Age and Body Type Dominate Social Bias in MLLMs Across 25K Images

StylisticBias benchmarks attribute-level social bias in MLLMs using 500 base faces with ~50 single-attribute variations each, totaling ~25K images, keeping identity fixed to isolate visual cue effects. Across six MLLMs and 25 binary judgment scenarios, age and body type drive the largest identity-level biases, while fashion style causes the biggest attribute-level shifts.