StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs

The article introduces StylisticBias, a benchmark designed to evaluate attribute-level social bias in multimodal large language models (MLLMs). It features a dataset of 25,000 images generated from 500 photorealistic faces with variations in 15 visual attributes, revealing that age and body type primarily influence identity-level effects, while fashion style and other cues drive significant attribute-level shifts. This benchmark is crucial for practitioners as it provides a controlled method to assess and mitigate biases in MLLMs, particularly in contexts where visual cues impact social judgments.