source ↗
Figure 4
Figure 4. Figure 4: End to end pipeline. Face images and occupation statements are embedded using contrastive VLMs to compute sim￾ilarity matrices Sm and Sf . Bias is measured as Bias(s) = Sm(s) − Sf (s) where positive values indicate male association and negative values indicate female association. Bootstrap confi￾dence intervals and null calibration validate results. 3. Results and Discussion We begin by comparing the obser… view at source ↗
Figure 5
Figure 5. Figure 5: ViT-B/32 statement bias. Top 25 female- and male￾associated statements [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Category-level bias across models. Mean bias scores with 95% CIs. 3.1. Top biased statements Top biased statements per model: ViT-B/32 associates males with firefighter, carpenter, truck driver and females with nurse, teacher, caregiver; ViT-L/14 shows pilot, CEO, engineer (male) vs. therapist, counselor, librarian (female); RN50 exhibits mechanic, builder, athlete (male) vs. coun￾selor, receptionist, care… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Vision-Language Models Suppress Female Representations Under Ambiguous Input

    cs.CV 2026-05 unverdicted novelty 6.0

    VLMs encode female associations internally for ambiguous images of female-stereotyped occupations but output male due to asymmetric layer-wise suppression, revealed by the new LALS metric across 15 occupations and fou...

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Exploring the frontier of vision- language models: A survey of current methodologies and future directions

    Akash Ghosh, Arkadeep Acharya, Sriparna Saha, Vinija Jain, and Aman Chadha. Exploring the frontier of vision- language models: A survey of current methodologies and future directions. arXiv preprint arXiv:2404.07214, 2024. 1

    arXiv 2024

  2. [2]

    Learning transferable visual models from natural language supervision

    Alec Radford et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2

    work page 2021

  3. [3]

    Vision-Language Models Performing Zero-Shot Tasks Exhibit Gender-based Disparities

    Melissa Hall, Laura Gustafson, et al. Vision-language mod- els performing zero-shot tasks exhibit gender-based dispari- ties. arXiv preprint arXiv:2301.11100, 2023. 1, 2, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    A survey on bias and fairness in machine learning

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys, 54(6):1–35,

    work page

  5. [5]

    Survey of social bias in vision-language models

    Nayeon Lee, Yejin Bang, Holy Lovenia, Samuel Cahyaw- ijaya, Wenliang Dai, and Pascale Fung. Survey of social bias in vision-language models. arXiv preprint arXiv:2309.14381, 2023. 1

    Pith/arXiv arXiv 2023

  6. [6]

    Vlstereoset: A study of stereotypical bias in pre-trained vision-language models

    Kankan Zhou, Yibin LAI, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. Association for Computational Linguistics, 2022. 1

    work page 2022

  7. [7]

    Can we talk models into seeing the world differently? arXiv preprint arXiv:2403.09193, 2024

    Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, M Jehanzeb Mirza, Margret Keuper, and Janis Keuper. Can we talk models into seeing the world differently? arXiv preprint arXiv:2403.09193, 2024. 1

    Pith/arXiv arXiv 2024

  8. [8]

    Unsupervised prompt learning for vision-language models

    Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 1

    Pith/arXiv arXiv 2022

  9. [9]

    Clip-adapter: Better vision-language models with fea- ture adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with fea- ture adapters. International Journal of Computer Vision , 132(2):581–595, 2024. 1

    work page 2024

  10. [10]

    Unveiling encoder-free vision-language models

    Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. Advances in Neural Information Processing Systems, 37:52545–52567, 2024. 1

    work page 2024

  11. [11]

    Visual-text cross alignment: Refining the similarity score in vision-language models

    Jinhao Li, Haopeng Li, Sarah Erfani, Lei Feng, James Bai- ley, and Feng Liu. Visual-text cross alignment: Refining the similarity score in vision-language models. arXiv preprint arXiv:2406.02915, 2024. 1

    Pith/arXiv arXiv 2024

  12. [12]

    Balancing the Picture: Debiasing Vision-Language Datasets with Synthetic Contrast Sets

    Brandon Smith, Miguel Farinha, et al. Balancing the picture: Debiasing vision-language datasets with synthetic contrast sets. In arXiv preprint arXiv:2305.15407, 2023. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Gender bias in neural natural lan- guage processing

    Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Aman- charla, and Anupam Datta. Gender bias in neural natural lan- guage processing. arXiv preprint arXiv:1807.11714, 2018. 1, 2

    Pith/arXiv arXiv 2018

  14. [14]

    Hadas Kotek, Rikker Dockum, and David Q. Sun. Gender bias and stereotypes in large language models.arXiv preprint arXiv:2308.14921, 2023. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    Vision-language models performing zero-shot tasks exhibit disparities between gender groups

    Melissa Hall, Laura Gustafson, Aaron Adcock, Ishan Misra, and Candace Ross. Vision-language models performing zero-shot tasks exhibit disparities between gender groups. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2778–2785, 2023. 2

    work page 2023

  16. [16]

    Se- mantics derived automatically from language corpora con- tain human-like biases

    Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. Se- mantics derived automatically from language corpora con- tain human-like biases. Science, 356(6334):183–186, 2017. 2

    work page 2017

  17. [17]

    On measuring social biases in sentence encoders

    Chandler May, Alex Wang, Shikha Bordia, Samuel R Bow- man, and Rachel Rudinger. On measuring social biases in sentence encoders. In Proceedings of NAACL-HLT, pages 622–628, 2019. 2

    work page 2019

  18. [18]

    Evaluating biased attitude associa- tions of language models

    S Omrani Sabbaghi et al. Evaluating biased attitude associa- tions of language models. In ACM, 2023. 2

    work page 2023

  19. [19]

    Gender detection & classification – face dataset

    Training Data Pro. Gender detection & classification – face dataset. https://www.kaggle.com/datasets/ trainingdatapro / gender - detection - and - classification-image-dataset, 2021. Accessed: 2025-08-15. 2

    work page 2021

  20. [20]

    Gender shades: Inter- sectional accuracy disparities in commercial gender classifi- cation

    Joy Buolamwini and Timnit Gebru. Gender shades: Inter- sectional accuracy disparities in commercial gender classifi- cation. In Conference on Fairness, Accountability and Trans- parency, pages 77–91, 2018. 2

    work page 2018

  21. [21]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. CVPR, pages 1521–1528, 2011. 2

    work page 2011

  22. [22]

    Revise: A tool for measuring and mit- igating bias in visual datasets

    Angelina Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. Revise: A tool for measuring and mit- igating bias in visual datasets. In ECCV, pages 733–751,

    work page

  23. [23]

    Fairface: Face at- tribute dataset for balanced race, gender, and age

    Kimmo Karkkainen and Jungseock Joo. Fairface: Face at- tribute dataset for balanced race, gender, and age. arXiv preprint arXiv:1908.04913, 2021. 2

    Pith/arXiv arXiv 1908

  24. [24]

    Fairness and bias in clip

    Sahil Ali et al. Fairness and bias in clip. arXiv preprint arXiv:2303.11408, 2023. 2

    Pith/arXiv arXiv 2023

  25. [25]

    GenderBias-\emph{VL}: Benchmarking Gender Bias in Vision Language Models via Counterfactual Probing

    Y Xiao et al. Genderbias-vl: Benchmarking gender bias in vision-language models via counterfactual probing. arXiv preprint arXiv:2407.00600, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Identifying implicit social biases in vision–language models

    Kimia Hamidieh et al. Identifying implicit social biases in vision–language models. AAAI/ACM Conference on AI, Ethics, and Society, 2024. 2

    work page 2024

  27. [27]

    A multi-dimensional study on bias in vi- sion–language models

    G Ruggeri et al. A multi-dimensional study on bias in vi- sion–language models. In Findings of ACL, 2023. 2

    work page 2023

  28. [28]

    Probing and mitigating intersectional social biases in vision–language models

    P Howard et al. Probing and mitigating intersectional social biases in vision–language models. In CVPR, 2024. 2

    work page 2024

  29. [29]

    Evaluating gender bias in multilingual mul- timodal ai models

    Kshitish Ghate. Evaluating gender bias in multilingual mul- timodal ai models. Findings of NL4AI, 2024. 3 5

    work page 2024

  30. [30]

    Racial and gen- der stereotypes encoded into clip

    Vatsal Baherwani and Joseph James Vincent. Racial and gen- der stereotypes encoded into clip. ICLR, 2024. 3

    work page 2024

  31. [31]

    Examining gender and racial bias in large vision-language models us- ing a novel dataset of parallel images

    Kathleen C Fraser and Svetlana Kiritchenko. Examining gender and racial bias in large vision-language models us- ing a novel dataset of parallel images. arXiv preprint arXiv:2402.05779, 2024. 3 6

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.