Reproduce “Ask Me Again Differently: GRAS for Measuring Bias in Vision Language Models on Gender, Race, Age, and Skin Tone” on OpenResearch

About this paper

Shaivi Malik, Hasnat Md Abdullah, Sriparna Saha, Amit Sheth

Researchers introduced GRAS, a new benchmark and an interpretable metric, to systematically evaluate demographic biases in Vision Language Models (VLMs) across gender, race, age, and skin tone. Their analysis revealed state-of-the-art VLMs exhibit extremely high levels of bias, consistently associating negative personality traits with darker skin tones and positive traits with lighter ones.

Key insights

• The linguistic formulation and framing of questions in VQA significantly affect VLM bias evaluations, highlighting the need for diverse question templates in bias assessment.
• State-of-the-art VLMs demonstrate pervasive and severe demographic biases, consistently attributing negative personality traits to certain groups (e.g., darker skin tones, older individuals) and positive traits to others.
• Bias manifests in specific patterns of valence-based trait attribution, such as associating male and Middle Eastern individuals with a higher percentage of negative traits, while White and lighter-skinned individuals receive more positive attributions.

Summary by alphaXiv.