Reproduce “Measuring Social Bias in Vision-Language Models with Face-Only Counterfactuals from Real Photos” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Haodong Chen, Qiang Huang, Jiaqi Zhao, Qiuping Jiang, Xiaojun Chang, Jun Yu

Researchers from Harbin Institute of Technology (Shenzhen) developed the FOCUS dataset and REFLECT benchmark to precisely measure social bias in Vision-Language Models. This method uses face-only counterfactuals derived from real photos, demonstrating that demographic disparities consistently persist across various decision tasks even with tightly controlled visual contexts.

Key insights

  • Demographic disparities consistently persist in state-of-the-art VLMs across all evaluation tasks, even under strict counterfactual control where only facial attributes vary.
  • The manifestation of bias is context- and task-dependent; gender preferences in 2AFC tasks, for instance, flipped between favoring male variants for 'Income' and female variants for 'Perceived Safety'.
  • Occupation acts as a strong moderator of bias, with the 'CEO' occupation consistently amplifying observed disparities, such as significant race penalties for non-White groups in salary recommendations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.