Papers
Topics
Authors
Recent
Assistant
AI Research Assistant
Well-researched responses based on relevant abstracts and paper content.
Custom Instructions Pro
Preferences or requirements that you'd like Emergent Mind to consider when generating responses.
Gemini 2.5 Flash
Gemini 2.5 Flash 89 tok/s
Gemini 2.5 Pro 53 tok/s Pro
GPT-5 Medium 26 tok/s Pro
GPT-5 High 25 tok/s Pro
GPT-4o 93 tok/s Pro
Kimi K2 221 tok/s Pro
GPT OSS 120B 457 tok/s Pro
Claude Sonnet 4 38 tok/s Pro
2000 character limit reached

Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs) (2410.19314v2)

Published 25 Oct 2024 in cs.CY and cs.CL

Abstract: Pre-trained LLMs have been reliably integrated with visual input for multimodal tasks. The widespread adoption of instruction-tuned image-to-text vision-language assistants (VLAs) like LLaVA and InternVL necessitates evaluating gender biases. We study gender bias in 22 popular open-source VLAs with respect to personality traits, skills, and occupations. Our results show that VLAs replicate human biases likely present in the data, such as real-world occupational imbalances. Similarly, they tend to attribute more skills and positive personality traits to women than to men, and we see a consistent tendency to associate negative personality traits with men. To eliminate the gender bias in these models, we find that fine-tuning-based debiasing methods achieve the best trade-off between debiasing and retaining performance on downstream tasks. We argue for pre-deploying gender bias assessment in VLAs and motivate further development of debiasing strategies to ensure equitable societal outcomes. Code is available at https://github.com/ExplainableML/vla-gender-bias.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (72)
  1. Phi-3 technical report: A highly capable language model locally on your phone. In arXiv, 2024.
  2. Gpt-4 technical report. In arXiv, 2023.
  3. Evaluating clip: towards characterization of broader capabilities and downstream implications. In arXiv, 2021.
  4. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022.
  5. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. In arXiv, 2023.
  6. Intersectionality in quantitative research: A systematic review of its emergence and applications of theory and methods. In SSM-population health, 2021.
  7. On the dangers of stochastic parrots: Can language models be too big? In ACM conference on fairness, accountability, and transparency, 2021.
  8. A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. In AACL-IJCNLP, 2022.
  9. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In ACM Conference on Fairness, Accountability, and Transparency, 2023.
  10. Internlm2 technical report. In arXiv, 2024.
  11. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. In arXiv, 2024a.
  12. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024b.
  13. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023.
  14. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, 2023.
  15. Mobilevlm v2: Faster and stronger baseline for vision language model. In arXiv, 2024.
  16. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In arXiv, 2024.
  17. All our n-gram are belong to you, 2006.
  18. Examining gender and racial bias in large vision–language models using a novel dataset of parallel images. In EACL, 2024.
  19. Mme: A comprehensive evaluation benchmark for multimodal large language models. In arXiv, 2023.
  20. Bias and fairness in large language models: A survey. In Computational Linguistics, 2024.
  21. Uncurated image-text datasets: Shedding light on demographic bias. In CVPR, 2023.
  22. Gemini: a family of highly capable multimodal models. In arXiv, 2023.
  23. Unboxing occupational bias: Grounded debiasing llms with us labor data. In arXiv, 2024.
  24. Visogender: A dataset for benchmarking gender bias in image-text pronoun resolution. In NeurIPS, 2024.
  25. Efficient multimodal learning from data-centric perspective. In arXiv, 2024.
  26. Madeline E Heilman. Gender stereotypes and workplace bias. In Research in organizational Behavior, 2012.
  27. Women at work: pathways from gender stereotypes to gender bias and discrimination. In Annual Review of Organizational Psychology and Organizational Behavior, 2024.
  28. Social bias evaluation for large language models requires prompt variations. In arXiv, 2024.
  29. Quantifying societal bias amplification in image captioning. In CVPR, 2022.
  30. Uncovering bias in large vision-language models at scale with counterfactuals. In arXiv, 2024.
  31. Unsupervised discovery of gendered language through latent-variable modeling. In ACL, 2019.
  32. LoRA: Low-rank adaptation of large language models. In ICLR, 2022.
  33. Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision–language models. In EACL, 2023.
  34. Kameni Florentin Flambeau Jiechieu and Norbert Tsopze. Skills prediction based on multi-label resume classification using cnn with model predictions explanation. In Neural Computing and Applications, 2021.
  35. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, 2021.
  36. Measuring bias in contextualized word representations. In First Workshop on Gender Bias in Natural Language Processing, 2019.
  37. Optimal brain damage. In NeurIPS, 1989.
  38. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021.
  39. Seed-bench-2: Benchmarking multimodal large language models. In arXiv, 2023a.
  40. Seed-bench: Benchmarking multimodal large language models. In CVPR, 2024a.
  41. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023b.
  42. Red teaming visual language models. In Findings of the Association for Computational Linguistics, 2024b.
  43. Improved baselines with visual instruction tuning. In CVPR, 2024a.
  44. Llava-next: Improved reasoning, ocr, and world knowledge, 2024b.
  45. Visual instruction tuning. In NeurIPS, 2024c.
  46. Mmbench: Is your multi-modal model an all-around player? In arXiv, 2023.
  47. Stable bias: Evaluating societal representations in diffusion models. In NeurIPS, 2024.
  48. Llm-pruner: On the structural pruning of large language models. In NeurIPS, 2023.
  49. Debiasing surgeon: fantastic weights and how to find them. In arXiv, 2024.
  50. Moses Olafenwa. Idenprof, 2018.
  51. Learning transferable visual models from natural language supervision. In ICML, 2021.
  52. Status incongruity and backlash effects: Defending the gender hierarchy motivates prejudice against female leaders. In Journal of experimental social psychology, 2012.
  53. A multi-dimensional study on bias in vision-language models. In Findings of the Association for Computational Linguistics, 2023.
  54. A unified framework and dataset for assessing gender bias in vision-language models. In arXiv, 2024.
  55. A step toward more inclusive people annotations for fairness. In AIES, 2021.
  56. Quantifying language models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. In ICLR, 2024.
  57. Quantifying social biases using templates is unreliable. In NeurIPS Workshop on Trustworthy and Socially Responsible Machine Learning, 2022.
  58. The bias amplification paradox in text-to-image generation. In NAACL, 2024.
  59. Dear: Debiasing vision-language models with additive residuals. In CVPR, 2023.
  60. SkunkworksAI. Bakllava, 2023.
  61. Aligning large multimodal models with factually augmented rlhf. In Findings of the Association for Computational Linguistics, 2024.
  62. Discovering and mitigating biases in clip-based image editing. In WACV, 2024.
  63. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. In arXiv, 2024.
  64. Stable diffusion exposed: Gender bias from prompt to image. In arXiv, 2023.
  65. Genderbias-vl: Benchmarking gender bias in vision language models via counterfactual probing. In arXiv, 2024.
  66. Gendered skills and unemployed men’s resistance to “women’s work”. In Gender, Work & Organization, 2021.
  67. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, 2024.
  68. Sigmoid loss for language image pre-training. In ICCV, 2023.
  69. Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model. In arXiv, 2024.
  70. Gender bias in coreference resolution: Evaluation and debiasing methods. In NAACL, 2018.
  71. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In AACL-IJCNLP, 2022.
  72. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024.

Summary

We haven't generated a summary for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Summarize Now
Lightbulb On Streamline Icon: https://streamlinehq.com

Continue Learning

We haven't generated follow-up questions for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Generate Now
List To Do Tasks Checklist Streamline Icon: https://streamlinehq.com

Collections

Sign up for free to add this paper to one or more collections.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up
X Twitter Logo Streamline Icon: https://streamlinehq.com

Tweets

This paper has been mentioned in 1 post and received 3 likes.

User Circle Single Streamline Icon: https://streamlinehq.com Sign Up to View