Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Stereotype Content Analysis on Color-related Social Bias in Large Vision Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Large vision-language models treat the color of a person's clothing as social information: blue-toned images are scored lower in competence and red-toned images lower in warmth, across all eight models tested.

desk verdict Useful early benchmark for color-related bias in LVLMs, with a clean blue-vs-red finding, but the white-baseline comparisons are confounded by an extra generation pass and the 'SCM is effective' claim needs human validation. read the letter →

arxiv 2505.20901 v1 pith:H7V752J6 submitted 2025-05-27 cs.CL cs.AI

Junhyuk Choi , Minju Kim , Yeseon Hong , Bugeun Kim This is my paper
classification cs.CLcs.AI
keywords largevisionlanguagemodelsstereotypesStereotypeContentModelcolorbiasgenderraceBASICbenchmarkevaluation
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

Large vision-language models (LVLMs) appear to carry color stereotypes that standard sentiment metrics miss. The paper builds BASIC, a paired-image benchmark in which the same person, pose, and scene are shown in white, blue, or red, and scores each model's free-form description with two new metrics derived from the Stereotype Content Model: competence and warmth. Across eight open models, blue-toned images are rated less competent than red- or white-toned ones, while red-toned images are rated less warm than blue-toned ones, with most differences reaching p < 0.001. The SCM metrics also surface gender and race differences that a sentiment metric does not, and the pattern changes with model family and parameter count rather than with architecture alone. The paper concludes that color should be treated as a first-class variable in stereotype audits of multimodal models.

What carries the argument

The load-bearing object is BASIC, a counterfactual paired-image benchmark. Starting from a seed portrait for each of 102 occupations and five action scenes, the pipeline edits gender and race (two genders by six racial groups), then recolors the image to blue or red while keeping pose, background, and composition as stable as generation allows, yielding 18,360 filtered images. The other load-bearing object is the SCM projection metric: sentence embeddings are projected onto a plane spanned by a warmth basis (the average embedding of words such as friendly, trustworthy, warm, sincere) and a competence basis (competent, confident, capable, efficient, intelligent, skillful), and the two projected coordinates are the warmth and competence scores. Pointwise mutual information between image attributes and SCM-relevant words is used to show which words drive the score differences.

What would settle it

Rebuild the paired image set while matching brightness, saturation, and non-target RGB channels between white, blue, and red versions (or restrict the color change to a small clothing region), then rerun the paired t-tests. If the blue-vs-red competence gap and the red-vs-blue warmth gap disappear, the reported color stereotypes are generation artifacts rather than model bias.

Watch

Extended reading notes

Core claim

The central claim is that an image's color tone is not a neutral background variable for LVLMs: it shifts the competence and warmth a model attributes to the person depicted. Blue images are scored lower in competence than red images in all eight models (p < 0.001) and lower than white images in seven of eight; red images are scored lower in warmth than blue images in all eight models. These findings come from the paper's SCM projection metrics, which are near-zero correlated with the sentiment metric (r = 0.014 for competence, r = 0.189 for warmth), so the color signal is largely invisible to sentiment-only evaluation. The paper also claims that neither architecture nor parameter size has an independent effect: Llama and Qwen families show more significant stereotypes with larger sizes, while InstructBLIP and Pixtral show fewer, so the two factors interact.

Load-bearing premise

The paired images are assumed to differ only in the attribute being tested—color, gender, or race—so any difference in a model's description is attributed to that attribute rather than to brightness, facial expression, pose, or background changes that the image generator introduced.

Editorial extensions

If this is right

  • Color must be controlled in future stereotype benchmarks: without matching color tones, gender and race comparisons can be confounded by the very effect this paper measures.
  • Sentiment-only audits undercount stereotypes, since the SCM warmth and competence scores are nearly uncorrelated with the sentiment score.
  • Scaling a model is not a reliable debiasing lever: larger versions of some families show more significant stereotypes and larger versions of others show fewer.
  • The SCM projection metric can be applied to any open-ended description, so existing model outputs can be re-scored without generating new images.
  • The paper's paired-image protocol gives an audit template: isolate one visual attribute, hold the rest fixed, and compare warmth and competence scores.

Reading between the lines

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

  • An extension the paper leaves implicit: the same projection metric could be applied to text-only language models, which would show whether the color effect lives in the vision encoder or in the language backbone's learned associations.
  • The RGB verification in the paper shows non-target color channels also shift, so a natural stress test is to regress the t-statistics on the residual RGB differences and ask whether the 'color' effect survives when non-target channels are statistically or generationally held constant.
  • The three chosen colors come from human perception studies; extending BASIC to hues such as green or yellow would test whether the warmth/competence pattern tracks specific cultural color symbolism or a more general hue-based effect.
  • Because the SCM anchor words are taken from English social psychology, the metric may not transfer to other languages; a multilingual version would need locally validated anchor terms.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces two SCM-based evaluation metrics—warmth and competence scores obtained by projecting sentence embeddings onto word-embedding basis vectors built from six or seven seed words drawn from Fiske et al. (2018)—and a new benchmark, BASIC, comprising paired images that vary color (white/blue/red), gender, and race across 102 occupations. Using eight open LVLMs, the authors run paired t-tests and PMI analyses on model-generated descriptions, reporting that (1) SCM metrics are more effective than VADER sentiment at surfacing stereotypes, (2) LVLMs show color stereotypes, e.g., blue images are rated lower in competence than red/white and red images lower in warmth than blue/white, and (3) architecture and parameter size interact in shaping stereotype expression. The dataset and generation code are released.

Significance. If the central counterfactual and metric-validation assumptions hold, the paper would make a useful contribution: it extends stereotype auditing to a visual attribute (color) that is often left uncontrolled in paired-image benchmarks, proposes a psychologically grounded scoring method that is independent of sentiment polarity, and provides a systematic comparison across eight models with varied architectures and sizes. The public release of BASIC, the detailed six-step generation protocol, and the use of an open-ended, non-leading prompt are concrete strengths. However, the empirical color findings rest on the assumption that paired images differ only in hue, which is not established; and the claim that the SCM metrics are 'effective' needs external validation beyond a near-zero correlation with VADER.

major comments (4)
  1. [§3, Steps 4–5 and Table 3] The Blue-vs-White and Red-vs-White comparisons in Table 3 are not counterfactual color comparisons. Per Appendix A.5, white images are the output of Step 4, while blue and red images are produced from those white images by an additional SDXL img2img pass (strength 0.8, guidance 11) with different positive/negative prompts. The white condition never undergoes this second diffusion pass, so any Blue-vs-White or Red-vs-White difference conflates color with an extra generation step that can alter pose, expression, lighting, or composition. This directly affects the competence Blue-vs-White rows (significant for 7 of 8 models) and the warmth Blue/Red-vs-White rows. The authors should either generate the white condition through an identical second pass or restrict the color-stereotype claim to Blue-vs-Red pairs that share the generation pipeline.
  2. [§3, Table 1 and Appendix A.5] Even for the Blue-vs-Red pairs, the paper does not verify that the images differ only in hue. Table 1 reports mean RGB channel changes relative to white (blue: R −5.37%, G −4.87%, B −0.52%; red: R +0.12%, G −3.06%, B −2.43%), which indicate systematic brightness/saturation shifts, and these are means over the dataset rather than per-pair checks. No verification is provided for facial expression, pose, background, clothing fit, or skin-tone rendering. The Limitations section explicitly concedes that 'facial expression, brightness, or posture may still have a potential influence on the model outputs.' Because the SCM scores are computed from free-form descriptions, a systematic low-level visual difference across color conditions could produce the observed warmth/competence gaps even if the LVLMs encode no color stereotype. This is load-bearing for the central empirical claim and needs per-image counterfactual verification (e.g., human or automated checks that pairs differ only in the targeted attribute, or a control analysis on images filtered to match in brightness and expression).
  3. [§4.2.2 and §6.1] The claim that the SCM-based evaluation is 'effective in capturing stereotypes' is not validated. The metrics are constructed from only six to seven seed words per dimension from Fiske et al. (2018), and the paper shows that SCM scores are nearly uncorrelated with VADER (r = 0.014 and r = 0.189), which demonstrates independence but not validity. To support the effectiveness claim, the authors should validate the scores against human SCM judgments on a sample of descriptions, or show that known stereotyped and counter-stereotyped descriptions shift the scores in the expected directions. As written, the definition of a stereotype as a significant difference in the SCM projection makes the effectiveness finding partly self-referential.
  4. [§5, Table 3 and §4.3] The statistical reporting needs effect sizes and multiple-comparison correction. Each model is tested on 19 pairwise comparisons across three metrics, and the '# of significant stereotypes' counts in Rows 5 of Table 3 and in Section 6.3 are based on uncorrected p < 0.05 thresholds. With 8 models and 57 tests per model, the counts and the architecture/parameter comparisons are inflated by expected false positives. The main Blue-vs-Red effects (p < 0.001 for all models) would likely survive a Bonferroni correction, but weaker rows (notably many VADER rows and some race/gender comparisons) may not. The paper should report effect sizes such as Cohen's d and either apply a correction or justify the uncorrected thresholds.
minor comments (5)
  1. [Table 2] The parameter-size row is difficult to read: the values are not aligned with the model columns, and 'Large' is used without a number (the surrounding text later mentions 124B for Pixtral Large). Please clarify which size belongs to which model and state the exact parameter counts used in the analyses.
  2. [Appendix C, Table 7] Table 7 is captioned 'Full experimental result for Competence metric in SCM,' but the text in Appendix C.1 and the values (e.g., positive warmth differences for male vs. female) indicate this table reports warmth, not competence. The caption should be corrected.
  3. [Appendix D, PMI tables] A few PMI word lists contain misspellings or spurious tokens, such as 'frendly' and 'handworking' (Tables 19 and 12 in the appendix). A light cleaning step for the PMI vocabulary, or a note that no spell-checking was applied, would improve readability and reproducibility.
  4. [§4.2.2, Eqs. (2)–(6)] The projection framework is clear, but the exact Sentence-Transformer checkpoint used for ST(·) is never named. Please specify the model (e.g., all-MiniLM-L6-v2 or a specific sentence-transformers version) so the metric can be reproduced exactly.
  5. [§4.1 and Appendix B] The generation temperature is set to 0 for deterministic responses, but it is not stated whether this applies to all eight models through the OpenRouter API and the InstructBLIP transformers calls, or whether some models ignore temperature settings. A brief per-model note would strengthen reproducibility.

Circularity Check

2 steps flagged · score 2.0 of 10

Minor self-referential evaluation claim; central color finding is independent.

  1. self definitional [Abstract and Section 4.2.2 (SCM-based metrics)]
    "(1) The SCM-based evaluation is effective in capturing stereotypes. ... we propose two new metrics following SCM: competence and warmth. ... we interpret them as scores quantifying warmth and competence in SCM."

    Stereotypes are operationally defined as statistically significant differences in the SCM projection scores between attribute conditions (Section 4.3 paired t-tests on alpha_w and alpha_c). Therefore, the finding that the SCM-based evaluation 'captures stereotypes' is largely entailed by the measurement definition: the metric defines the stereotype differences it then detects. No external ground truth, such as human stereotype ratings of the same descriptions or images, is used to validate that SCM score gaps correspond to stereotypes. This makes contribution (1) self-referential, although the color-specific differences remain an independent empirical observation about LVLM outputs.

  2. self definitional [Section 4.3 (PMI analysis) and Section 5 (PMI results)]
    "we filtered words based on the cosine similarity. We computed cosine similarity between w := ST(w) and basis vectors and discarded words whose similarity is in between -0.5 and 0.5. Then, we compute PMI ... PMI result supports this observation; red and white were associated with high competence words, such as BRIGHT, STRONG, COMPETITIVE (for red), CLEAN, TECHNOLOGICAL, and EXPERTISE (white)."

    The 'supporting' PMI words are selected by thresholding their cosine similarity to the same warmth and competence basis vectors used to define the SCM metrics. Thus, the PMI analysis does not provide independent confirmation of the SCM findings; it re-identifies words close to the metric's own axes and reports their PMI with the attribute. It is illustrative rather than independent evidence, but the paper presents it as support ('PMI result supports this observation').

full rationale

The central empirical claim, that LVLMs show color-related differences in SCM-based warmth and competence scores, is not circular: the metrics use externally grounded word lists from Fiske et al. (2018) and Sentence-Transformer embeddings, and the image pairs are generated by an independent procedure. The paired t-tests compare generated descriptions across color conditions, so the observed differences are a real property of the model outputs rather than an artifact of fitting. The two flagged issues are auxiliary. First, the 'effectiveness' of SCM-based evaluation is partly self-referential because stereotypes are operationally defined as SCM score differences, with no external validation. Second, the PMI 'support' uses the same SCM basis vectors to preselect words, so it does not independently corroborate the metric. Neither issue undermines the color-stereotype finding itself, which stands on the t-tests and the externally grounded metric construction. The white-vs-color image generation confound noted in the Limitations section is a validity threat rather than a circularity, so it does not raise the circularity score. Overall, the derivation is self-contained apart from the minor self-referential evaluation claim, giving a score of 2.

Assumptions & free parameters 3 free parameters · 8 assumptions · 0 invented entities

The paper relies on assumptions about embedding geometry, image counterfactual validity, and the validity of its SCM word lists. The most critical is counterfactual validity: if the paired images differ systematically beyond the targeted attribute, the attribution of observed score differences to color, gender, or race is invalid.

free parameters (3)
  • PMI word filtering threshold = 0.5
    Hand-chosen at cos(60°) to discard words nearly orthogonal to the SCM axes; affects PMI word lists in Appendix D but not the main t-test results.
  • CLIP semantic similarity threshold = 0.2
    Hand-chosen to retain images whose CLIP embedding matches the text prompt; affects which occupations survive Step 6 filtering.
  • Image generation strengths and guidance = strength 0.7/0.8, guidance 9/11, 50 steps
    Manual settings for SDXL text-to-image and image-to-image; could influence image properties, but are standard and not fit to the stereotype results.
assumptions (8)
  • domain assumption Sentence embeddings can be written as a linear combination of independent warmth and competence basis vectors plus orthogonal components.
    Invoked in Section 4.2.2 to justify interpreting projection coefficients as SCM scores; no geometric justification is given.
  • domain assumption The normalized average embedding of six competence words and seven warmth words provides valid axes for these dimensions.
    Section 4.2.2, Eqs. (2)-(3); relies on the chosen words from Fiske et al. (2018) adequately spanning the intended constructs.
  • domain assumption Generated paired images differ only in the targeted attribute (color, gender, or race) while holding pose, background, and other visual properties constant.
    Section 3 Steps 4-5 and Table 1 attempt to verify this via RGB changes, but image-to-image generation may still introduce subtle differences in expression, brightness, or composition.
  • domain assumption VADER sentiment scores are a meaningful baseline for comparison.
    Section 4.2.1; the paper treats VADER as representative of prior sentiment-based stereotype metrics.
  • domain assumption The open prompt 'Describe this person objectively' elicits descriptions reflecting stereotypes rather than purely superficial visual features.
    Section 4.1; no experiments validate this assumption.
  • domain assumption The binary gender and six race categories used are appropriate operationalizations for stereotype analysis.
    Section 3; categories adopted from Karkkainen and Joo (2021) and Howard et al. (2024); authors acknowledge broader spectra exist.
  • standard math Paired t-tests are valid for the comparison of metric values across image pairs, assuming independence of pairs and approximate normality.
    Section 4.3; with large sample sizes (18,360 images), normality is likely satisfied, but pairs from the same occupation may not be fully independent.
  • domain assumption Excluding an entire occupation if any image fails filtering does not bias the remaining occupation set in a way that correlates with stereotypes.
    Section 3 Step 6; the filtering could preferentially remove occupations with certain visual properties, potentially affecting racial or color comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Stereotype Content Analysis on Color-related Social Bias in Large Vision Language Models." pith.science (2026). https://pith.science/paper/H7V752J6

@misc{pith2026250520901,
  author       = {Pith},
  title        = {Pith review of: A Stereotype Content Analysis on Color-related Social Bias in Large Vision Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7V752J6}},
  note         = {Machine review of arXiv:2505.20901}
}
read the original abstract

As large vision language models(LVLMs) rapidly advance, concerns about their potential to learn and generate social biases and stereotypes are increasing. Previous studies on LVLM's stereotypes face two primary limitations: metrics that overlooked the importance of content words, and datasets that overlooked the effect of color. To address these limitations, this study introduces new evaluation metrics based on the Stereotype Content Model (SCM). We also propose BASIC, a benchmark for assessing gender, race, and color stereotypes. Using SCM metrics and BASIC, we conduct a study with eight LVLMs to discover stereotypes. As a result, we found three findings. (1) The SCM-based evaluation is effective in capturing stereotypes. (2) LVLMs exhibit color stereotypes in the output along with gender and race ones. (3) Interaction between model architecture and parameter sizes seems to affect stereotypes. We release BASIC publicly on [anonymized for review].

Figures

Figures reproduced from arXiv: 2505.20901 by the authors.

Figure 1
Figure 1. Examples from our BASIC dataset. Top row [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Process of Image Generation in BASIC for Asian male images. Inspired by previous work (Howard et al., 2024), we also used the same terms Black/White to represent African￾American/Caucasian people. This yields 12 white-clothing images per occupation. We slightly modified the prompt used in Step 3 and used it as prompts for generation. We re￾placed ‘Asian’ and ‘man’ with placeholders race and gender, as follows: {race… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Cultural Counterfactuals: Evaluating Cultural Biases in Large Vision-Language Models with Counterfactual Examples

    cs.CV 2026-03 conditional novelty 6.0 of 10

    Cultural Counterfactuals — same person placed in different cultural contexts — shows that LVLMs vary salary, rent, and character judgments with the depicted religion, nationality, and income level.

Reference graph

Works this paper leans on

56 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hajo Adam and Adam D Galinsky. 2012. Enclothed cognition. Journal of experimental social psychology, 48(4):918--925

    work page 2012

  2. [2]

    Tosin Adewumi, Lama Alkhaled, Namrata Gurung, Goya van Boven, and Irene Pagliai. 2024. https://arxiv.org/abs/2406.19097 Fairness and bias in multimodal ai: A survey . Preprint, arXiv:2406.19097

    arXiv 2024

  3. [3]

    Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. 2024. Pixtral 12b. arXiv preprint arXiv:2410.07073

    arXiv 2024

  4. [4]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923

    arXiv 2025

  5. [5]

    Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. 2024. Understanding the impact of negative prompts: When and how do they take effect? In European Conference on Computer Vision, pages 190--206. Springer

    work page 2024

  6. [6]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. https://lmsys.org/blog/2023-03-30-vicuna/ Vicuna: An open-source chatbot impressing gpt-4 with 90\

    2023

  7. [7]

    Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. 2023. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070

    arXiv 2023

  8. [8]

    Jasmin Cloutier, Malia F Mason, and C Neil Macrae. 2005. The perceptual determinants of person construal: reopening the social-cognitive toolbox. Journal of personality and social psychology, 88(6):885

    work page 2005

Show all 56 references
  1. [9]

    what if it’s not just an item of clothing?

    Coline Crutzen and St \'e phane Adam. 2022. “what if it’s not just an item of clothing?”--a narrative review and synthesis of the white coat in the context of aged care. Psychologica Belgica, 62(1):62

    2022

  2. [10]

    Kanjar De and Marius Pedersen. 2021. https://doi.org/10.1109/ICCVW54120.2021.00009 Impact of colour on robustness of deep neural networks . In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 21--30

    2021

  3. [11]

    Jennifer L Eberhardt, Phillip Atiba Goff, Valerie J Purdie, and Paul G Davies. 2004. Seeing black: race, crime, and visual processing. Journal of personality and social psychology, 87(6):876

    2004

  4. [12]

    Andrew Elliot and Markus Maier. 2013. https://doi.org/10.1146/annurev-psych-010213-115035 Color psychology: Effects of perceiving color on psychological functioning in humans . Annual review of psychology, 65

    2013 doi

  5. [13]

    Andrew J Elliot, Jessica L Tracy, Adam D Pazda, and Alec T Beall. 2013. Red enhances women's attractiveness to men: First evidence suggesting universality. Journal of Experimental Social Psychology, 49(1):165--168

    2013

  6. [14]

    Susan T Fiske, Amy JC Cuddy, Peter Glick, and Jun Xu. 2018. A model of (often mixed) stereotype content: Competence and warmth respectively follow from perceived status and competition. In Social cognition, pages 162--214. Routledge

    2018

  7. [15]

    Kathleen Fraser and Svetlana Kiritchenko. 2024. https://aclanthology.org/2024.eacl-long.41/ Examining gender and racial bias in large vision -- language models using a novel dataset of parallel images . In Proceedings of the 18th Conference of the European Chapter of the Assoc...

    2024

  8. [16]

    Fraser, Isar Nejadgholi, and Svetlana Kiritchenko

    Kathleen C. Fraser, Isar Nejadgholi, and Svetlana Kiritchenko. 2021. https://doi.org/10.18653/v1/2021.acl-long.50 Understanding and countering stereotypes: A computational approach to the stereotype content model . In Proceedings of the 59th Annual Meeting of the Association f...

    2021 doi

  9. [17]

    Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas K Sharma, Tom Hartvigsen, and Marzyeh Ghassemi. 2024. Bendvlm: Test-time debiasing of vision-language embeddings. Advances in Neural Information Processing Systems, 37:62480--62502

    2024

  10. [18]

    Timo Gnambs, Markus Appel, and Aileen Oeberst. 2015. Red color and risk-taking behavior in online environments. PloS one, 10(7):e0134033

    2015

  11. [19]

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

    2024 arXiv

  12. [20]

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. 2024. https://doi.org/10.1609/aaai.v38i3.27963 Anomalygpt: detecting industrial anomalies using large vision-language models . In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Inte...

    2024 doi

  13. [21]

    Fraser, Anahita Bhiwandiwalla, and Svetlana Kiritchenko

    Phillip Howard, Kathleen C. Fraser, Anahita Bhiwandiwalla, and Svetlana Kiritchenko. 2025. https://aclanthology.org/2025.naacl-long.305/ Uncovering bias in large vision-language models at scale with counterfactuals . In Proceedings of the 2025 Conference of the Nations of the ...

    2025

  14. [22]

    Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. 2024. https://doi.org/10.1109/CVPR52733.2024.01138 SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Ex...

    2024

  15. [23]

    Tiancheng Hu, Yara Kyrychenko, Steve Rathje, Nigel Collier, Sander van der Linden, and Jon Roozenbeek. 2025. Generative language models exhibit social identity biases. Nature Computational Science, 5(1):65--75

    2025

  16. [24]

    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

  17. [25]

    Keito Inoshita and Xiaokang Zhou. 2024. Sentiment bias and security analysis in training datasets of large language models. In 2024 IEEE International Conference on Big Data and Cloud Computing (BDCloud), pages 1--8. IEEE

    2024

  18. [26]

    Sepehr Janghorbani and Gerard De Melo. 2023. https://doi.org/10.18653/v1/2023.eacl-main.126 Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision -- language models . In Proceedings of the 17th Conference of the European C...

    2023 doi

  19. [27]

    Yukun Jiang, Zheng Li, Xinyue Shen, Yugeng Liu, Michael Backes, and Yang Zhang. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.713 ModSCAN : Measuring stereotypical bias in large vision-language models from vision and language modalities . In Proceedings of the 2024 Confere...

    2024 doi

  20. [28]

    Jayendra Kantipudi, Shiv Ram Dubey, and Soumendu Chakraborty. 2020. https://doi.org/10.1109/TAI.2020.3046167 Color channel perturbation attacks for fooling convolutional neural networks and a defense against such attacks . IEEE Transactions on Artificial Intelligence, 1(2):181--191

    2020

  21. [29]

    Kimmo Karkkainen and Jungseock Joo. 2021. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1548--1558

    2021

  22. [30]

    Naz Kaya and Helen H Epps. 2004. Relationship between color and emotion: A study of college students. College student journal, 38(3):396--405

    2004

  23. [31]

    Jun Seong Kim, Kyaw Ye Thu, Javad Ismayilzada, Junyeong Park, Eunsu Kim, Huzama Ahmad, Na Min An, James Thorne, and Alice Oh. 2025. https://arxiv.org/abs/2503.16826 When tom eats kimchi: Evaluating cultural bias of multimodal large language models in cultural mixture contexts ...

    2025 arXiv

  24. [32]

    Michelle YoungJin Kim and Kristen Johnson. 2025. Korean stereotype content model: Translating stereotypes across cultures. In Proceedings of the 3rd Workshop on Cross-Cultural Considerations in NLP (C3NLP 2025), pages 59--70

    2025

  25. [33]

    Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.358 VLF eedback: A large-scale AI feedback dataset for large vision-language models alignment . In Procee...

    2024 doi

  26. [34]

    Yijun Liang, Ming Li, Chenrui Fan, Ziyue Li, Dang Nguyen, Kwesi Cobbina, Shweta Bhardwaj, Jiuhai Chen, Fuxiao Liu, and Tianyi Zhou. 2025. https://arxiv.org/abs/2504.10514 Colorbench: Can vlms see and understand the colorful world? a comprehensive benchmark for color perception...

    2025

  27. [35]

    Abhishek Mandal, Susan Leavy, and Suzanne Little. 2023. Measuring bias in multimodal models: Multimodal composite association score. In International Workshop on Algorithmic Bias in Search and Recommendation, pages 17--30. Springer

    2023

  28. [36]

    Nooree Na and Hyeon Jeong Suk. 2014. The emotional characteristics of white for applications of product color design. Industrial applications of affective engineering, pages 253--264

    2014

  29. [37]

    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

  30. [38]

    Ranjita Naik and Besmira Nushi. 2023. https://doi.org/10.1145/3600211.3604711 Social biases through the text-to-image generation lens . In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, AIES '23, page 786–808, New York, NY, USA. Association for Computi...

    2023

  31. [39]

    Gandalf Nicolas, Xuechunzi Bai, and Susan T Fiske. 2021. Comprehensive stereotype content dictionaries using a semi-automated method. European Journal of Social Psychology, 51(1):178--196

    2021

  32. [40]

    Ali Omrani, Alireza Salkhordeh Ziabari, Charles Yu, Preni Golazizian, Brendan Kennedy, Mohammad Atari, Heng Ji, and Morteza Dehghani. 2023. https://doi.org/10.18653/v1/2023.acl-long.227 Social-group-agnostic bias mitigation via the stereotype content model . In Proceedings of ...

    2023 doi

  33. [41]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. 2023. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    2023 arXiv

  34. [42]

    Haoyi Qiu, Zi-Yi Dou, Tianlu Wang, Asli Celikyilmaz, and Nanyun Peng. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.520 Gender biases in automatic evaluation metrics for image captioning . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Proce...

    2023 doi

  35. [43]

    Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan, Antonios Anastasopoulos, and Ziwei Zhu. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.611 B ias D ora: Exploring hidden biased associations in vision-language models . In Findings of the Association for Computational Ling...

    2024 doi

  36. [44]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084

    2019 arXiv

  37. [45]

    Ashutosh Sathe, Prachi Jain, and Sunayana Sitaram. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.66 A unified framework and dataset for assessing societal bias in vision-language models . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1...

    2024 doi

  38. [46]

    John Oliver Siy and Sapna Cheryan. 2016. Prejudice masquerading as praise: The negative echo of positive stereotypes. Personality and Social Psychology Bulletin, 42(7):941--954

    2016

  39. [47]

    Gary R VandenBos. 2007. APA dictionary of psychology. American Psychological Association

    2007

  40. [48]

    Guankun Wang, Long Bai, Wan Jun Nah, Jie Wang, Zhaoxi Zhang, Zhen Chen, Jinlin Wu, Mobarakol Islam, Hongbin Liu, and Hongliang Ren. 2025. https://arxiv.org/abs/2405.10948 Surgical-lvlm: Learning to adapt large vision-language model for grounded visual question answering in rob...

    2025 arXiv

  41. [49]

    Sibo Wang, Xiangkui Cao, Jie Zhang, Zheng Yuan, Shiguang Shan, Xilin Chen, and Wen Gao. 2024 a . https://arxiv.org/abs/2406.14194 Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model . Preprint, arXiv:2406.14194

    2024 arXiv

  42. [50]

    Wenxuan Wang, Haonan Bai, Jen-tse Huang, Yuxuan Wan, Youliang Yuan, Haoyi Qiu, Nanyun Peng, and Michael Lyu. 2024 b . https://doi.org/10.1145/3664647.3681433 New job, new gender? measuring the social bias in image generation models . In Proceedings of the 32nd ACM Internationa...

    2024

  43. [51]

    Lois B Wexner. 1954. The degree to which colors (hues) are associated with mood-tones. Journal of applied psychology, 38(6):432

    1954

  44. [52]

    Shengguang Wu, Fan-Yun Sun, Kaiyue Wen, and Nick Haber. 2025. https://arxiv.org/abs/2502.13928 Symmetrical visual contrastive optimization: Aligning vision-language models with minimal contrastive images . Preprint, arXiv:2502.13928

    2025 arXiv

  45. [53]

    Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. 2024. https://arxiv.org/abs/2406.17974 Evaluating fairness in large vision-language models across diverse demographic attributes and prompts . Preprint, arXiv:2406.17974

    2024

  46. [54]

    Peng Xu, Wenqi Shao, Kaipeng Zhang, Peng Gao, Shuo Liu, Meng Lei, Fanqing Meng, Siyuan Huang, Yu Qiao, and Ping Luo. 2024. Lvlm-ehub: A comprehensive evaluation benchmark for large vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  47. [55]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  48. [56]

    write newline

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

Pith tools

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